Modeled data. No real people, clients, or companies.
The purpose
Why did I build this?
Here it is: a dashboard a branch manager can understand, and an email that tells each branch what to do next.
What I learned about the industry
Scroll both ways, or pinch to zoom.
| visits accepted | typed in by hand | sent to the state | turned back | the manual half of the score | the rejected half of the score | the office: not on the report |
|---|---|---|---|---|---|---|
| 3,034 | 254 | 3,146 | 103 | 54.98 | 38.69 |
Its report counts visits by contract. An agency with seven offices has to work out for itself where the trouble is.
Scroll both ways, or pinch to zoom.
Every contract El Paso holds is heading to pass.
81 worked
Yet of every 100 hours of care planned at El Paso, 13 had nobody sent.
this quarter so far
A visit nobody worked leaves no record for the score to count. The state's reports count such a visit one member and one month at a time. Nothing adds it up into hours and dollars for one office.
Scroll both ways, or pinch to zoom.
the last of the quarter's visits can still be corrected
Jun 1the quarter starts
Jul 15today
Aug 31the quarter ends
Dec 4the grade is final, and a payer can open a review
No payer reviews a quarter before then. While the quarter runs, the state tells agencies to read their usage report at least once a month.
What this page does
The agency is modeled. The rules are real.
Scroll both ways, or pinch to zoom.
Each rule, and where it comes from
Behind “The state grades each contract, never an office.”
Source: Handbook 12040, Rev 24-1, effective Sep 12, 2024 and Handbook 11010, Rev 26-1, effective Jan 30, 2026 and Appendix G, v4.1, effective Feb 1, 2026
Behind “The score cannot see a visit nobody was sent to.”
Source: Handbook 12010, Rev 26-1, effective Jan 30, 2026 and CCSE 4674, Rev 25-3, effective Jun 1, 2025
Behind “A quarter's grade is final 95 days after the quarter ends.”
Source: Handbook 11000 and 11010, Rev 26-1, effective Jan 30, 2026
The stored sentences, word for word
For the visit nobody was sent to: what the quality review measured
the share of members whose attendant services were rendered in an amount equal to or greater than the amount approved
18 to 90 members in 100, never hours
Five managed care plans, research conducted September 2023 to August 2024, published in the state fiscal 2024 annual technical report by an external quality review organization. It counts members, not hours. It is statewide and per plan, never per agency and never per branch.
The state's report: every figure it prints
EVV Usage Report (the EVV Portal)
Link 2 · The rules
I learned these rules by observing and talking with the AI agents I directed, as we researched this industry and as we built the data for the modeled agency, where every rule had to work. Then I had other agents try to prove what we had learned wrong.
53 of 887 claims fell.
The rules themselves
The rule
A contract's grade has two parts: the share of visits clocked in, and the share of attempts to send a visit that the state does not turn back. 80 points pass.
Scroll both ways, or pinch to zoom.
At Tyler
The most a contract can get
What STAR+PLUS Plan A is heading to in each part, in our model
Added together, against the line
The rule
A visit can be typed in or corrected for 95 days. Then it locks.
Scroll both ways, or pinch to zoom.
At the agency
the quarter
so far
95 days
Jun 1the quarter's first day
today, Jul 15
Aug 31the quarter's last day
Dec 4its last visit locks, and a payer can review the quarter
The rule
When a contract finishes under the line, the health plan or the state, whichever pays for it, can send the agency a notice. The first asks for training. A second asks for a written plan to fix the problem. After that, the payer may propose ending the contract.
Scroll both ways, or pinch to zoom.
At Tyler
No notice
First notice
Training
within 20 business days of receiving the notice
Second notice
A written plan to fix the problem
within 10 business days of receiving the notice
After that
The payer may propose ending the contract
The rule
When an agency stops using a client's home phone for clocking in because it is not a true landline, it must make sure the attendant has another approved way to clock in.
Scroll both ways, or pinch to zoom.
At Tyler
The first half
Home phones no longer used
20 homes
done
The second half
Another way to clock in
0 devices ordered
not done
How the rules were checked
What the agents did, in order
Find the questions the first ten missed
3 agents, 8 questions
Each was told to think like a different expert (a director of operations, a compliance reviewer, a data scientist), and their questions were merged into the ones worth a report.
Read the sources
35 reports
over three rounds
One agent per question read the published sources and tagged every claim: confirmed in a source it fetched, not confirmed, or contradicted.
Try to prove each claim wrong
887 claims
One checker per report fetched the sources itself and tried to prove wrong every claim the report leaned on.
Keep what survived
746 held
The lead agent went through every check, let the check win wherever it differed from its report, and wrote up what held.
The lead agent: the AI session that planned the work, directed the other agents and answered to me.
Every checker answered in the same fixed form, one verdict for each claim, so a program could count them.
77 agents over three rounds. One checker per report. Each answer followed a fixed form and was saved in the workflow's record, and the counts below are tallied from those records.
What came out
Scroll both ways, or pinch to zoom.
A claim that could not be checked was marked that way. It was never guessed.
Counted by a program from each checker's answer in the workflow records.
What a caught mistake looks like
The claim
What the check found
What it changed
proved wrong
About 24 in 100 adults 65 and over, across the country, have a home phone line.
The report left out people who have both a mobile and a home line; added up from the source's own table, the figure is about 41 in 100.
Tyler's share of visits clocked in by home phone was lowered to fit it.
proved wrong
About 5 in 100 Texas adults have a home phone line.
The report left out the same group in every home phone figure; for Texas it is about 16 in 100.
The modeled agency was reshaped on the corrected figures, and home phones are a minority way to clock in at every branch.
proved wrong
A state appendix on visit records was the current one.
A later edition is in force, and its rule on time zones differs.
Each branch's visit times are kept in its own time zone, so El Paso's are in Mountain Time.
What this does not claim
Nobody who works in Texas home care has checked my reading of these rules. Every rule here was read from published sources by AI agents I directed, and checked by other agents set to prove it wrong. If you work in this field and see something wrong, that is the conversation I want.
The rules as they are stored, with their sources
Source: Handbook 12010, Rev 26-1, effective Jan 30, 2026Handbook 12020, Rev 26-1, effective Jan 30, 2026
Up to 60 points go to the share of visits clocked in and out electronically, and up to 40 to the share of attempts to send a visit that the state did not reject. 80 passes.
Source: Handbook 12010, Rev 26-1, effective Jan 30, 2026Handbook 12020, Rev 26-1, effective Jan 30, 2026
A visit counts as typed in by hand when the office types in the clock in, the clock out, or both. The score is rounded to a whole number, so 79.5 passes. A contract under 80 will be reviewed, and a payer may review every one or only a sample. We say "visit" where the handbook says "transaction".
Source: Handbook 9050, Rev 25-1, effective Mar 12, 2025Handbook 11000 and 11010, Rev 26-1, effective Jan 30, 2026
Each visit can be typed in or corrected for 95 days, then it locks.
Source: Handbook 9050, Rev 25-1, effective Mar 12, 2025Handbook 11000 and 11010, Rev 26-1, effective Jan 30, 2026
A quarter is final only once all its visits have locked. Once a visit locks, only its payer can reopen it. Since Jan 30, 2026, a payer may not reopen a visit when the attendant missed the clock and nothing stopped the office from keying it in time. No payer reviews a quarter before all its visits have locked. Tyler's quarter runs Jun 1 to Aug 31, and 44 of its 92 days have passed. It settles on Dec 4, and a payer's review can open from that day at the earliest.
Source: Handbook 11010, Rev 26-1, effective Jan 30, 2026 and UMCM 8.7.1, revision 2.7, effective Oct 1, 2024
A contract that finishes a quarter under 80 can draw a notice from its payer.
Source: Handbook 13000, Rev 24-1, effective Sep 12, 2024
A visit typed in by hand and accepted within 95 days is still paid.
Source: Handbook 11010, Rev 26-1, effective Jan 30, 2026 and UMCM 8.7.1, revision 2.7, effective Oct 1, 2024Handbook 13000, Rev 24-1, effective Sep 12, 2024
The state sends the notice for fee for service, and the health plan for its own contracts. The count runs over 24 months. A first notice asks for more EVV training within 20 business days. A second asks for a corrective action plan within 10 business days. A payer may propose ending the contract only after those steps and three failed quarters inside 24 months. If the training or the plan is missed, the payer may temporarily withhold Medicaid claims payments until requirements are met. Nobody publishes how long a payer takes to send a notice. Payers deny or take back a claim that does not match an accepted visit.
Source: Handbook 7030, Rev 26-1, effective Jan 30, 2026 and Handbook 7040, Rev 26-1, effective Jan 30, 2026
When the agency removes a phone that is not allowed, the handbook says to also 'make sure a valid landline or another approved clock in and clock out method is used.'
Source: Handbook 7030, Rev 26-1, effective Jan 30, 2026 and Handbook 7040, Rev 26-1, effective Jan 30, 2026
A landline used to clock in must be a true landline. When the attendant has no smartphone or tablet and the home has no usable landline, the home qualifies for an alternative device. The agency should order it within five business days of finding that, and the state's vendor has five business days to ship it. Phone signal is not part of the test. Until a device is in use, each visit is written down and typed in by hand within the 95 days.
The stored notes on these rules, word for word
How the claims were counted
node scripts/home-care-ops/receipts.ts --countWho the agents were
The modeling choices
Link 3 · The data
This is how I sourced, pooled and verified my information about this world, the industry and its data.
The agency is modeled. The rules are real.
Scroll both ways, or pinch to zoom.
713 agents in 101 workflows and 30 lead sessions, Sep 17 to 30 (692 of them in the workflows, 21 launched by the lead directly)
Every square is one agent, and every round mark is one of Ben's messages or one of the lead's checks, each on the day it happened.
Press any mark to open it. An agent shows what it was asked, what it did in its own words, and the pages it opened. Point at an agent in a workflow to light the rest of that workflow.
His own messages to the lead, each filed by what it did to the work.
Checks the lead made itself, before acting on what it had been told.
Brought information in from outside: the state's handbooks and rules, government data, published reports, real-world facts.
Put the information together: the rules written down, the modeled agency built, the programs that work out the figures, the sentences that state them.
Checked another agent's work, or tried to prove it wrong.
Built how the pages look and work: layout, colours, design references, tools, and keeping the plans in order.
A lane
His own messages to the lead, each filed by what it did to the work.
164 messages
A lane
Checks the lead made itself, before acting on what it had been told.
86 checks
A lane
Brought information in from outside: the state's handbooks and rules, government data, published reports, real-world facts.
47 agents
A lane
Put the information together: the rules written down, the modeled agency built, the programs that work out the figures, the sentences that state them.
233 agents
A lane
Checked another agent's work, or tried to prove it wrong.
318 agents
A lane
Built how the pages look and work: layout, colours, design references, tools, and keeping the plans in order.
115 agents
A message of Ben's
A message of Ben's
A message of Ben's
A check of the lead's
First a quick check of what main's instruction files look like, since the build session should start from main, not the cockpit branch.
A message of Ben's
A check of the lead's
I re-checked the usage-score formula against the handbook page itself (§12010, Revision 26-1, effective 2026-01-30). The 60/40 weights, rounding to a whole percent, and "below 80% gets reviewed" are all correct.
A message of Ben's
A check of the lead's
Before I hand back, one cheap check on the premise I flagged as unverified: what real agencies actually score against the 80% line.
A check of the lead's
A message of Ben's
A check of the lead's
The summarizer balked, but the PDF itself was saved locally. Reading it directly.
A check of the lead's
Determinism proven: two runs and an alternate output directory all produce byte-identical manifests.
A check of the lead's
48 checks, not 40: fixing my README rather than hard-coding a number that will drift.
A message of Ben's
A check of the lead's
Two of my quoted ranges were slightly off (Austin's repeat rate is 1.29, and Fort Worth's first-export rate dips to 95%). Correcting the doc to match the data.
A message of Ben's
A message of Ben's
A check of the lead's
All 26 figures resolve to exactly what the memo says.
A message of Ben's
A check of the lead's
Nothing moved: all checks pass, no git diff.
An agent
EVV report layout research
SourcedThe industry
Find out from the state's own documents whether the Texas EVV Usage Report's layout can be shown as it really is, and what the state calls each thing the dashboard names.
It pulled the two sample images out of a state job aid PDF, read every label on the state's usage report, and listed the state's own words for the dashboard's terms, saying plainly where it found none and what it could not reach.
HHSC's own "EVV Compliance Job Aid for Program Providers and FMSAs" (Revised: March 2024) embeds two sample images of the EVV Usage Report (program provider version and FMSA version), with every label and column legible. I extracted the images from the PDF and read them directly.
and 1 page not shown
6 web searches
A message of Ben's
A check of the lead's
The agency's first lock is July 17, and August 4 belongs only to the 24 members' visits. I corrected the memo, and the money tile now says "$9,640 locks within 30 days" instead.
A check of the lead's
My first tamper had a typo and never changed the file. Done properly, the reconciler catches a one-cent change, names the figure, and exits 1.
A check of the lead's
It's good. I ran the reconciliation myself and all 460 figures agree.
A message of Ben's
An agent
Incumbent home-care dashboards research
SourcedThe industry
Find, from public material only, what the vendor and state tools already show a Texas agency's office staff, and what practitioners complain about.
It opened 32 pages of vendor help sites, state notices, reviews and the Texas Register, found nothing public that projects the quarter, draws the 80 line or puts dollars on the problem, and found that branch comparison already exists.
Nothing I found projects forward, shows a target line, or puts dollars on EVV problems. "What to do next" exists only for a single record (fix this field, resubmit).
22 web searches
An agent
Dashboard craft and analogues research
Built
Find how the best tools show a current score, where it is heading and the target, give lessons from chart authorities, and pressure-test five choices in the lead's mock.
It opened 42 pages from Google Ads, Tableau, Linear, the SRE books and chart authorities, confirmed the chart form, and told the lead two of its choices were likely wrong.
Two things are likely wrong: six marks per row at 360px (over Few's own budget for a bullet graph), and the mark encoding (every verified analogue makes ACTUAL the solid mark and FORECAST the light/dotted one; you invert it).
29 web searches
A message of Ben's
A message of Ben's
A check of the lead's
Ground confirmed: all four commands pass, 460 figures agree, nothing moved under data/.
A check of the lead's
2,783 figures agree in SQL and pandas. Per the brief's trap list, watching the new checks fail once before trusting them:
An agent
review:contract
VerifiedThe data
Check whether a page that works nothing out could be built from the stored file as commit e86d1eb left it, and list what is missing.
It went through what the two pages must print or draw and found twelve values the stored file did not yet hold ready to read, two of them worse than gaps: nothing stored the order of the rows, and a status word and an at-risk flag could contradict each other.
Twelve values the exec read or the branch page has to print or draw are still missing, and two of them are contract defects rather than omissions: nothing in the file carries the row order, and projection.status and projection.at_risk can contradict each other
An agent
review:arithmetic
VerifiedThe data
Check the arithmetic in the code that projects each branch's quarter-end score and ranks it against the other branches (commit e86d1eb, step 6a), read-only, and report only real problems with evidence.
It ran the analysis at 30 rewound dates, found six of seven areas sound, and caught a crash, a self-check that fails on a state the code allows, a wrong line in the build brief, and two smaller gaps, one of them a case not yet in the data where the page would print a false sentence.
Area (5) is the one real problem: usageScore(pace) can throw and takes the whole run with it. Four findings below, one medium and three low, plus a checkable correction to the build brief.
An agent
review:recompute
VerifiedThe data
Recount the figures step 6a added, straight from the raw tables with a program of its own, and report only real problems with evidence.
Before reading the first program's queries it wrote its own count from the raw tables and matched all 113 figures it compared exactly, then raised six findings: figures the Python second program never checked, a wrong line in the build brief, a claim about the 'close' status that a later checker refuted, and three smaller faults.
I wrote my own read-only node:sqlite script against the raw tables (visit, visit_transaction, client, eligibility, quarter, meta) before reading 21-levers.sql, 22-landline-removals.sql or reconcile.py, and compared 113 figures against the committed analysis.json. EVERY ONE AGREED EXACTLY.
An agent
verify:arithmetic:currentPace: usageScore(pace) aborts the
VerifiedThe data
Try to refute the finding that the new pace figure crashes the whole analysis when the last four weeks hold no visits.
It could not refute it: it reproduced the crash on the real data, pinned the first date it happens (Aug 11), and showed the code before the commit did not crash.
I could not refute it; it reproduces exactly as described.
An agent
verify:recompute:statusOf labels a slice "close" when its
VerifiedThe data
Try to refute the finding that the status word 'close' contradicts the page design.
It reproduced every fact but showed the finding was wrong: the two-sided 'close' status is the recorded design, and the page's sentence code, already committed, names the close case correctly.
REFUTED. The finding's facts reproduce, but its claim of wrongness does not: the two-sided statusOf is the accepted, explicitly-ruled design, and the harm it predicts does not occur in the built page.
An agent
verify:recompute:The brief's fixture table says "close ne
VerifiedThe data
Try to refute the finding that the build brief wrongly says a 'close' state never occurs.
It ran the analysis at twelve dates, confirmed 'close' occurs on six of them, and showed the brief's row was already wrong when it was written.
So the row was checkable and wrong at writing time, not merely overtaken by a new field.
An agent
verify:recompute:Three new figures are not covered by rec
VerifiedThe data
Try to refute the finding that the Python second program does not check three new figures.
It planted five wrong values in a copy of the data and the Python program still said every figure agreed, proving the gap for the event week, and argued the day count did not belong in that program.
CONFIRMED for events[].week_start; the second and third bullets of the finding are weaker than stated.
An agent
build:7c
PooledThe data
Build the 'what it costs' and 'what to do' cards of the branch page from stored figures, working nothing out on the page.
It built both cards and caught two places where the specified numbers would have argued, one card apart and by a dollar.
I departed from the spec text in eight places, each for a stated reason; revert any you disagree with.
An agent
build:7d
PooledThe data
Build the three small weekly charts of the branch page (one per lever, against the other six branches) from stored figures, working nothing out on the page.
It built the three charts with a table view and a touch readout, and left out each chart's low point rather than work it out on the page.
Bottom line: step 7d is done, including the optional touch readout, and both files are clean.
An agent
build:7a
PooledThe data
Build the top of the branch page (the verdict sentences, branch chips, points meters and strips against the other six) from stored figures, working nothing out on the page.
It wrote the verdict for every case, tested it on today's data and four rewound states, and caught a rounding trap that would have printed 60.1 points out of 60.
I did not compute anything: I added a formatter, points(), that rounds a tie to the even tenth, so a kept and lost pair always prints to the term's weight (checked on all seven branches in today's data and all four fixtures).
An agent
build:7b
PooledThe data
Build the three 'why and who' panels of the branch page (rejections, clocking in, hours delivered) from stored figures, working nothing out on the page.
It built the three panels and changed several specified sentences where two figures on the page would have argued with each other.
A panel is open only when b.focus equals its Panel key. On a healthy branch all three are closed.
A check of the lead's
Useful findings. The biggest: with status now read from the range, "close" does occur on real days, so the hand-made fixture is unnecessary. Checking the dates the reviewer found.
An agent
critic:7a
Verified, also pooledThe data
Check the top-of-page section a builder just made against the data, the memo and the brief, fix what is wrong in place, and report.
It applied the builder's suggested rewrite so Tyler's page no longer shows three scores in three lines with nothing saying which is which, and confirmed the builder's rounding fix on all seven branches, finding two more where plain rounding would have printed a total over the maximum.
It now reads "Over the last 28 days it ran at 77. It has to run at 83 for the rest of the quarter.", taking 28 from analysis.method.pace_window_days (a lookup, not a computation).
An agent
critic:7b
Verified, also pooledThe data
Check the 'why and who' panels a builder just made against the data, the memo and the brief, fix what is wrong in place, and report.
It found the builder's claim that a field was missing was false, so a lock date no longer prints without its time left, and, building on the builder's own reasoning, flagged that three places read the stored last day to fix as the lock date, one day early.
Good section, one real defect, now fixed. The build is honest, computes nothing, and holds up on a phone.
An agent
critic:7c
Verified, also pooledThe data
Check the cost and what-to-do cards a builder just made against the data, the memo and the brief, fix what is wrong in place, and report.
It fixed wording faults only visible in the rendered words, such as a lock date marked '(a day)' when it meant today and a pronoun pointing at the wrong dollar figure, and left a $1 rounding gap for the lead.
What it missed were four things only visible by reading the rendered words rather than the data
An agent
critic:7d
Verified, also pooledThe data
Check the weekly charts a builder just made against the data, the memo and the brief, fix what is wrong in place, and report.
It found the charts' 32 weeks argued with the 'this quarter' figure beside them and added a caption naming the window, and fixed a legend that promised a mark not drawn.
The serious find was a number that could argue with another: the charts span 32 weeks across four quarters while the title row's figure counts only the six weeks from Jun 1, and nothing said so
A message of Ben's
A message of Ben's
A message of Ben's
A message of Ben's
A message of Ben's
A message of Ben's
An agent
Explanatory graphics references research
Built
Find checked references and research on showing a pass line, notes on charts and countable units such as '1 in 5' to readers who are not technical.
It opened 19 pages and brought back the physicians study behind writing '1 in 5' instead of '20%', with ten rules, and flagged three sources it could not read in full.
Given natural frequencies, 16 of 24 physicians solved the Bayesian problem; given conditional probabilities (same math), only 1 of 24 did. This is the strongest single number to justify "1 in 5" over "20%."
29 web searches
An agent
Light data-UI identity references
Built
Find public style guides, typefaces and examples for a distinct light-page look, with real numbers and links.
It opened 17 pages of government and newsroom style guides and font specimens and recommended Atkinson Hyperlegible Next or Public Sans, with colour contrasts it worked out.
Braille Institute of America, released 2025 as successor to the original Atkinson Hyperlegible. Variable, weight axis 200-800 (7 static instances ExtraLight-ExtraBold), upright + italic, and confirmed OpenType tnum (tabular numerals) + frac support.
40 web searches
An agent
Interactive chart tech research
Built
Research how to rebuild the dashboard's pictures with interaction, from the site's framework docs and the web, without changing any file.
It checked chart libraries, browser support and the site's installed packages, and recommended hand-drawn pictures plus the ready-made Base UI controls already installed.
18 web searches
A check of the lead's
All three threads are in. The code-stack report makes a few load-bearing claims about this repo, so I'm checking them before they go into the doc.
A check of the lead's
I ran the palette checker. The two colors score 12.0 for normal vision and 5.3 for red-green colorblindness, and both are below the threshold.
A check of the lead's
I had written "three steps per branch." The code shows only San Antonio has three, and Tyler's list varies. I fixed it in all three places.
A message of Ben's
A check of the lead's
R0 gate passes: generate, validate, analyze, reconcile all green, 2,870 figures agree, nothing under data/ moved.
An agent
build:r2
PooledThe data
Add two stored figures to the first program (what happened to every 100 visits and approved hours, and phrases like '1 in 5'), each with a Python twin, and prove a planted error fails.
It stored both on every branch with Python twins, proved nothing old moved, and flagged that 'typed in by hand' now named two different counts and that some phrases were loose.
The brief's own sentence reproduces exactly: San Antonio bounces back 845 of 4295 first exports = 0.1967, stored as "1 in 5"; the other six pooled are 411 of 20057 = 0.0205, stored as "1 in 50".
A message of Ben's
An agent
verify:additive-and-tamper
VerifiedThe data
Prove or break two claims: nothing old in the data file changed, and the Python second program would catch a wrong value in any new figure.
It proved nothing old moved, then made 28 small changes to copies of the file one at a time; nine changes to new figures got past the Python program, among them a one-digit change to San Antonio's bounced-back share behind its '1 in 5' line, and two agency totals had no check at all.
18 of 22 first-round tampers were caught AND named the exact figure (branch id, scope, group key, measure).
An agent
verify:raw-recompute
VerifiedThe data
Give an independent second opinion on the two new stored figures by recounting them from raw rows before reading the builder's queries.
Its own recount, with exact fractions, agreed with every new figure once it counted rejections the way the state's rule does; it found labels that could mislead a reader and confirmed the brief was wrong to treat every undelivered hour as never staffed.
RESULT: NO DISAGREEMENTS. Every new figure matches analysis.json.
An agent
verify:code-review
VerifiedThe data
Review the builder's uncommitted change for correctness bugs and rule-breaking.
It found every stored number correct but failed the change because 'typed in by hand' named two counts, worst on the first day of a quarter.
VERDICT is fail only because of finding 1: the stored numbers are all correct and reconciled, but typed_in_by_hand names two different populations under one word
An agent
cold:squint
Built
Look at a deliberately blurred screenshot of the mock and say, from shape and color alone, how many rows there are, which differ, and what stands out.
It could not tell whether the two failing rows and the five passing rows sat on one axis, which made the comparison unreadable at a glance.
The chart uses two different visual encodings for the same comparison, and that is the real problem.
An agent
cold:grayscale
Built
Look at a screenshot of the mock with all color removed and say whether trouble, groups and changes can still be told apart.
It got San Antonio's whole story without color but found the grid's two dark groups merge into one.
The business problem is stated plainly and I got it without color: on May 1, 24 clients changed insurance plans and nobody updated their setup, so every visit for those clients bounces back, and each resend counts against the grade again.
An agent
cold:exec-5s
VerifiedThe data
Look at one screenshot of the mock for five seconds as an operations manager with no context, then longer, and say what it is about, what is wrong and what is unclear.
It read the page right in five seconds and caught that every branch's arrow points down with nothing on the page saying so, and that the rows did not share one scale.
In five seconds I got: we're mostly fine, two branches are going to fail, and it's costing money.
An agent
cold:layperson
VerifiedThe data
Look at one screenshot of the mock as someone who never worked in healthcare and say what is going on, who is in trouble, and what the two dollar figures mean.
It read the page right but could not tell what the two big dollar figures measured, since neither carried a label.
It's checking whether the company's 7 branches will pass a state grading check on how home visits get "clocked in" (recorded), scored 0-100 with 80 as the passing line.
An agent
critic:design
Built
Critique the mock's design by eye against the brief and the rejected first build, and rank what to fix before the owner sees it.
It judged the mock clearly beats the rejected build, gave its pick on each of the three variants, and ranked five fixes led by Tyler's contradicting numbers, money cut off by the fold, and a grid that failed without color.
Verdict: this beats the rejected build decisively and it is worth showing Ben, but it has one credibility bug that an ops reader would catch in ten seconds and two hierarchy errors that cost it the first screen.
An agent
critic:mechanical
VerifiedThe data
Run the redesign brief's mechanical tests on the mock (desktop only), read-only, and report with evidence.
It ran the tests in a real browser and by searching the source; the mock passed the usual AI-dashboard tells but showed two figures for one phrase, keyboard stops hidden in closed panels, and an incomplete list of numbers worked out on the page.
The mock passes the tests that usually catch AI-made dashboards: zero non-ASCII bytes in all four files, so no em dash, en dash, middle dot or arrow glyph anywhere
An agent
cold:opened-row
VerifiedThe data
Look at a screenshot of San Antonio's opened row as a branch manager with no context and say why it is failing, what it costs, and what to do by when.
It got the cause, the cost and the steps, but could not tie the '1,308 visits' to anything else on the page and saw the number 80 used three ways.
A data-entry failure, not a care failure. 24 clients switched insurance May 1; their plan details were never updated in the system.
An agent
critic:facts
VerifiedThe data
Check every sentence and number the mock shows against the stored figures and the repo's rules document, as a Texas operations leader who knows the rules would.
It found every figure right to the cent and seven sentences an expert would reject, several of them about the state's rules.
Bottom line: the numbers are safe to show, the sentences are not. Fix the seven must-fix lines and Ben can judge the design on its merits instead of arguing with the copy.
An agent
fix:r2
Pooled, also verifiedThe data
Decide whether each of nine findings from three checkers is real by checking it, fix the real ones, and rerun every gate.
It confirmed and fixed seven findings and rejected two remedies with evidence, one of which would have failed a true figure.
No tolerance can both forgive a half-last-place rounding and catch a one-unit tamper, because they are the same distance. Rounding both sides and comparing exactly is the only fix that does both, so that is what I did.
A message of Ben's
A message of Ben's
A check of the lead's
Now the brief's factual corrections. Docs are intent, code and data are truth, and the data disagreed with it on El Paso.
A check of the lead's
I re-ran the gates myself: the Python second opinion agrees on 4,069 figures
A message of Ben's
A check of the lead's
I'll ground on the ledger and the mock first, so the questions I ask are about the actual pixels and not about my memory of them.
An agent
Capture reference board screenshots
Built
Capture and check screenshots of excellent products and articles for a reference board Ben would judge by eye.
It captured 27 screenshots in five groups and wrote their list, leaving alone three sites whose bot checks blocked it rather than getting around them, and a fourth whose pages had moved.
Done. 27 shots captured to docs/project3-reference-board/shots/, manifest.json written and verified to match exactly (no stray files, _test.jpg deleted, no git commands run).
7 web searches
A check of the lead's
Board is built and renders at your window size. I'm checking the rest of the captured shots for weak ones before handing it over.
A message of Ben's
A message of Ben's
An agent
R1 codebase and framework feasibility
Built
Study the site's own code and framework to say what the rebuild must work around, and write one research file.
It measured the first build's frame and the site's styles, found a width cap and a style trap, and noted that the second program had no twin for the projection.
1. Width cap. v1's main carries max-w-[1180px] (src/components/demo/home-care-ops/ui.tsx:24). Measured with shot.mjs, the content column is 1116 px at both 1440 and 2560, with 690 px empty on each side at 2560.
An agent
R3 how the references are built
Built
Measure what the references Ben kept are made of (type, colour, spacing, chart code) and write one research file with close-up shots.
It measured the live pages and read the source code of three of them, finding that the best data pages draw their own charts and that all the references but one keep their content to a narrow column on a wide screen.
Best finds: Plausible's chart is a hand-built SVG component (verified in their GitHub source, not a library), and OWID's real label-collision code is a simple "hide the lower-emphasis label on bounding-box overlap" algorithm with 4px/2px minimum spacing, no nudging.
5 web searches
An agent
R2 interaction and chart engineering
Built
Research how to build the dashboard's interactions and charts in the site's code, with browser support looked up, and write one research file.
It looked up browser support in MDN's data and 48 pages, set out how to light up linked parts without redrawing the page, and showed that a premise in its own instructions needed a caveat.
1. Linked highlighting needs zero re-renders: one stateless wrapper sets data-active/data-pinned, seven generated CSS selectors do the rest, and it lights up server-rendered SVG that ships no JS.
3 web searches
An agent
R4 pictogram construction and evidence
Built
Research how to draw a small family of pictograms by hand and how to use them as units of data, and write one research file.
It read Neurath's Isotype sources and icon-array studies, set drawing rules for ten signs, and warned that a person-shaped sign in the grid of 100 would read as a headcount.
(1) quantity is repetition of one same-size sign, never resizing, confirmed on three independent sources; (2) one base shape plus one added attribute distinguishes roles (Neurath's worker+coal rule)
26 web searches
An agent
R5 dissection page engineering
Built
Research how to build a case-study article whose pictures are live pieces of the dashboard, and write one research file.
It opened 28 pages on explorable articles, read the site's own data files, and found the gaps that would block four of the page's sections, including that the second program saved no file of its results.
2. Word-to-part linking is the one interaction with a controlled study (Zhi et al., 180 people): it erased the vertical article's comprehension penalty.
6 web searches
A check of the lead's
queries.json holds 25 queries (counted; scripts/home-care-ops/sql/ has 25 files). The docs that say 21 are stale.
A message of Ben's
An agent
draft:F
Built
Write appendix F of the v2 build brief (the pictograms) with real drawing data, from the research and the reference board.
It drew the signs with bounds-checked path data and found the stored data needs a sign the research had missed.
It gives real, bounds-checked SVG path data for every sign, splits the set into a people family (told apart by one attribute) and an event family (told apart by outer silhouette, because attributes vanish under blur)
An agent
draft:B
Built
Write appendix B of the v2 build brief (the token system: colours, type, spacing, motion) with real values, from the research and reference close-ups.
It wrote the colour and type tokens with a contrast script it ran, so every pair's contrast is computed, not asserted.
A contrast script was written and run: every token is inside sRGB and every text pair clears 4.5 to 1 and every mark 3 to 1 in both themes, with color-blind separation measured.
An agent
draft:D
PooledThe data
Write appendix D of the v2 build brief: the fields the first program must store for the v2 page, each with its twin in the Python second program.
It makes reconcile.py write reconcile.json, gives the projection ten honest invariants instead of a fake twin, and closes the old tolerance hole at twelve named call sites, plus a hole nobody had named: the whole grade has no second opinion today.
An agent
draft:G
Built
Write appendix G of the v2 build brief: the case study that explains how the numbers were worked out, with live pieces of the dashboard embedded.
It planned the case study page and found the research's recommended address would not scroll, so it moved the page out of the demo section.
It is built last, behind an optional homepage card field so it can be cut in one line.
An agent
draft:C
PooledThe data
Write appendix C of the v2 build brief: every module of the kit, with the stored paths each reads, resolved against the real data.
It listed 36 pieces of the kit and resolved 110 module-to-path rows with a script, finding that every drawing module waited on mark positions nobody stored.
Appendix C lists 36 pieces of the v2 kit: 16 composed modules grouped by the five questions, 4 shell modules, 16 primitives.
An agent
draft:E
Built
Write appendix E of the v2 build brief (interaction and chart techniques), checked against the installed packages.
It wrote the interaction techniques checked against the installed packages and found the tooltip library gives screen readers nothing.
One spine: a single data attribute (data-hck-key) on one scope element, a generated stylesheet, one stateless delegated listener, and a tiny external store.
An agent
draft:A
Built
Write appendix A of the v2 build brief (architecture and framework) from the research files, checked against the real code and installed packages.
It wrote the architecture plan, tested its lint rules on a sample file, and listed stale facts in the older documents and code.
Modules render on the server from a pure view file that the live island reuses, so twin and figure cannot drift; only a route page imports the data door.
A message of Ben's
An agent
verify:G
Built
Try to break appendix G by checking every claim against the code, the stored data and the installed packages, and report each defect with evidence.
It resolved every stored path the plan cites and found one blocker, a code sketch that would fail on its first render, and four contradictions a builder could not resolve.
What it gets wrong is the composition, not the facts. One blocker: page.tsx passes component functions into a client ContentsRail, which React refuses outright, and that is in the first slice.
An agent
verify:F
Built
Try to break appendix F by checking every claim against the code, the stored data and the installed packages, and report each defect with evidence.
It re-ran the drawing measurements from the path data and found three blockers, including a component name no other appendix used.
What fails is integration and self-consistency. F invents a component name and folder that neither A nor C uses, so a cold builder writes a <Pictogram> no module imports
An agent
verify:D
VerifiedThe data
Try to break appendix D by checking every claim against the code, the stored data and the installed packages, and report each defect with evidence.
It re-checked about 96 claims, ran the first program at a rewound date, and found the plan wrong about the 'too late' test state and about how many data checks there are.
Appendix D is unusually well grounded: I re-checked roughly 96 load-bearing claims and the overwhelming majority hold exactly.
An agent
verify:C
VerifiedThe data
Try to break appendix C by checking every claim against the code, the stored data and the installed packages, and report each defect with evidence.
It resolved all 110 module-to-path rows and found one blocker: an axis tick of 90 that no query produces and no stored position places, on the first picture.
97 YES all resolve with the stated sample values, 13 NO all genuinely absent, so "110 paths checked, 13 missing" is exact
An agent
revise:G
Built
Confirm each defect the checker reported on appendix G, then fix the appendix in place.
It fixed 14 defects and rejected three of its checker's points, among them a suggested fix that would have printed 4,356 where the page means 100.
The big change is subtraction. G no longer builds its own embedding machinery.
An agent
verify:A
Built
Try to break appendix A by checking every claim against the code, the stored data and the installed packages, and report each defect with evidence.
It checked 95 claims and proved in a scratch copy that the plan's test setup would turn the project's typecheck red on the first slice.
Appendix A is accurate on almost everything checkable: 95 load-bearing claims checked, and the file-and-line citations, the JSON paths, the Base UI and Tailwind package facts, the Next 16 doc references and the research attributions all hold.
An agent
verify:E
Built
Try to break appendix E by checking every claim against the code, the stored data and the installed packages, and report each defect with evidence.
It ran the plan's sketches through the repo's React and found a misspelled attribute that would have left every chart without a name for screen readers.
The defects are not in the library research; they are in the parts the drafter reasoned about rather than ran.
An agent
revise:C
Pooled, also verifiedThe data
Confirm each defect the checker reported on appendix C, then fix the appendix in place.
It fixed 15 defects, wrote every chart position into the plan as a figure to be stored rather than worked out on the page, and rejected three of its checker's corrections, one of them a wrong claim about a line citation.
The load-bearing correction: every chart position is a stored value, so the axis ticks, the pass line, the weekly points, the lock timeline's date axis and every bar width are owed fields, not page arithmetic.
An agent
revise:F
Built
Confirm each defect the checker reported on appendix F, then fix the appendix in place.
It fixed 16 defects and turned down five of the checker's claims or repairs, redrawing the caregiver's bag and the hour tick itself where the proposed fixes fell short.
Two drawings were redrawn to clear it: the caregiver's bag and the approved-hour tick.
An agent
revise:D
Pooled, also verifiedThe data
Confirm each defect the checker reported on appendix D, then fix the appendix in place.
It confirmed all 16 of its checker's defects, rejected none, and rewrote the plan so the 'too late' state carries a full series of fix-by days.
The slider's every stop is stored, including the too-late case, where the series is full and no stop passes.
A check of the lead's
An agent
verify:B
Built
Try to break appendix B by checking every claim against the code, the stored data and the installed packages, and report each defect with evidence.
It reproduced all 86 contrast ratios, but found every column width in the canvas table worked out on the wrong box, and measured a reference image to show the appendix, and the research it cited, had a proportion wrong.
The color work is the strongest part and I could not break it: I re-implemented section 5's algorithm and every one of the 86 reported contrast ratios reproduces to the second decimal
An agent
revise:A
Built
Confirm each defect the checker reported on appendix A, then fix the appendix in place.
It fixed 10 defects and disproved one of its checker's claims by compiling the stylesheet twice.
The variable is not emitted. I compiled the exact block twice: through tailwindcss 4.2.2's compile() API, and through the real @tailwindcss/postcss 4.2.2 over the appendix's exact two-line globals.css with kit.css in place.
An agent
revise:E
Built
Confirm each defect the checker reported on appendix E, then fix the appendix in place.
It fixed 14 defects and rejected two repairs and one count, cutting a grid animation that could not work rather than invent a feature.
Appendix E now specifies thirteen interaction techniques that ship four small client islands (link-scope, breakdown, fixed-by, term-trigger) plus one optional fifth.
An agent
revise:B
Built
Confirm each defect the checker reported on appendix B, then fix the appendix in place.
It fixed 14 defects and rejected several of its checker's claims, including the stored field it proposed for the group at issue.
Contrast is computed, not asserted: 98 pairs pass in both themes.
An agent
critic:consistency
Built
Read the seven revised appendices as a set and list every contradiction, gap and rule break between them.
It found 20 contradictions, 20 gaps and 7 rule breaks between the seven plans, among them places where the plans would have had the page work out a number itself.
The module file layout is specified two incompatible ways.
An agent
critic:cold-builder
Built
Try to start the build from the seven appendices alone, as a builder with no conversation, and list every question you would have to ask.
It found 25 questions, several blocking the very first file, and concluded the seven plans do not fit together without one document that picks winners.
The appendices are unusually strong individually and are not yet a brief.
An agent
harmonize:A
Built
Apply the spine's rulings to appendix A (architecture), editing only that file.
It worked the spine's rulings into the architecture plan and caught the spine miscounting the format helpers (eleven where the file has fifteen).
K1: A's folder per module, with C's kebab-case ids.
An agent
harmonize:F
Built
Apply the spine's rulings to appendix F (pictograms), editing only that file.
It worked the spine's rulings into the pictogram plan and moved the pictogram sheet into the workbench so no second copy of the drawings is kept.
K20: the hollow twin's stroke is 2 grid units.
An agent
harmonize:E
Built
Apply the spine's rulings to appendix E (interaction), editing only that file.
It worked the spine's rulings into the interaction plan and measured that the no-arithmetic grid ruling roughly doubles the page elements each grid needs.
K7 costs more than the spine seems to price. The old markup was one SVG with 100 <use> elements at about 110 nodes
An agent
harmonize:C
Built
Apply the spine's rulings to appendix C (the module list), editing only that file.
It worked the spine's rulings into the module list, giving every module its folder, its props and its slice.
K1: every module is now a folder, modules/<id>/ with reads.ts, view.tsx, index.tsx, table.tsx only when TWIN = true, live.tsx only when it ships a client island.
An agent
harmonize:G
Built
Apply the spine's rulings to appendix G (the case study), editing only that file.
It worked the spine's rulings into the case-study plan, including its new address, and flagged a gap the rulings left for narrow windows.
K3: the route is now stated as ruled, not recommended.
An agent
harmonize:B
Built
Apply the spine's rulings to appendix B (tokens), editing only that file.
It worked the spine's rulings into the token plan, turned the tokens into the three skins with measured contrast, and found the spine's accent colours disagree with its own named hue.
I measured the dim floor (the smallest opacity at which every mark still clears 3:1 on every legal ground) and got 0.950 ledger, 0.721 night, 0.920 field.
An agent
harmonize:D
PooledThe data
Apply the spine's rulings to appendix D (the stored-fields plan), editing only that file.
It worked the spine's rulings into the data plan, added ten new stored-field cards, removed the grid's index arithmetic, and found the plan's code line numbers had drifted.
Ten new cards, D17 to D26, each with path, type, plain definition, a derivation checked in the code today, its reconcile.py twin and its rounding rule
An agent
data:D-a
PooledThe data
Land the first data commit: move two blocks with no change in output, then store each branch row's mark positions and arrow direction with Python twins, and prove the check can fail.
It proved the move changed nothing, stored the new figures (4,153 agreeing), watched two planted errors fail, and found a false claim in the plan about the test states.
FAIL 'the row each branch draws', named 'SAT finish_pct 13.334 vs 13.333', exit 1. Restored by re-running analyze.
An agent
check:conformance
Built
Check, without editing, that the seven appendices now carry each of the spine's rulings and that a builder of the first slice would meet no conflicting instructions.
It began checking the rulings across the seven files but was stopped mid-run when a message was cancelled, and gave no answer.
Now let me verify the claimed-existing stored paths against the real data file.
An agent
check:conformance
Built
Check, without editing, that the seven appendices now carry each of the spine's rulings and that a builder of the first slice would meet no conflicting instructions.
It confirmed nearly all rulings landed and the data paths resolve, but found four contradictions in the first hour of slice S1, one of which would leave the kit's shared context empty, and stopped the foundation build.
What did not get resolved is the S1 component contract.
An agent
data:D-b+D-t
PooledThe data
It landed two data commits (7,425 figures agreeing), watched a tampered copy fail each new check, and caught the plan wrongly treating a date helper as deciding no word.
7,425 figures compared by reconcile.py, all agreeing (4,153 at the start of this slice; 7,380 after D-b; 7,425 after D-t).
An agent
fix:B+03
Built
Correct two stale rows in the design-tokens appendix and add a dated correction to one research note's measurement of a design reference.
It rewrote the stale rows about page widths after checking the architecture appendix, added the correction note, and confirmed two rulings were already stated correctly.
It now records that A was conformed on 2026-09-19 (verified: A:1060-1066 and A:1215-1216 give 44rem and 64rem citing K12; the only surviving 56rem/100rem in A is the provenance note at A:1255 about an old compile test)
An agent
build:S1-tokens
Built
Write the kit's colour values for three candidate looks, load its fonts, and build a script that checks every colour pair for contrast.
It wrote the colours and fonts and a checker that found no contrast failures across 252 colour pairs, keeping the stylesheet and the checker's table in step.
252 rows, 84 per skin, all three skins: 0 contrast failures, 0 out of gamut, 0 disagreements between kit.css and token-table.mjs.
An agent
build:S1-foundation
Built
Build the kit's foundation: the new page route, the doors that read the stored figures, linked highlighting, the honesty guards, and the workbench page.
It built the foundation in eight commits, showed its two test guards catch mistakes (a stored path that does not exist failed in all six data states, and a module that mis-declared its table failed the registry check), and measured one grid column at 87.66 pixels as the plan predicted.
And the stored-path guard catches what it is meant to: a path that does not exist failed in all six states, and a TWIN mismatch failed the registry test.
An agent
fix:A
Built
Apply rulings K31 to K38 and a checker's leftovers to the architecture appendix of the build plan, checking each new claim against the code.
It rewrote the appendix's code sketches and tables to match the rulings and checked its claims about a code library against the installed package.
Verified today in node_modules/@base-ui/react/tooltip/provider/TooltipProviderContext.js that it is a plain React.createContext, tagged [pkg].
An agent
fix:D
Built
Record in the build plan's data appendix that the latest data work landed (the second program agrees on 7,425 figures) and fold in four ways the code differs from the plan, checking each against the code.
It recorded the landing and the four differences in the data appendix, reproducing one of them (a Python library error) before writing it down.
Deviation 2 (checked, reproduced under pandas 3.0.6: ValueError: value_name (value) cannot match an element in the DataFrame columns): the axis-mark group measures 'grade' not 'value'
An agent
fix:E
Built
Apply rulings K31 to K34 to the interaction appendix and settle two leftover questions about where code lives.
It rewrote the linked-highlighting sketch and the file table of the interaction appendix to match the rulings.
K32 applied: KitScope takes the skin and never the theme; the theme is a fixed lookup (ledger light, night dark, field light)
An agent
review:S0+S1
Built
Review the first two build slices without editing: run every check, prove the old version untouched, check the code against the rulings, and report rendering defects in the workbench.
It ran every check but the production build, which it settled by reading the code (the second program agreed on 7,425 figures), showed on scratch copies that the stored-path test and the colour checker fail on bad input, and found one blocker and three defects in the workbench.
Both guards were proven to bite, on copies in the scratch directory, not taken on trust.
An agent
repair:S1
Built
Fix the reviewer's blocker and three defects in the kit and the workbench, then re-run the checks.
It fixed all four in four commits and measured the page width at each switch setting against the plan's predictions, which matched.
Measured at 2560, specimen canvas then chrome rail, one column: 1280 -> 1216/1216, colW 79.33 | 1440 -> 1360/1360, 87.66 | 1600 -> 1504/1504, 85.14 | no cap -> 2464/2464, 153.70. Appendix B section 7.3 predicts 79.33, 87.67, 85.14 and about 153.7.
A check of the lead's
The build passed. One thing in the route table needs a look: the workbench page /lab is listed in the production build. I'm checking whether it prerenders as a 404, as the builder claimed.
An agent
s2:docfix-A
Built
Correct four places where the architecture appendix no longer matched the code after the first review and repair, checking each against the code.
It rewrote the appendix's rows on the kit's text colour, border colour, width rules and look switches to match the code.
Section 5.2, the globals.css 183-188 row: rewritten to name both inherited properties, color as well as line-height, with the failure it causes (a bare li, td or span drawing the site's near-white on light paper)
An agent
s2:pass-rows
PooledThe data
Build the module that draws each branch's grade and where it is heading on a 70-to-100 track from stored positions, with a table twin and its sentences.
It built the grade-track module so every mark sits at a stored position and the table twin repeats the same stored values.
The button's aria-label is the whole row as one sentence containing those exact stored values, and the table twin carries them again as addressable text; without this each row is announced three times.
An agent
s2:signs
Built
Build the sign component and draw exactly three signs to the pictogram rules: a visit, the dollar-with-padlock unit, and the agency's monogram.
It drew the three signs on the pictogram grid and proved in the browser that they take their colour and line weight from the page.
F's last unverified claim is now settled and F should be told: currentColor AND stroke-width both reach the cloned artwork through <use> in Chrome.
An agent
s2:primitives
Built
Build the first four building blocks (panel frame, big figure, status pill, named scale) and the file that holds the kit's words.
It built the four pieces and their styles, shot them in all three looks, and fixed two defects it found itself.
Three skins were shot at 1600 and read by me, and the two defects I found (stacked lead figures colliding at 64px on a 60px line, and a unit that had to be checked for cap flush) are fixed.
An agent
s2:hero
Built
Compose the hero sample panel and the three-across view from the pieces three builders had just committed.
It composed one San Antonio panel from the new pieces and built the views that show it in all three looks, drawing no row of signs for a count stored only as approximate.
A row of exact signs is a claim about an exact count, so a loose k-in-n fit is said in the sentence and not pictured.
An agent
loop:critic-1
VerifiedThe data
As fresh eyes who may not read the code, shoot the sample panel in all three looks and list its defects beside the reference close-ups.
It listed 30 defects and checked every drawn position and printed figure against the expected values, all of which matched.
Checked every stored position in the plot against the values (zone 0-33.333%, line at 33.333%, dot 26.667% = 78, arrowhead 13.333% = 74; Austin dot 66.667% = 90, head 60% = 88)
An agent
loop:repair-1
Built
Fix the first critic's defects at their root and reject any that do not survive measurement.
It fixed most of the 30 defects in one commit and rejected six in whole or in part, one because the critic had measured the wrong column.
Measured in the page: Field's lit step computes lab(95.22 -1.25 +3.13), a warm cream, against Ledger's lab(96.28 -0.50 -1.81). The critic sampled the wrong column.
An agent
loop:critic-2
VerifiedThe data
As fresh eyes, shoot the repaired sample panel again and list its remaining defects beside the reference close-ups.
It listed 26 defects; the worst was that the '80 passes' line never moved, so in one test state the picture put Tyler's 77 on the passing side of 80.
This is a number arguing with its picture, in the panel that is the whole gate.
An agent
loop:repair-2
Built
Fix the second critic's defects at their root and reject any that do not survive measurement.
It fixed 20 of the 26 defects, including moving the '80 passes' line to where 80 falls, and rejected six with measurements.
Twenty are fixed here, at their root; six did not survive being measured and are rejected with the numbers, in the notes below.
An agent
gate:S2
Built
Close slice S2: build the gate 1 page for Ben, run every mechanical check, and search the kit for banned style habits.
It ran every check green (the first program rewrote the stored file byte for byte and the second agreed on 7,425 figures) and built the gate 1 page.
exit 0; wrote analysis.json and queries.json (25 queries); git status shows data/home-care-ops unchanged, so the rewrite was byte for byte identical
An agent
rev1:build
Built
Build the lead's revision 1 of the sample panel: bands, a horizontal named scale, figures on the track's marks, and the sign row in all three looks.
It built revision 1 in eight commits; it could not reproduce a 24-pixel overflow the order named, and fixed its cause anyway.
I could not reproduce a 24px figure at any ruled width: measured at 1280, 1440 and uncapped, and scanned every element of both plots against the figure's own box.
An agent
rev1:critic
VerifiedThe data
As fresh eyes, shoot revision 1 in all three looks and list its defects beside the reference close-ups.
It listed 23 defects, among them a row of signs placed beside a sentence it did not picture and a close branch whose 'points short' line vanished.
The reader's eye reads the houses as a picture of 74 out of 100.
An agent
rev1:repair
Built
Fix the revision-1 critic's defects at their root and reject any that do not survive measurement.
It fixed most of the rest in two commits, left a few for the lead, and rejected five findings on measurement, among them a 'broken' pass line that was only picture compression.
Five of the twenty-three findings were REJECTED on measurement and are written up at the head of kit.css part 17 with the numbers
An agent
rev1:gate
Built
Close revision 1: retake the gate 1 picture, restart the judging room if needed, and run every mechanical check.
It ran every check green (the second program again agreed on 7,425 figures), retook the gate picture, and corrected two lines of the gate page that still described the old panel.
index.html needed two corrections, not zero: it still described the pre-revision panel.
A message of Ben's
An agent
freeze:field-guide
Built
Make Field guide the only look, delete the losing looks and their fonts, and fix three rough spots the first gate left.
It froze the look in four commits and left one rough spot unfixed, because fixing it would have meant the page comparing two figures.
Dropping the label would have meant comparing two data values to decide whether the reader is shown a word.
An agent
pieces:grid-100
PooledThe data
Build the grid of a hundred (visits and care hours) in both colourings for Ben to choose between, from stored cell strings.
It built the grid module from stored strings, gave both colourings equal treatment so Ben's choice was fair, and declined to draw a pointer the stored data could not place.
Both colourings share the lit-object treatment on the group at issue, so the only difference Ben judges is whether the OTHER groups carry colour. Giving the object to (b) alone would have rigged the comparison toward the recommendation.
An agent
pieces:signs-sheet
Built
Draw the full sign set and the sheet Ben judges it on at gate 2, and make the judging room stop sending him to port 3000.
It drew the twelve signs, found that one drawing in the pictogram appendix did not match its own stated measurements, and made the judging room serve the workbench itself.
The phone is drawn at F's fallback width, and F's printed fallback path is wrong.
An agent
pieces:answers
PooledThe data
Build the first screen's two answers, the verdict sentence and the money pair, from stored figures and sentences.
It built both modules and declined to print a status word above the verdict, since no single status is stored for seven branches.
M2's blind-spot note supersedes earnedLine's money sentence with the stored hours figures, per the work order.
An agent
pieces:shell
Built
Build the first screen's shell: the header, the branch strip, the question blocks, the glossary of the state's terms, the table twin and the source footer.
It built the shell and its words, including the glossary pairing the page's plain words with the state's terms, and left out a quarter progress bar whose position is not stored.
Every status hue is spoken for (red means failing and nothing else) and the accent means 'you can touch this', so neither was honest for 'this is not the real file'.
An agent
compose:v2
Built
Compose the first screen at /demo/home-care-ops/v2 from the finished pieces, following the lead's composition.
It composed the first screen, fixed a duplicate-id fault inside the grid module, and left one answer slot empty rather than invent a sentence nobody had ruled.
Question 3's answer={null} is mine: the two money figures and the two sentences under them are the answer, no authored sentence for that slot exists in labels.ts, and inventing one would be page copy nobody has ruled.
A check of the lead's
I searched the repo, the ledger, every doc and my memory files for "airport", and the word appears nowhere
An agent
loop:critic-1
VerifiedThe data
As fresh eyes, shoot /v2 at two window sizes and in two fixture states and list its defects beside the reference close-ups and the earlier mock.
It listed 35 defects and confirmed every expected figure on the page (five of seven pass, San Antonio 78 heading to 74, $97,331, $29,045) was present and correct.
Expected values: five of seven pass, San Antonio 78 to 74, Tyler 79 to 76, 80 passes, $97,331 collectable, $29,045 gone, Jul 29 and Aug 14, El Paso 77 of every 100 - ALL present and correct in the live fixture
An agent
loop:repair-1
Built
Fix the first critic's defects in the first screen at their root and reject any that do not survive measurement.
It fixed 29 of 35 defects in one commit, rejected six with measurements, and reported a figure that can print on the wrong side of the pass line.
Twenty-nine are fixed at the root; six did not survive being measured and are written up in kit.css part 23 so nobody spends a second round on them.
An agent
loop:critic-2
VerifiedThe data
As fresh eyes, shoot the repaired /v2 again at two window sizes and two fixture states and list the remaining defects.
It listed 31 defects, among them that on a day when every branch passes, the loudest thing on the chart was an empty block labelled 'failing'.
Verified every expected figure in the live state: five of seven pass, San Antonio 78 today to 74, Tyler 79 to 76, 80 passes, $97,331 collectable, $29,045 gone for good, Jul 29 and Aug 14, El Paso 77 of every 100 approved care hours. No figure contradicts its headline
An agent
loop:repair-2
Built
Fix the second critic's defects in the first screen at their root and reject any that do not survive measurement.
It fixed the other 22 of 31 defects (one only in part) and rejected nine with measurements, two of them because a fix would have needed the page to compare stored figures.
Nine of the thirty-one defects were rejected after being measured, each written up with its measurement in kit.css part 24 so a third round does not revisit them
An agent
gate:S3+S4
Built
Close slices S3 and S4: build the gate 2 page and the judging room's cards, and run every mechanical check.
It ran every check green (the stored file came out byte for byte and the second program agreed on 7,425 figures) and built the gate 2 page.
wrote analysis.json and queries.json; git status on data/ stayed empty, so the stored file is byte for byte unchanged
An agent
Fill the v2 rail, add width switch
PooledThe data
Build El Paso's blind-spot tile as a real page module from stored figures, close the empty space beside the chart, and add a screen-width switch for development only.
It built the tile so it picks its branch from a stored field, never by comparing figures, closed the row, added the switch, and passed every check in three commits.
El Paso's blind spot is now a real kit module in the rail under the money pair, in its own QuestionBlock. The branch is selected from the stored branches[].focus field (analyze.ts's out_of_pack flag, which reconcile.py recomputes), never by comparing figures.
A message of Ben's
A message of Ben's
A check of the lead's
Gate 2 is up and the room is already running. Let me verify the data layer actually holds what the next slice needs before I commit agents to it.
A message of Ben's
An agent
S5a:data
Pooled, also verifiedThe data
Store, for every branch, which group of visits its problem lives in, add an independent twin of it to the second program, and run both.
It stored the new field and its twin, the second program agreed on 7,439 figures with none disagreeing, and it fixed a test file that would have said 'nothing needs fixing' over a close grade.
A grid of DELIVERED visits cannot hold hours that were never delivered (unstaffed visits and visits cut short are not in it at any size), so there is no honest group to name.
An agent
research:voice
Sourced, also verifiedThe industry
Check every term in the page's glossary and its other domain words against the state's own wording and the words Texas agency staff use.
It found three glossary cells wrong, among them British 'per cent' and two of the page's own definitions printed as if they were the state's terms.
Nine of the twelve glossary rows are right. Three are not.
7 web searches
An agent
research:blindspot
Sourced, also verifiedThe industry
Check whether the state's score really cannot see hours that were never delivered, what agencies call this, and whether 77 of 100 delivered is unusual.
It confirmed the score cannot see hours nobody worked, found the state's own report field for them, and warned that, against older figures from outside Texas, 77 of 100 is poor but ordinary, so the contrast is the story, not the level.
every input to the EVV Usage Score is a property of a visit transaction, so an hour nobody worked is not a zero in the denominator, it is absent from the denominator.
and 1 page not shown
15 web searches
An agent
research:fault
Sourced, also verifiedThe industry
Using the state's EVV rules and related sources, try to prove the dashboard's main story wrong: does a rejection for a client's changed health plan count against the agency, and who must fix it?
It found the state's own rejection code for a payer mismatch in the list the handbook says counts against the score, and that every fix step sits in the agency's own system, so the story survived.
The dashboard's story survives, and it gets sharper. The payer mismatch is a real Aggregator edit, Ex00031C, and it sits in Appendix M
7 web searches
An agent
research:fix
Sourced, also verifiedThe industry
Check the page's three-step fix (correct the plan, stop resending, resend oldest first) against how Texas EVV rejections are really worked.
It found step 2 told the reader to stop a resend loop that no rule or vendor guide it found describes, and step 3's 'resend' wrong, since correcting a visit is what sends it and each stuck visit must be fixed one by one.
Steps 1 and 3 hold up and get sharper, more recognizable names. Step 2 is wrong as worded and must be reworded.
and 1 page not shown
8 web searches
An agent
research:clock
Sourced, also verifiedThe industry
Check whether the page's 'safe date' for the grade and its earlier 'money' date are two real, different deadlines under Texas rules.
It confirmed two different 95-day clocks, found the money date is driven by a billing rule the page had not found, and found 'money permanently lost' too strong.
Both clocks are 95 days and both run from a date of service, yet they start on different dates.
12 web searches
An agent
S5:step-list
PooledThe data
Build the plan for a failing branch: three steps with stored figures, a drawer of numbers behind each step, and two dates.
It built the plan from stored fields and left the comparing of its two dates to the reader; a later reader found its lock date came from the whole branch while its words spoke of the 24 clients.
The money's lock date is sooner than the grade's safe date on every branch in every one of the six states, and no field says so; comparing days_until_first_lock with days_until_must_be_fixed_by is exactly the arithmetic the honesty rule bans.
An agent
S5:blind-spot-block
PooledThe data
Build the block that sets El Paso's graded visits beside its authorized care hours, reusing the grid of a hundred.
It built the block to read the stored group rather than the grid's derived one, so El Paso's visits grid would not light nine squares on a panel saying nothing was wrong.
Both figures are stored, the two windows are different and the sentences name them ("today" against "this quarter"), so I printed it honestly rather than hiding the block on a zero, which would have been the page deciding.
An agent
S5:cause-card
PooledThe data
Build the cause card that states a failing branch's cause from stored fields: the plan mismatch for San Antonio, the clock setup for Tyler.
It built the card and caught that its specification paired two different stored measures as if comparable, which would have put a false comparison on the page.
Nothing the module needed was missing, so nothing was computed on the page and there are no disable comments.
An agent
S5:outcome-set
PooledThe data
Build the module showing where a branch's quarter finishes if fixed today, if fixed by the safe date, or not fixed.
It built the three outcomes on either side of a real pass line and refused to place them on a to-scale track, since their positions are not stored.
I did NOT compare a figure to 80. Each row's side and word are selected from a stored enum
An agent
S5:weekly-cliff
PooledThe data
Build the week-by-week chart for a branch's room from stored line and band paths, with its table twin.
It built the chart from stored paths and positions and refused to print axis end numbers, which would have meant multiplying a stored value on the page.
Nothing is derived: the line and the band are stored strings, every left and top is a stored percent, and the two printed counts are stored integers.
A message of Ben's
An agent
audit:page-sentences
VerifiedThe industry
Turn the five researchers' findings into a ranked list of exact changes to the words on the page, and say what must be left alone.
It found eight sentences that must change, the worst that the page said the state grades branches when it grades contracts, plus eleven to reword and fourteen to leave alone.
The page survives, but not untouched. Eight sentences are wrong. The worst is not a domain subtlety: the v2 explainer tells the reader the state grades branches.
An agent
comp:b-b-cause-chain
PooledThe data
Build design comp B for 'Why is San Antonio failing?': the cause drawn as a chain from the plan change to the lost points.
It built comp B and showed the state's rejection code on screen as a quotation from the state's manual, labelled as not a value in the dataset.
Two things it prints are authored citations rather than stored values, and both are visibly attributed on screen rather than passed off as data
An agent
comp:a-a-points-split
PooledThe data
Build design comp A for 'Why is San Antonio failing?': the grade drawn as its two parts (60 and 40 points) with the broken part lit.
It built comp A from stored figures (55.13 and 19.18 points, 74.31 in all, printed as 74) and asked for stored display strings rather than pad figures on the page.
In particular I did NOT need a pooled "other six" figure: I drew each of the six branches' own stored term as its own mark, in headline.branch_order, so nothing is averaged.
An agent
comp:c-c-break-point
PooledThe data
Build design comp C for 'Why is San Antonio failing?': the weekly line that breaks at the plan change, beside the other six branches.
It built comp C and refused to print axis numbers, because turning a stored 0.7 into '70 of 100' would have been the page doing arithmetic.
Two, both of which I refused to compute and left off the picture rather than fake
An agent
S5:compose
Built
Compose the branch rooms under the strip, one per branch, from the new modules, opened by picking a branch, and run the checks.
It composed the seven rooms and took four rulings itself, among them pointing the grid at the stored field after Austin's room lit fourteen red squares under 'Nothing in Austin needs fixing'.
The stored field is null on exactly those five, which is analyze.ts saying no group is at issue. Red means failing and nothing else.
An agent
why:gate-page
Built
Build the dev-only page that stacks the three comps full width for Ben, run the checks, take the screenshots, and list it in the judging room.
It built the page and the screenshots and, asked which comp was strongest, argued for C because its picture itself carries the cause.
C is the only one whose PICTURE carries the cause.
A check of the lead's
I was wrong, and it propagated into the build. Fixing the source of the error first.
An agent
correct:copy
PooledThe industry
Apply the grounding audit's must-change and should-change lists to the page's words without moving any figure.
It rewrote all eight wrong sentences and the eleven to reword (one only in part), computing no figure and leaving one billing question for Ben.
Nothing was computed and nothing was hardcoded.
An agent
read:fresh-ops-reader
VerifiedThe data
As a fresh reader posing as a Texas operations director, read the rendered page and report every figure, claim and sum that does not hold.
It found the corrected page still did not add up: figures about the same stuck visits disagreed by a factor of two, and a lock date contradicted the drawer above it.
A director multiplies: 1,308 x 5 = 6,540. Or 1,308 x 4.65 = 6,082. The page says 2,956, which is 2.26 each. Nothing on the page can be made to give 2,956.
A check of the lead's
Both confirmed exactly. May 1 + 95 days is Aug 4, and that date is stored on the cohort object the page isn't using.
An agent
fix:scope-rule
Pooled, also verifiedThe data
Apply one rule across the rooms: a panel about the 24 clients takes every figure from their own stored record, a panel about the branch from the branch's; fix what crosses.
It moved the crossing figures onto the right records, wrote the rule into the ground-truth file, and found two crossings the reader had missed.
Total: seven crossings. Four the reader named (two page-level, two it could not run down), plus the cause card's 4.65 row which it named as a voice flag and which was also a scope crossing, plus two the audit found alone (MoneyPair's subtitle, and M14's leading figure).
An agent
recheck:fresh-ops-reader
VerifiedThe data
As a fresh reader, re-run a director's arithmetic on four branches with every drawer open, and check which stored record each figure comes from.
It confirmed the two earlier failures were fixed on San Antonio, found the lock-date one still live on Tyler, and found a new one: 24 clients were said to hold 1,308 stuck visits right above a panel saying the whole branch held 781.
24 clients are said to hold 1,308 stuck visits; the whole branch is said to hold 781. A part cannot be 167% of its whole.
An agent
sweep:classes
PooledThe data
Fix each remaining kind of error (unnamed windows, lock dates, Tyler's two counts, the duplicate grid) on all seven branches, not only where it was found.
It fixed seven kinds of error across all seven branches, read every branch's rendered text to confirm, and wrote the unshown failing contract up as an open question for Ben.
Method: fixed the class at the one selector that decides it, then read the rendered text of every branch and state rather than trusting the code.
An agent
final:fresh-ops-reader
VerifiedThe data
As the final fresh reader, check every sum, pair of counts and date on all seven branches before the page goes in front of Ben.
It found every sum on the page now comes out, and one stopper: each weekly chart's 'now' figure belonged to a different week than its own table's last row.
Every sum on the page comes out. I ran them all and found no arithmetic error.
A message of Ben's
A message of Ben's
A check of the lead's
Slices are made. Now I look at the page myself, San Antonio's room and the why comps, before any reader does.
An agent
read:sat-ops
VerifiedThe data
Read the dashboard cold from screenshots, as a Texas home-care operations director with a calculator, San Antonio open; say why it fails and list what does not add up.
It explained San Antonio's failure in its own words, ran 22 sums (16 came out) and raised 15 problems; the first 10 were checked against the stored figures.
An agent
read:sat-design
VerifiedThe data
Read the dashboard cold as a senior product designer judging whether it has a visual system of its own, San Antonio open, and list what does not add up.
It judged the page its own thing (4 of 5), explained San Antonio's failure, ran 19 sums (15 came out) and raised 14 problems; the first 10 were checked.
It is a one-line change, it is above the fold in the right rail, and it is the only arithmetic on the page that fails in a single glance with no cross-referencing.
An agent
read:elp-ops
VerifiedThe data
Read the dashboard cold as a Texas home-care operations director with a calculator, El Paso open; say what is wrong there and list what does not add up.
It found El Paso passing but losing hours nobody was sent to work, ran 17 sums (12 came out) and raised 15 problems; the first 10 were checked.
El Paso is not failing. It grades 93 and passes, and the page says plainly that nothing in the way El Paso records its visits is costing it the grade
An agent
read:aus-general
VerifiedThe data
Read the dashboard cold as a general reader, Austin open; say whether opening a passing branch was worth it and list what does not add up.
It found Austin passing and the failing branches' cause never named on its screens, ran 10 sums (6 came out) and raised 13 problems; the first 10 were checked.
is the sentence a reader will carry out of this page, and its subject is a pronoun with no antecedent.
An agent
VerifiedThe data
It explained San Antonio's failure in its own words, ran 19 sums (14 came out) and raised 16 problems; the first 10 were checked.
Right now the page's two most prominent objects say 74 while the track underneath says 78, and the only thing reconciling them is a grey caption.
An agent
read:tyl-ops
VerifiedThe data
Read the dashboard cold as a Texas home-care operations director with a calculator, Tyler open; say why it fails and list what does not add up.
It explained Tyler's failure but refused to call the plan workable, since no step dealt with the 20 removed landlines; it ran 22 sums (17 came out) and raised 17 problems, the first 10 checked.
As it stands I could execute this plan in full and Tyler would keep bleeding manual visits, because the larger half of the cause was never assigned.
An agent
read:tyl-general
VerifiedThe data
Read the dashboard cold as a general reader, Tyler open; say why Tyler fails and list what does not add up.
It explained Tyler's failure in plain words, ran 14 sums (10 came out) and raised 19 problems; the first 10 were checked.
The care happened; only the record of it had to be typed in afterwards by hand.
An agent
read:sat-general
VerifiedThe data
Read the dashboard cold as an educated general reader with no healthcare or data background, San Antonio open; say why it fails and list what does not add up.
It explained San Antonio's failure in plain words, ran 16 sums (10 came out) and raised 18 problems; the first 10 were checked.
As the reader that headline is aimed at, my first and last question is whether the company clears 80, and I finish the page unable to say.
An agent
read:sat-calc
VerifiedThe data
Read the dashboard cold as a data analyst checking every total, subset, date and scope against every other figure, San Antonio open, and report every sum.
It ran 30 sums (25 came out) and raised 18 problems; the first 10 were checked and 8 confirmed as real defects.
By the page's own 35-day rule and its own Jul 14 cut-off, neither Jun 8 nor Jun 9 is the right cut-off.
An agent
verify:aus-general
VerifiedThe data
Test a reader's confusion over the weekly chart's grey band, keyed only as 'the other six branches', and its unexplained dive.
It recomputed the band for all 33 weeks: it is the lowest-to-highest range of the other six, a definition the page wrote but showed only to screen readers; the dive is San Antonio.
So the band is the FULL RANGE, lowest to highest, not a middle and not an average.
An agent
verify:aus-general
VerifiedThe data
Test a reader's claim that the settling footnote calls the week of Jun 8 both unsettled and settled, against the stored figures.
It confirmed the footnote wraps a stored day (Jun 9) in words about weeks, and showed the reader's other half wrong: the 'week of Jun 8' clause is correct.
Verified against the stored data; the defect is real but NARROWER than the reader states. One of the reader's two halves is wrong.
An agent
verify:aus-general
VerifiedThe industry
Test a reader's claim that the page says the state grades the agency, then never shows an agency grade.
It found the agency's projected grade (86) and three contract grades stored but never shown, showed an average of the branches gives 87, and found one contract heading to 79, under the line.
The reader is right, and the gap is wider than they could see.
An agent
verify:aus-general
VerifiedThe data
Test a reader's claim that the header's 'Wed Jul 15, day 44 of the state's 92-day quarter' is off by a day.
It showed both day counts are counted to the data's last day (Jul 14) but printed beside Jul 15, so the label, not the figure, is off by one.
Both 44 and 48 are anchored to data_through (Jul 14), not to as_of (Jul 15).
An agent
VerifiedThe data
It confirmed a stored day (Jun 9) sits in words about weeks, and that the footer's 'Weeks up to Jun 9' is false by the page's own 35-day rule.
No figure is wrong; the words wrap a DAY-level stored value in WEEK-level language
An agent
VerifiedThe data
It found the third group stored (703 hours missed or cut short), traced the drop to code that picks two of the three groups, and noted the tile shows whichever branch is open.
THE READER'S ARITHMETIC IS CORRECT AND THE THIRD BUCKET EXISTS, STORED AND NAMED.
An agent
verify:aus-general
VerifiedThe data
Test a reader's claim that Austin's 102 unpaid visits are fewer than the about 110 its grid of a hundred implies.
It showed the true counts nest correctly (94 this quarter inside 102 over a longer window) and traced the 110 to rounded squares the footnote invites you to scale.
The reader's two figures are both right, and the reader's arithmetic is the arithmetic the page's own footnote hands them.
An agent
verify:aus-general
VerifiedThe data
Test a reader's claim that the weekly chart's '97 of 100 now' and the Jul 13 date under it belong to different weeks.
It confirmed the 97 is the week of Jul 6 printed over Jul 13, and warned the fix must not call it 'settled', because at Austin that week is not.
So the printed 97 is Jul 6 and the axis label under it is Jul 13, whose stored figure is 99.
An agent
verify:aus-general
VerifiedThe data
Test a reader's claim that El Paso's tile ('77 of every 100 authorized hours: 15 had no attendant') leaves 8 hours in 100 unnamed.
It found all three hour groups stored (77, 15 and 8 make 100) and judged that the tile's colon promises a full account the sentence does not give.
Every figure in the sentence is stored and correct; the gap is in the words, not the data.
An agent
verify:aus-general
VerifiedThe data
Test a reader's confusion that the weekly chart has two date labels and starts six months before the quarter without saying so.
It ruled the two-label axis deliberate (the table gives every week) but confirmed the chart never names its seven-month window, breaking the build's own rule that each figure names its window.
SPLIT VERDICT: the reader is wrong on half and right on the half that matters.
An agent
VerifiedThe data
It showed the stored day counts run to Jul 14, not Jul 15, and that the page's own footer already says the figures count visits through Jul 14.
The label pairs the as-of date with a day count that belongs to a different day.
An agent
verify:aus-general
VerifiedThe data
Test a reader's confusion that Austin's money panel ($5,306, 102 visits, $163) mixes three time windows.
It showed all three figures come from one stored set and one window, so the reader's premise was wrong, but found the panel's sentence names no window while the row under it says 'not only this quarter'.
All three figures in that strip come from ONE object and ONE window; the reader's premise that they are three scopes is false.
An agent
verify:aus-general
VerifiedThe data
Test a reader's confusion that every branch's arrow points down from today's grade to a lower finish, with no word why.
It found the two ends of each arrow measured differently (grade posted so far against grade after settling) and that, compared like for like, six of seven branches are flat.
There is no agency-wide slide. Every visible arrow on the first screen is the settling lag drawn as motion.
An agent
VerifiedThe data
It confirmed the footnote prints only San Antonio's total under both grids though the other six's 20,569 is stored, and found it hits exactly San Antonio and Tyler, the two failing branches.
An agent
VerifiedThe data
It confirmed the 80 is the week of Jul 6 printed as 'now' over Jul 13, with the true week given only to screen readers.
The figure is right; the label under it is wrong.
An agent
VerifiedThe data
It confirmed the 5 hand-typed visits are drawn like the 76 clean ones, ruled the one-row second grid deliberate, and found the shared footnote gives the second grid the wrong total.
So no figure is wrong and both hundreds add. The reader's claim is not an arithmetic claim.
An agent
VerifiedThe data
It showed 74 is the projected finish and 78 today's grade, found the mid-page banner prints 74 with no tense word, and found the drop from 78 to 74 is almost all settling, not a trend.
but one panel does print a projected figure with no tense word on it.
An agent
VerifiedThe data
That is friction, not falsehood, and it is capped by the fact that the correcting words sit in the same sentence as 1,308.
An agent
VerifiedThe data
It showed the seven cards print each branch's projected finish (San Antonio's 74), not today's grade, with 'heading to' hidden from sighted readers, while the words needed already existed in the table view.
So 74 is the projected quarter finish and 78 is the quarter-to-date score as of Jul 14. Nothing is wrong and nothing is mixed.
An agent
verify:tyl-general
VerifiedThe data
Test a reader's claim that the footer's 'Weeks up to Jun 9' names a mid-week day and breaks the 35-day rule printed beside it.
It confirmed the footer prints a day where it says weeks, that the stored per-week flags give the week of Jun 1, and that the second program does not recheck the settled date or the rule behind it.
That is a WEEK-level sentence printing a DAY-level value.
An agent
verify:tyl-general
VerifiedThe data
Test a reader's claim that Tyler has more hand-entered visits (566 + 164 = 730) than visits with no way to clock in (637).
It showed 637 is one of four stored causes of hand-entered visits (731 in all), so nothing contradicts, but found the page never says other causes exist, the build's own notes held the same wrong idea, and the 637 called the 7 attendants' visits also counts an eighth who left on Jul 11.
The 637 is ONE ROW of that 731, 87% of it, not the set the 566 and 164 come out of.
An agent
verify:tyl-general
VerifiedThe data
Test a reader's claim that 'Wed Jul 15, day 44 of the state's 92-day quarter' is off by one.
It confirmed 44 is the number of the data's last day (Jul 14) printed as today's, a third case of the same labelling fault as the two already known.
The reader's arithmetic is right and the diagnosis is exact.
An agent
verify:tyl-general
VerifiedThe data
Test a reader's claim that El Paso's hours figures leave 8 in 100 unexplained (77 + 15 = 92).
It confirmed the tile opens a hundred and closes only 92 of it, the one such sentence on the page that does not add up, and that the words for the missing group already existed.
The tile is the only place on the page where a hundred is opened and not closed.
An agent
VerifiedThe data
It confirmed both: the 8 is a stored group the tile never prints, and the tile sits among agency figures, so El Paso's 77 can be read as the agency's (whose own figure is 90).
Both halves check out, and the second half is a hundred that does not close on the first screen.
An agent
verify:sat-ops
VerifiedThe data
Test a reader's claim that the settling footnote contradicts itself and that by the 35-day rule the last final week is the week of Jun 1.
It confirmed it from the stored weekly flags, the same on all seven branches (the last final week is the week of Jun 1), noted the reader's own date sum was a day off but immaterial, and found the second program did not recheck the weekly settled flags.
CONFIRMED against the stored data, and the stored data agrees with the reader's conclusion more strongly than his own arithmetic did.
An agent
verify:tyl-general
VerifiedThe data
Test a reader's claim that Tyler's '60 of 100 now' sits over the Jul 13 date and the line rises past it.
It confirmed the 60 is the week of Jul 6 over a Jul 13 date whose stored figure is 66, and showed the chart's left end teaches the opposite reading to its right end.
A reader who learns the rule at the left end and applies it at the right end lands on the wrong week.
An agent
verify:sat-ops
VerifiedThe industry
Test a reader's claim that San Antonio's grade cannot be rebuilt from the page's 60/40 split and its grid of 100 visits (89 or 78, never 74).
It rebuilt the grade from the stored counts, showed the 40-point half is counted per send, not per visit, and found the page's defining sentence says visits, an 11-point trap.
That is an 11-point invited error, and it is the whole of the reader's gap.
An agent
verify:tyl-general
VerifiedThe data
Test a reader's claim that Tyler's footnote ('1,879 visits this quarter') sits under two grids, one of them six branches pooled.
It confirmed the pooled grid stands for 23,046 visits, stored and checked but never printed, so a reader multiplying gets figures 12 times too small.
The reader is right, and the figure they say is missing is stored, reconciled, and simply not printed.
An agent
verify:tyl-general
VerifiedThe data
Test a reader's confusion that Tyler's two 'landlines were removed' labels overlap so the later one reads as the first.
It confirmed the dates and their order are right in the data, and found the two near-identical labels overlap with the later one on top, inviting the reading that the fall began in May.
Every stored value behind the two callouts is correct, and the reader's *second* reading is the right one. What is wrong is the arrangement.
An agent
verify:tyl-general
VerifiedThe data
Test a reader's confusion over what the grey band on Tyler's chart means and why it stops short of the line.
It checked the band against the six branches' own figures for two weeks and in the program that stores it (lowest to highest), found its definition shown only to screen readers, and showed by the pixels that the band does not stop short.
So the edges are literally the best and the worst branch that week, i.e. the first of the reader's three guesses, and they had no way to know it.
An agent
verify:tyl-general
VerifiedThe data
Test a reader's claim that Tyler's money panel stacks '191' visits over '$193' like a copy error, and that $193 does not fit a first lock in 23 days.
It showed every figure consistent ($193 is the 4 visits that lock between day 23 and day 30) and found the layout invites the misread: a bare 191 over $193, and an 'Of it' that sits right under the visit count though it means the dollar total.
Nothing is a copy error and nothing fails to add.
An agent
verify:tyl-general
VerifiedThe data
Test a reader's claim that Tyler's finish panel draws 85 and 80 as identical boxes on what looks like a scale.
It found every figure right but the panel dressed as a number line that places nothing by value, which the code itself admits, so the 80 box sits clear of the '80 passes' rule.
No wrong figure, no wrong label, no wrong window, no invited sum. So the reader did not catch an arithmetic defect.
An agent
verify:sat-ops
VerifiedThe data
Test a reader's claim that El Paso's hours panel explains 15 of its 23 missing hours in 100 and never names the other 8.
It confirmed the 8 is stored and named only in El Paso's own room, so with San Antonio open the page never closes the hundred.
The defect is an incomplete partition presented as if complete.
An agent
verify:sat-ops
VerifiedThe industry
Test a reader's claim that the page asks 'Will we pass?' and defines the grade per contract, yet shows no agency or contract grade.
It found the contract grades stored but never shown on the page, checked the project's rules file ('Not each branch'), and showed one contract heading to 79 while an average of the branches reads a comfortable 87.
VERIFIED, and the data makes it worse than the reader could see.
An agent
verify:sat-ops
VerifiedThe data
Test a reader's confusion that the other-six grid shows one legend row, so it seems they have no hand-typed visits.
It confirmed both figures (5 here, 11 there) are stored, found the filter dropping three rows deliberate but hiding the comparison that backs the page's own story, and separately found the one footnote under both grids gives the second grid San Antonio's total.
The visible picture says strictly less than its own accessible description.
An agent
verify:sat-ops
VerifiedThe data
Test a reader's confusion that the plan names run backwards (the new plan is MCO 1, the stale one MCO 2).
It checked the generator and confirmed the 24 clients really moved from plan 2 to plan 1, and found the page never says which code to key in, so the direction must be inferred.
It is a panel whose wording leaves the direction to be inferred, in a place where the ordinals actively argue the other way.
An agent
verify:sat-design
VerifiedThe data
Test a reader's claim that El Paso's hours tile leaves 8 hours in 100 without a name.
It confirmed the third stored group closes the hundred and that the tile's colon, meaning 'namely', turns a chosen part into what reads as the whole account.
So there is no missing 8. The third group exists, is stored, and is the whole of the remainder: 77 + 15 + 8 = 100.
An agent
verify:sat-ops
VerifiedThe data
Test a reader's claim that the header's 'day 44' is off by one and disagrees with '48 days left'.
It confirmed the off-by-one in the label, showed the reader's second charge wrong (44 and 48 agree), and named it the third case of one labelling fault.
The reader caught a real off-by-one in the label; the reader's second charge (that the two figures disagree) is wrong, and the distinction matters for the fix.
An agent
verify:sat-design
VerifiedThe data
Test a reader's claim that 'day 44' is off by one against the page's own quarter.
It confirmed 44 counts days of data through Jul 14 while the caption glues it to Jul 15: two conventions inside one small caption, in the header shown whichever branch is open.
The label is off by one; the figure is not.
An agent
verify:sat-design
VerifiedThe data
Test a reader's claim that the settling footnote calls the same week both unsettled and settled.
It confirmed the footnote glues a week-level date and a day-level date into one sentence, checked that the stored weekly flags are the same on all seven branches, and showed why the chart's hatching agrees with the first half.
CONFIRMED, and it is the known L2.
An agent
verify:sat-ops
VerifiedThe data
Test a reader's claim that the grid's scale note names one total under two grids and does not square with the 1,308 stuck visits.
It split the claim: the footnote under both grids is a real defect, while the 1,308 against 784 half was already closed and is the reader's error.
TWO CLAIMS IN ONE. The first is a real, unrecorded defect on all seven branches.
An agent
verify:sat-ops
VerifiedThe data
Test a reader's confusion that the grid's legend names four kinds of visit but the picture draws three.
It confirmed by eye and in the code that hand-typed and clean visits draw alike, worse on quiet branches (97 identical cells at Austin), hiding San Antonio's real strength on the 60-point half.
THE READER IS RIGHT, AND UNDERSTATED IT. Every figure checks; the picture is what fails.
An agent
verify:sat-design
VerifiedThe data
Test a reader's claim that the page footer repeats the day-level cutoff as 'Weeks up to Jun 9'.
It confirmed no week in the data ends on a Tuesday, so 'Weeks up to Jun 9' is impossible on its face, and that it sweeps in a week the chart marks as still arriving.
The claim is correct, and it is the page-wide restatement of the known L2 defect at a second, independent render site.
An agent
verify:sat-ops
VerifiedThe data
Test a reader's claim that '80 of 100 now' over Jul 13 is either an untrusted part week or a different week than labelled.
It showed the 80 is the last full week (Jul 6), not the part week, and that the footnote itself creates the dilemma; it also noted that at San Antonio the quarter's figure and the last full week's both print as 80, a coincidence that masks the fault there.
The reader is right, and the second horn of their dilemma is the true one: 80 is NOT the week of Jul 13, and nothing in the visible layer says so.
An agent
verify:sat-design
VerifiedThe data
Test a reader's confusion over which week '80 of 100 now' belongs to.
It confirmed every visible cue points the 80 at Jul 13 while it is Jul 6, a week given only to screen readers.
Every visible cue points the 80 at Jul 13 and the 80 is Jul 6.
An agent
verify:sat-design
VerifiedThe data
Test a reader's claim that the grid footnote calls its marks squares when they are houses.
It read the drawing code and the pixels: visit marks are houses with a door (the hours marks really are squares), so the sentence that teaches the picture points at the wrong thing.
CONFIRMED, and narrower and worse than the reader put it.
An agent
verify:sat-design
VerifiedThe data
Test a reader's claim that the card shows San Antonio's 74 with no label while the track below shows 78 today.
It found the reader wrong about the cards (a visible line above them says where they are heading) but found the projected grade mislabelled twice elsewhere: 'El Paso grades 93' in the present tense, and a mid-page banner with no tense word.
THE READER IS WRONG ABOUT THE CARD STRIP ITSELF.
An agent
verify:sat-design
VerifiedThe data
Test a reader's claim that one grid footnote is attached to two grids it cannot both describe.
It confirmed the footnote's 4,356 is San Antonio's alone, the pooled 20,569 is stored but never printed, and a reader's 44 stuck visits at the other six should be 103.
An agent
verify:sat-design
VerifiedThe data
Test a reader's confusion that the grid draws only grey, red and hatched houses for four legend categories, so the 5 hand-typed visits cannot be found.
It confirmed clean and hand-typed visits render as one unbroken grey block, found the same on all branches but Tyler, and ruled out merging legend rows because 81 is not stored.
The reader's observation is factually correct and the defect is in the picture, not the figures.
An agent
verify:sat-general
VerifiedThe data
Test a reader's claim that the chart's 'now' figure sits inside the zone the chart marks as unsettled and names no week.
It confirmed all three parts, one of them not on record before: at San Antonio the week the 80 belongs to sits inside the hatched still-settling zone, so the chart prints as 'now' a figure its own table marks still settling.
VERIFIED AGAINST THE STORED FILE. The reader's figures are all real and their reading is the one the picture invites.
An agent
verify:sat-general
VerifiedThe data
Test a reader's claim that the header's day count and days-left count use different conventions.
It showed the reader's reasoning wrong (both counts pivot on Jul 14) but the label real: the page states Jul 15 is day 44 when it is day 45.
So the reader's stated mechanism is wrong. There is ONE convention here, consistently applied
An agent
verify:sat-design
VerifiedThe data
Test a reader's confusion that the other-six grid prints one legend row and so reads only as a contrast.
It judged the one-row legend a deliberate choice with no wrong figure, and found the real defect beside it: the single footnote gives the lower grid San Antonio's total.
The reader's literal observations all hold, and none of them is a figure defect; adversarial verification of the same panel found a real one beside it.
An agent
verify:sat-general
VerifiedThe data
Test a reader's claim that the sidebar's 'Weeks up to Jun 9' names a mid-week day.
It confirmed Jun 9 is a Tuesday in a dataset of Monday-start weeks and that the stored per-week flags give the week of Jun 1, the reader's own answer.
VERIFIED REAL. The reader is right on both halves of the claim
An agent
verify:sat-general
VerifiedThe data
Test a reader's claim that the settling footnote contradicts itself and both its halves break the 35-day rule.
It found the first half correct (the week of Jun 8 is still coming back), the second a day value in week words, and the footer's 'Weeks up to Jun 9' a plain wrong label.
The claim is half right, and the half it gets wrong is the half it was loudest about.
An agent
verify:sat-general
VerifiedThe data
Test a reader's claim that El Paso's missing authorized hours do not account for themselves.
It confirmed the 8 is stored and rechecked by the second program, printed one click away in El Paso's room, and that the colon turns a chosen part into an implied full account.
The reader's arithmetic is exactly right, and the missing 8 is a stored value the page prints one click away.
An agent
verify:sat-general
VerifiedThe data
Test a reader's claim that one footnote's 4,356 is shared by two grids that cannot share it.
It confirmed the footnote is built from the branch's own total only, on the two branches whose view draws two grids (San Antonio and Tyler), while at San Antonio the table one click away prints the other six's counts at about five times what that footnote implies.
The module declares it and never prints it, which is the bug.
An agent
verify:sat-general
VerifiedThe data
Test a reader's claim that the grid's 18 in 100 (about 784 visits) and the plan's 1,308 stuck visits cannot both be right.
It queried the project's own database read-only and showed the figures fit (1,308 since May 1, 777 of them inside the quarter), with both windows stated in the sentences, so the reader was wrong.
The two figures are consistent, and the page names both windows in the sentences that carry them.
An agent
verify:sat-general
VerifiedThe data
Test a complaint from a reader who knew nothing about the project: San Antonio's card says 74 while the chart just below says 'today 78'. Rule on it against the stored figures.
It found both numbers stored and right, showed in the page's code that the words 'heading to' were hidden from sight so the big 74 read as today's grade, and warned that a draft would add a third San Antonio grade (74.31) printed as 'to date'.
THE INVERSION IS THE WHOLE DEFECT. The one temporal word drawn anywhere in this region is attached to the figure that needs it least (78, which appears once, small and grey) and withheld from the figure that carries all the visual weight
An agent
verify:sat-general
VerifiedThe industry
Test a reader's claim that San Antonio's grade cannot be rebuilt from the grid and the 60/40 split.
It rebuilt all seven branches' grades the reader's way (five within a point, Tyler about 2.5 off, San Antonio about 11 off), traced the gap to the 40-point half counting every send, repeats included, and checked the rule against the project's rules file.
An agent
verify:sat-general
VerifiedThe data
Test a reader's complaint that El Paso's 'What it needs' cell says only 'The grade cannot see this.' where the other rows say what each branch needs.
It confirmed every figure was right and agreed with the reader that the cell answered a question the column did not ask, with a 'this' that pointed at a panel further down the page, out of view.
So the reader is right on the facts and right about the effect. No figure, window or sum is wrong. What is wrong is that the cell answers a question the column did not ask, and its deictic "this" has no antecedent in the panel it sits in.
An agent
verify:tyl-ops
VerifiedThe data
Test a reader's claim that the header's 'Wed Jul 15, day 44 of the state's 92-day quarter' is a day off.
It confirmed the header printed the number of days of data (through Jul 14) as today's place in the quarter, and warned against changing the stored 44, which the forecast depends on.
Jun 1 is day 1, so Wed Jul 15 is day 45 of the quarter. Day 44 is Tue Jul 14. The sentence binds today's date to the day count of the cutoff, and that sentence is simply false.
An agent
verify:tyl-ops
VerifiedThe data
Test a reader's claim that the footer's 'Settled through: Weeks up to Jun 9' breaks the page's own 35-day rule, which makes the week of Jun 1 the last final week.
It confirmed the stored cutoff is a single day, Jun 9, dressed in words about weeks, and that the page's own stored week flags agree with the reader: the week of Jun 1.
No figure is wrong. A day-level stored cutoff is wearing week-level words, in the one panel whose job is to certify how every figure on the page was made.
An agent
verify:tyl-ops
VerifiedThe data
Test a reader's claim that El Paso's panel accounts for only 92 of every 100 authorized hours (77 delivered, 15 with nobody to send).
It showed the stored hours split into three groups that make 100 (77, 15, and 8 missed or cut short) and that the sentence named only two, though El Paso's own page already printed all three.
NO FIGURE ON THE PAGE IS WRONG. 77, 15 and 93 are stored and correct. The defect is an incomplete partition presented as a breakdown.
An agent
verify:tyl-ops
VerifiedThe data
Test a reader's claim that Tyler's weekly chart prints '60 of 100 now' above the Jul 13 date while its own line rises past it.
It confirmed from the stored chart shapes that the 60 is the week of Jul 6 printed as 'now' beside the Jul 13 date, while the table one click away says Jul 13 was 66.
The reader's optical claim is confirmed to the pixel by the stored geometry, and it is the already-recorded defect L1 seen from the outside.
An agent
verify:tyl-ops
VerifiedThe data
Test a reader's claim that the grid for the other six branches lists four groups but draws its 4 bounced-back visits in the same grey as the clean ones, under a one-row legend.
It counted the squares in the screenshot and read the code: all four groups are stored, but the picture drew two of them the same grey and printed a single legend row.
The 4 bounced_back squares (row 10, positions 5-8) are pixel-identical to the 86 recorded_cleanly squares.
An agent
verify:elp-ops
VerifiedThe data
Test a reader's claim that the 93 printed over El Paso's visit picture cannot be rebuilt from that picture with the page's own 60/40 formula.
It found the 93 is the projected finish captioned as the grade for the quarter over a picture of visits so far whose own grade is 94, and that the picture's 'bounced back' square counts open visits, not the share of sends that bounced, which misleads badly at San Antonio.
Reconstructing SAT's waffle the way this reader reconstructed El Paso's gives 89.1 on a panel printing 74, on the branch whose whole argument is "6 points short of the line."
An agent
verify:tyl-ops
VerifiedThe data
Test a reader's claim that one footnote ('The hundred stands for 1,879 visits') sits under two grids whose hundreds differ.
It confirmed the footnote gave only Tyler's 1,879 visits while the other six branches' grid stands for 23,046, so a reader scaling that grid would be off twelvefold.
Not a reader error and not merely an invited misreading: as positioned, the sentence makes a false statement of fact about the grid it sits under.
An agent
verify:elp-ops
VerifiedThe data
Test a reader's claim that the footer's 'Weeks up to Jun 9' breaks the page's own 35-day rule, and that the honest answer is the week of Jun 1.
It confirmed the reader on every step and recommended storing the settled week with an independent recount; the lead had already put that repair on its list, and the next build made it.
No figure is wrong. No flag is wrong. The window is wrong only in the words: the page claims a settled horizon one week longer than its own data says, and says it twice, once page-wide in the provenance box a skeptic reads first.
An agent
verify:tyl-ops
VerifiedThe data
Test a reader's note that two different 77s sit inches apart on the first screen, El Paso's '77/100' hours figure and El Paso's 77 percent share of the unstaffed money, which the reader had already checked were not one number restated.
It showed the two 77s are a coincidence (one is a share of El Paso's own hours, the other a share of the agency's money that the page never prints) and found nothing to fix.
Both figures are stored, both are right, and the reader's own conclusion ("they are not the same figure restated") is correct. The collision is arithmetic coincidence, not a page defect.
An agent
verify:elp-ops
VerifiedThe data
Test a reader's claim that the weekly chart's note calls the week of Jun 8 unsettled and then says everything through Jun 9 has settled.
It confirmed the contradiction on all seven branches: the stored cutoff is a day, Jun 9, inside the week of Jun 8, which the page's own flags mark unsettled, and the footer repeats it as 'Weeks up to Jun 9'.
So: every figure is right, both flags are right, and the prose states a week-level claim using a day-level boundary.
An agent
verify:elp-ops
VerifiedThe data
Test a reader's claim that the header's 'Wed Jul 15, day 44' is wrong because Jul 15 is day 45.
It confirmed 44 counts the days of data through Jul 14, and warned that printing 'day 45' would break the sum with the '48 days left' beside it (45 + 48 = 93).
It takes a count of ELAPSED days and prints it as today's ORDINAL position. The date beside it is today; the count beside it belongs to yesterday.
An agent
verify:tyl-ops
VerifiedThe data
Test a reader's complaint that Tyler's money panel stacks a bare 191 over '$193', so the two read as one quantity.
It found both figures stored and right, agreed with the reader that a missing unit on 191 and an 'Of it' pointing at the wrong line invited the misread, and checked all seven branches: only Tyler's two figures collide.
A reader who takes "it" at its nearest antecedent and 191 as dollars gets $193 out of $191: a subset bigger than its set, which is exactly class C1 from the 2026-09-20 sweep
An agent
verify:tyl-ops
VerifiedThe data
Test a reader's complaint that the '77/100' in the 'What is the grade not showing us?' panel carries no branch name, so it could be the agency's rate or El Paso's.
It found every figure in the panel is El Paso's and right and that the agency's own figure (90) is stored but printed nowhere, and agreed with the reader that a label with no owner invites reading 77 as the agency's.
So a reader who takes 77 as the agency's rate is off by 13 squares on a page-level fact, and there is no figure anywhere on the page to catch them at it.
An agent
verify:tyl-ops
VerifiedThe data
Test a reader's complaint that the seven-branch chart prints the word 'today' only on its top row, leaving two bare numbers on the others.
It found the word deliberately hidden on six rows to avoid overprinting the pass line, judged that the reader read it right but had to work it out, and found the chart's own title hidden from sight too.
SO: the reader inferred correctly and had to. On the two rows where the pair is close and the stake is highest, Tyler 79/76 and El Paso 94/93, a glance without following the arrow admits the inverted reading
An agent
verify:elp-ops
VerifiedThe data
Test a reader's claim that El Paso's 67 unpaid visits is smaller than the roughly 71 its own visit picture implies for this quarter alone.
It showed every figure right and the nesting sound (13 and 49 this quarter, within 14 and 53 overall); the 71 came from scaling rounded squares, which the picture invites because it hides the counts.
EVERY FIGURE IS RIGHT, AND THE NESTING HOLDS CAUSE BY CAUSE. The reader's ~71 exists nowhere in the data; it is their own back-computation from rounded squares.
An agent
verify:elp-ops
VerifiedThe data
Test a reader's suspicion that El Paso holding $22,484 of the $29,045 in unstaffed care forces an odd shape on the other six branches that the page does not show.
It confirmed the reader's arithmetic to the hour ($6,561 and 383 hours at the same $17.13 an hour) and that the stored split for the other six (1 in every 100 hours with nobody to send, 8 missed or cut short) is the shape the reader deduced; the page printed only '91 in every 100'.
The reader's arithmetic is not just plausible, it is exactly right to the hour, and the stored data confirms every step.
An agent
verify:elp-ops
VerifiedThe data
Test a reader's claim that El Paso's weekly chart figure (98) cannot be reconciled with the visit picture below it.
It found the 98 right but counted over the 2,317 visits sent, while the picture's hundred is the 2,366 visits delivered and its 'bounced back' square counts only visits still open; the chart never said which hundred it used.
THE FIGURE IS RIGHT. THE RECONCILIATION IS IMPOSSIBLE FROM THE PAGE.
An agent
verify:elp-ops
VerifiedThe data
Test a reader's claim that El Paso's two grids draw fewer groups than their legends list: 98 identical visit marks, and only the 15 red hours stand out.
It traced the one-grey colouring rule in the code, counted the same collapse across all seven branches, and found the failing case missing from the sample sheet Ben was about to rule on.
The page tells the reader this picture is how the visits were recorded and hands them a four-row legend; the picture draws one state.
An agent
verify:sat-calc
VerifiedThe data
Test a calculator-minded reader's claim that El Paso's hours sentence explains a shortfall of 23 hours in 100 with only 15.
It confirmed the sentence named two of three stored groups on a panel that shows on every branch's page, and that the gap is 8 or 9 in the live data and in all five test states.
the blind-spot branch is El Paso in every state and delivered + at_issue is 92, 91, 92, 91, 92 - never 100.
An agent
verify:elp-ops
VerifiedThe data
Test a reader's complaint that the grey 'other six branches' band on El Paso's weekly chart collapses two-thirds of the way across with no label or explanation.
It found the plunge is San Antonio's own line after its May 1 plan change, not the settling window the reader guessed, and that the one sentence defining the band existed only for screen readers.
The pooled low IS San Antonio's line, and it broke because of San Antonio's stored event
An agent
verify:sat-calc
VerifiedThe data
Test a calculator-minded reader's claim that by the page's own 35-day rule the last final week is the week of Jun 1, not Jun 8 or Jun 9.
It confirmed the reader's rule is the program's own, corrected one step (Jun 8 is rightly the first unsettled week), and noted that the per-week settled flag had no independent recount.
The reader's arithmetic is the generator's own arithmetic, and the stored data already agrees with them.
An agent
verify:sat-calc
VerifiedThe data
Test a calculator-minded reader's claim that San Antonio's '80 of 100 now' is drawn from inside the stretch the chart shades as unsettled and names no week.
It confirmed the 80 is the week of Jul 6, printed as 'now' beside the Jul 13 date, and that this week sits inside the chart's own 'still settling' shading, which the note never mentions.
The headline is 12.5 points of the plot box INSIDE the band the same picture tags "still settling".
An agent
verify:sat-calc
VerifiedThe data
Test a calculator-minded reader's claim that the settling note calls the week of Jun 8 unsettled and then says everything through Jun 9 has settled.
It confirmed the contradiction, found the right answer (the week of Jun 1) already stored as a flag on every weekly row, and pointed to the same day-in-a-week sentence in the footer, which the lead and the readers had already raised.
The page's rule and the page's number disagree in adjacent sentences. That is on the page, not in the reader.
An agent
verify:elp-ops
VerifiedThe data
Test a reader's claim that El Paso's '98 of 100 now' sits over the Jul 13 date and the 'still settling' shading while the note says the Jul 13 point can still move.
It confirmed the 98 is the week of Jul 6 and that the week was named only for screen readers; at El Paso both weeks read 98 so nothing visibly breaks, but at Tyler the same label is six points off.
So: figure right, week right, choice of week right, visible attribution wrong.
An agent
verify:sat-calc
VerifiedThe data
Test a calculator-minded reader's claim that the header dates the read one day off its own counter: day 44 is Jul 14, the data cutoff.
It confirmed 44 counts the days of data through Jul 14 and that the header contradicted the page's own footer by one day.
So the header contradicts the page's own footer by one day, in the first line above the fold, while the footer states the truth at the very bottom.
An agent
verify:sat-calc
VerifiedThe data
Test a calculator-minded reader's claim that San Antonio's most repeated number, 74, is a forecast printed as a fact, with the current 78 just below.
It found future-tense words printed above the cards ('will finish', 'Where all seven are heading'), so it judged the claim too strong there, but found the branch page's banner printed the forecast with no qualifier at all, and it flagged a draft that would print 74 as the grade 'to date'.
That one instance is the reader's grievance in its honest form: a lone unqualified forecast, and one copy, not seven, so the "seven copies" ruling does not cover it.
An agent
verify:sat-calc
VerifiedThe data
Test a calculator-minded reader's claim that one footnote, 'The hundred stands for 4,356 visits', is made to serve both San Antonio's grid and the other six branches' grid.
It confirmed the footnote prints San Antonio's own total under the other six branches' grid, whose hundred is 20,569 visits, so the picture and its table disagree for anyone who multiplies.
REAL. The footnote is rendered once per module, always from the branch's own total, and it lands directly under the pooled grid.
An agent
verify:sat-calc
VerifiedThe data
Test a calculator-minded reader's complaint that El Paso's '77/100' hours figure is dressed like the grades and set beside the 93/100 grade, inviting a meaningless subtraction.
It re-checked the El Paso and San Antonio figures, found no correctness fix needed, and offered only an optional change of the unit's form for Ben to decide.
No correctness fix. The figures, windows, denominators and words are all right, and the reader's subtraction is invited by nothing on the page.
An agent
verify:sat-calc
VerifiedThe data
Test a calculator-minded reader's claim that the footer's 'out of the pack' rule is never used and would apply to both San Antonio (15 points off) and Tyler (14 off).
It found both of the reader's claims false against the data (the rule is used, on three measures per branch, and flags exactly the branches the page singles out), but found the sentence itself wrong: it names the wrong set (the other six, not all seven), says 'off' where the rule means under, and never says which measures it applies to.
The reader's two assertions are both FALSE against the data, but the sentence that produced them is genuinely wrong in a way the reader did not catch.
An agent
verify:sat-calc
VerifiedThe data
Test a calculator-minded reader's claim that the agency's '$97,331 earned, still collectable' names no time window while the branch panel says 'not only this quarter'.
It found the reader's reason wrong (the two figures are one measure over one window and add up to the cent) but the conclusion right: the $97,331 reaches back to Apr 11, 51 days before the quarter, and the first screen never says so.
THE READER'S REASON IS WRONG, THE READER'S CONCLUSION IS RIGHT.
An agent
VerifiedThe data
It picked A as the only draft that shows the number, the line and the gap, wrote the same three-layer combination as the other two readers, and listed nine things that did not add up.
Net: C opens, A lands, one borrowed line from B explains the magnitude.
An agent
read:why-general
VerifiedThe data
As an educated general reader with no healthcare, data or software background, read three draft answers (A, B, C) to 'Why is San Antonio failing?', explain each, rank them, pick one, and list every figure or word that does not add up.
It picked draft C as the only one it could retell from memory, still proposed A's picture and one line from B under it, and listed ten things that did not add up, including B's broken multiplier.
C is the only one I could retell without the page, and retelling is the whole job of a dashboard panel
An agent
read:why-ops
VerifiedThe data
As a Texas home-care operations director with no other context, read three draft answers (A, B, C) to 'Why is San Antonio failing?', explain each, rank them, pick one, and list every figure or word that does not add up.
With only the screenshots and the page's text to go on, it picked A for its checkable sum (55.13 + 19.18 = 74.31), proposed the combination the lead later recommended, and listed ten things that did not add up.
B is more insightful about mechanism but its chain does not multiply (781 x 'about 5' = 3,905, not 2,956), and a chain panel whose arithmetic fails is worse than no panel
An agent
verify:why-general
VerifiedThe data
Test a general reader's claim that draft B's 2,948 rejections and 2,946 'on this reason' leave 2 unexplained on a panel arguing a single cause.
It ran the project's own queries on the modeled agency's database, read-only, and showed the two counts are different sets (2,933 in both, 15 from other reasons, 13 on visits since cleared), so the gap of 2 means nothing; it still ruled the page at fault, because only one of the two counts names its window.
The reader caught a real crack, and the real crack is five times wider than they thought.
An agent
verify:why-general
VerifiedThe data
Test a general reader's claim that draft A's two point bars do not share a scale (4.87 lost looks like a quarter of the 60-point bar).
It measured the live page with a screenshot script, found both bars drawn at 4 pixels a point, and traced the misreading to only the broken half's lost part being painted.
On the 60 side there is no shape to measure, so the eye takes the only marked thing, and that marking sits OUTSIDE the quantity
An agent
verify:why-general
VerifiedThe data
Test a general reader's claim that draft C's 'has stayed there' and '80 of 100 every week since' overstate a line that moves.
It listed the eleven stored weeks since the break (six at 80, five at 81), showed the words false for five of them, and found Tyler's version spanning 58 to 78.
The figure is right; the words pinned to it are a constancy claim the page's own stored weeks contradict.
An agent
verify:why-general
VerifiedThe data
Test a general reader's claim that draft C's end label '80 of 100' sits above the Jul 13 date while its note says the last point is a week not yet over.
It confirmed the 80 is the week of Jul 6, printed with no week above the Jul 13 date (the table says Jul 13 was 81): a new copy, in draft C's own files, of a chart-label fault the lead had found on the branch page.
So: the 80 is Jul 6's stored figure, printed with no week, in the rail beside an axis whose last tick says Jul 13, over a table that says Jul 13 was 81.
An agent
verify:why-general
VerifiedThe data
Test a general reader's claim that draft B gives three answers for how often a stuck visit is resent: 'about 5 times', 4.65, and 2,956 / 781 = 3.79.
THE 781 AND THE 2,956 ARE THE SAME SET BY CONSTRUCTION.
An agent
verify:why-general
VerifiedThe data
Test a general reader's claim that the drafts never say which period is graded, so a reader cannot tell whether the May 1 break falls inside it.
It established that the graded quarter opened Jun 1, a month after the break, found the real page states this in its header but the drafts' page has none, and flagged a draft that would print a second 'now' grade.
May 1 is 31 days BEFORE the quarter opened. The reader's question has a real and load-bearing answer, and the answer changes how the panel reads.
An agent
verify:why-general
VerifiedThe data
Test a general reader's claim that draft B's chain from 781 visits to 2,956 sends works with neither multiplier it prints, 5 or 4.65.
It confirmed the multiplier misses by 949 sends (32 percent), showed the 5 was measured mostly in the previous quarter, and found that the 2,956 had no independent recount.
VERDICT IN ONE LINE: the page prints a multiplier on the arrow between two figures it also prints, and the multiplication misses by 949 exports (32%).
An agent
verify:why-general
VerifiedThe data
Test a general reader's claim that draft B's figures imply 772 visits on the payer reason while its chain says 781.
It found the reader's 772 is a stored figure the page never prints, and that 772 (this reason, cleared or not) and 781 (every reason, still not cleared) are different sets; the six reasons' uncleared visits add up to exactly 781.
An agent
verify:why-general
VerifiedThe data
Test a general reader's claim that draft C shows 80 of 100 accepted first time beside 19.18 of 40 points kept (48 percent) with nothing reconciling them.
It showed from the grade formula in the code that the 80 counts each visit once on its first try while the 40-point half counts every send (3,031 rejections in 6,577 sends), and that calling that half 'visits the state accepted' is the wrong noun.
So 80% is 845 bad first tries over 4,295 visits; 48% is 3,031 rejections over 6,577 submissions of those same visits. Nothing in C reconciles them.
An agent
verify:why-general
VerifiedThe data
Test a general reader's claim that drafts A and B give two counts, 1,308 and 781, of San Antonio's stuck visits.
It found both counts stored and right but over different sets and windows (the 24 clients since May 1, against every cause this quarter), with neither draft naming its window, a fault already fixed on the branch page.
BOTH FIGURES ARE STORED AND BOTH ARE ARITHMETICALLY RIGHT. The defect is the window labelling around them, and it is the C1 class (S5-ground-truth section 9) surviving in the two why comps
An agent
VerifiedThe data
It pinned the grade formula on the stored quarter (40 x (6,577 - 3,031) / 6,577 = 21.57) and showed the 40-point half counts every send while the chart counts each visit once, so the shared noun 'visits the state accepted' is the fault.
Both are right. They sit 32 points apart because one counts visits once and the other counts every send.
An agent
VerifiedThe data
It confirmed every figure right and the reader's point: the number 80 is used page-wide for the pass line and, on draft C, for a first-try acceptance rate set in the biggest red type with only 'of 100' under it.
A correct rate wearing the grade's uniform, at the grade's own passing value, on the page whose whole claim is "74, and the line is 80."
An agent
VerifiedThe data
It found the reader's subtraction lands exactly on a stored figure (772) and that the bridge closes both ways, and flagged that the 2,946 is not a part of the 2,948 printed beside it.
THE READER'S ARITHMETIC IS CORRECT AND LANDS ON A STORED FIGURE.
An agent
VerifiedThe data
It found both right but counting different sets over different windows, with draft B's arrow handing the 24 clients' cause to the branch's all-cause count; unasked, it also pointed to B's 'about 5 times', which the same reader had already raised in another claim.
The noun on card 2 is wrong, not the number.
An agent
VerifiedThe data
It showed that one word, 'these', ties a rate measured on 542 visits whose bouncing had settled to the 781 visits still open, and that this fault had been swept from the branch page but not from the drafts.
All three figures are stored and true, but two populations are being crossed inside one panel, with the crossing carried by a demonstrative pronoun.
An agent
VerifiedThe data
It re-ran the project's queries on the modeled database, read-only, showed the 8 are rejections the state itself caused, which do not count, and that the page's 'the state counts every rejection' is false by its own stored flags.
The 8 are the state's own fault (the two counts_against_score=0 reasons), so they do not score. That is not an unexplained gap, it is the handbook rule the panel exists to teach.
An agent
verify:why-ops
VerifiedThe data
Test an operations-director reader's claim that draft B's 2,946 rejections minus 2,174 repeats (772 visits) against its 781 stuck visits is a defect.
It found 772 is a stored figure the page withholds, that the six reasons' uncleared visits add up to exactly 781, and that on quieter branches the same subtraction would show a part bigger than its whole.
So the figure the reader derived is a real stored value the page withholds.
An agent
VerifiedThe data
It judged the omission deliberate and correct (the branch-wide 4.65 would bring back a fault already swept out) and, while checking, found draft A's equation printing Houston as 54.22 + 38.98 = 93.21, where a calculator gives 93.20.
THE CLAIM ITSELF: not a defect. Every figure on A is stored, and the only sum A invites comes out.
An agent
verify:why-ops
VerifiedThe data
Test an operations-director reader's claim that draft B's '4.65 times the average rejected visit bounced' cannot fit its own 781 visits and 2,956 sends.
It found the 4.65 is measured on 542 visits over 28 days, 19 of them in the previous quarter, so it and the 781 count almost separate sets, and that neither window was named.
The reader's arithmetic is right, and the defect is bigger than the one they caught.
An agent
VerifiedThe data
It found the real mismatch inside draft C, where a points strip counted from Jun 1 sits under a chart running from Dec 1, 2025 with no year shown and no window named.
The actual crossing is inside C, not between A and C.
An agent
verify:why-ops
VerifiedThe data
Test an operations-director reader's claim that draft C's '80 of 100' is read off the one unfinished week its own note disclaims, the optimistic end of the line.
It showed the 80 is the last full week (Jul 6), not the unfinished week (Jul 13, which reads 81), so the page had printed the cautious end and the reader had it backwards, but the label names no week.
The reader's figure is right and their attribution is inverted.
An agent
verify:why-ops
VerifiedThe data
Test an operations-director reader's claim that draft B's 2,956 sends, 2,948 rejections and 2,946 'on this reason' leave gaps of 8 and 2.
It re-ran the queries on the modeled database, read-only, and broke the 2,956 down by reason: every one was rejected, 8 were the state's own errors that do not count, and the page stated only half the state's rule.
THE THREE FIGURES ARE THREE DIFFERENT POPULATIONS. All stored, all correct, none computed in the browser.
An agent
verify:why-ops
VerifiedThe data
Test an operations-director reader's claim that draft C's chart (80 of 100) and its footer (19.18 of 40, which is 48 of 100) cannot both describe the branch.
It showed the chart counts visits on their first try and the points count every send (repeats turn a 20 percent first-try failure into 52 percent of sends), and that draft C calls both 'visits the state accepted'.
Both figures are stored and both are arithmetically right; what is wrong is that panel C calls them the same thing.
An agent
VerifiedThe data
It found neither figure wrong but neither draft naming its window, warned that draft A moved onto the branch page unchanged would bring back the part-bigger-than-whole contradiction, and also pointed to B's multiplier, which the same reader had already raised in another claim.
The knowledge exists in the build and never reached the rendered words. That is the defect.
An agent
verify:why-ops
VerifiedThe data
Test an operations-director reader's claim that 781 stuck visits times 'about 5' trips does not give draft B's 2,956 sends.
It confirmed from the query definitions that the 2,956 are the sends of exactly those 781 visits (3.79 each), that the 5 comes from a different 542 visits, and that the panel's own note contradicts its arrow.
The reader's arithmetic is right and the defect is real: the connector between the 781 card and the 2,956 card asserts a rate that the page's own two adjacent figures contradict.
An agent
VerifiedThe data
It showed the claim stands on five fully settled weeks (May 4 to Jun 1, all 80 or 81), so the reader's premise was false, while confirming that the big 80 sits inside the shading and names no week.
THE READER'S PREMISE IS FALSE; THE OBSERVATION UNDER IT IS TRUE.
An agent
verify:why-ops
VerifiedThe data
Test an operations-director reader's claim that drafts A and B give 1,308 and 781 for the same stuck visits with nothing to tell them apart.
It found draft A's header 'SFY26 Q4 to date' false for the 1,308, which is counted from May 1 in the previous quarter, and draft B's 781 carrying no window and no 'every cause'.
That is a false window on a shipped label, not an invited misreading.
An agent
verify:why-ops
VerifiedThe data
Test an operations-director reader's claim that the drafts' periods conflict: a May 1 cause outside a quarter starting Jun 1, and no date range stated.
It found the reader's calendar right but the inference wrong (a cause before the quarter can produce rejections inside it), and under it a real fault in draft A's 1,308; it also pointed to B's multiplier, which the same reader had already raised in another claim.
The reader's premise about the calendar is right and their inference about the contradiction is wrong, but one real labelling defect sits underneath it.
An agent
verify:why-ops
VerifiedThe data
Test an operations-director reader's claim that 80 means two things on the page: the pass line and draft C's acceptance rate.
It found the clash real but a clash of units, not a wrong number, and worse on the real page: three 80s in two units within one scroll, one of them '80/100 Passes'.
And the reader undercounted: on the shipping room there are THREE 80s, in two different units.
An agent
verify:why-ops
VerifiedThe data
Test an operations-director reader's claim that 20.82 of 40 points lost is asserted on all three drafts and derived on none.
It found the 20.82 stored and right, but draft B's 2,948 card claims to be the grade's count when it counts only rejections on visits not yet cleared (the grade counted 3,031), so no stored reading turns 2,948 into 20.82.
The reader's headline complaint is half right and the half that is right is worse than they said.
An agent
merge
PooledThe data
Merge the twelve readers' reports and the 120 checkers' verdicts into one deduplicated report for the lead, in ten set sections, never softening a verdict.
It wrote one report: 9 of 12 readers could say why San Antonio is failing, 24 confirmed defects plus three labelling faults (two already known), the drafts' tally, and four things no reader raised.
Both lists point the same way. Every strongest object is a piece of writing or a panel whose argument is carried by a sentence. Every weakest object is a picture.
A message of Ben's
A check of the lead's
Before I hand the build to the agents I need a few facts for the work order: the stored fields behind B's sentence, how the comps take a branch, what the kit's registry demands of a new module, and the next free ruling number.
An agent
data
PooledThe data
Add the new figures the overnight page repairs needed to the stored results, each with an independent matching check in the second program (reconcile.py, which works the figures out again on its own).
It stored the new fields (such as the last week the data treats as settled, and the 777 of the 24 clients' visits not yet accepted that fall inside the graded quarter) and gave each a second-program check, taking the agreeing figures from 7,425 to 7,796.
Instead reconcile.py now remeasures the whole curve from the raw transactions (the 8 horizons, the 99 per cent rule), so settle_days = 35, settled_through = 2026-06-09 and settled_through_week_start = 2026-06-01 are all second opinions
An agent
build:contract
Pooled, also verifiedThe data
Build the new contract line: one sentence and a row of four grades (the agency and its three payer contracts), chosen only from stored figures.
It built the contract line with 39 tests, fixed its own first footnote that wrongly said the state issues all four grades, and flagged one word in the lead's approved sentence that the stored data did not back.
It does not. docs/project3-world-and-rules.md section 3 is explicit: payers score each provider contract, not the provider as a whole. Three of the four cells are grades the state issues and the agency cell is our own roll-up.
An agent
build:grid
PooledThe data
Repair the grid of a hundred visits: name both totals it stands for, show all four groups, give each group its own mark and print each group's count.
It gave the grid a distinct mark for each kind of visit, printed every group's stored count and both hundreds' totals, and tested it in all six data states.
Both cuts are 18.0 square units, drawn to the same area on purpose so a run of one group is not visibly heavier than another's (that would be a quantity claim nobody stored).
An agent
build:chart
PooledThe data
Redraw the week-by-week chart the way Ben's chosen draft drew it and repair its wording from the cold read's list, using only stored figures.
It redrew the weekly chart, named the week each end figure belongs to, and rewrote the footnote to say which weeks have settled and what the chart's hundred counts, all from stored figures, with 71 test checks.
R2: the settling footnote reads the tranche's method.settled_through_week_start and is said in weeks: "Weeks up to the week of Jun 1 have settled; from the week of Jun 8 on, rejections are still coming back."
An agent
build:screen
Pooled, also verifiedThe data
Apply the cold read's repairs to the first screen: the date line, the badges, the track's labels, the glossary and several sentences.
It shipped the repairs with 61 new test checks, and refused two sentences in the lead's work order after checking them against the code and every test state, shipping true ones instead.
R30 was wrong in a way the work order did not catch, and my rewrite differs from its proposed wording.
An agent
build:words
Pooled, also verifiedThe data
Apply the cold read's repairs to the page's sentences (cause card, blind-spot tile, money panel, step list), including a new line saying how many of the stuck visits fall in the graded quarter.
It rewrote about a dozen sentences from stored figures, and refused the work order's '777 of the 781' line after querying the modeled database and finding 14 of the 777 were never sent.
THE DEFECT WORTH YOUR ATTENTION. The work order's R19 sentence, "777 of the 781 bounced this quarter are theirs", is not true and I did not ship it.
An agent
build:checks
PooledThe data
Build a new panel showing the grade's two halves as two bars with the sum written out, and name the 40-point half by what it actually counts.
It built the two-bars panel, named the 40-point half 'Sends the state accepted' instead of visits, and tested it across the live data and five test states.
R24: the 40-point half is named "Sends the state accepted" in HALF_NAME, in the finding sentence and in the table twin; the equation chip reads "sends accepted". The word "visits" appears nowhere near the 40, on any branch, in any state, and the test asserts it.
An agent
room
Built, also pooled
Place the six builders' new pieces on the branch page and first screen as ruled, run every check, and shoot the pages at 2,560 wide.
It composed the pages, wrote one footnote from the projection code, caught two layout faults in its own screenshots, and left the remaining design calls to the lead.
Nothing in this stage computes. The one place it came close: the quarter tick's collision could have been fixed by testing whether two stored x percents are within N of each other. That is arithmetic on stored values to decide what the page prints, so it is out.
A message of Ben's
An agent
critic:1
VerifiedThe data
As a fresh reader who has not seen the code, read only the screenshots and list every defect: what is wrong, where, how bad, and the smallest fix.
It listed 23 defects (6 blocking) from the screenshots, including placeholder text on the page, a sentence counting three finishes over two rows, and figures on one page that contradicted each other.
Two finish chips ship a placeholder instead of a verdict: each shows the score (77, 75) above a dashed grey box reading "wording not ruled"
An agent
fix:1
Pooled, also verifiedThe data
Fix every real defect the first critic listed, checking each claim against the code and data first, and list the ones judged not real.
It fixed 15 defects with 14 new tests, showed two of the critic's claims were wrong, and refused to work out on the page anything the data did not store.
Two of the twenty-two are not defects. "The other six companies" at Austin does not exist in the source or in the pixels; they read the provenance panel two inches below.
An agent
critic:2
VerifiedThe data
As a second fresh reader, read only the screenshots after the fix round and list every defect.
It listed 17 defects, including El Paso's hours picture counting 85 delivered under a printed 77, and several places where two figures for the same thing disagreed without explanation.
The square array has a three-row legend and only two marks.
An agent
Final S5b pass: nine lead rulings
Pooled, also verifiedThe data
Apply the lead's nine rulings to the branch room's sentences, grids and footnotes from stored values, and check a word two critics reported.
It applied all nine rulings in six commits with every check green, confirmed in the first program that the grid's shares are rounded to add to 100 before adding a note that says so, and found that the phrase two critics reported was not on the page.
F9 needed no fix: "companies" renders nowhere except the honest "no real people, clients, or companies" disclaimer.
A check of the lead's
The re-shoot is on disk. My final look, on the three things the pass changed most: the band with the money moved out of the rail, El Paso's hours grid, and the first screen.
An agent
gates:fast
VerifiedThe data
Re-run, read-only, every check the overnight session claimed passed (the kit tests, the colour check, both type checks, lint and the second program) and report exactly what they print.
It re-ran all the checks, and every result the overnight report claimed for them held exactly: 589 tests passing and the second program agreeing on all 7,796 figures.
Every overnight claim verified exactly: - 589 tests claimed, 589 tests printed, 589 pass, 0 fail (duration 372.82 ms).
An agent
audit:data
VerifiedThe data
Try to prove wrong the report's data claims: the four contract grades, the new stored fields and their second-program checks, the 24,925 visit total, and that the page computes nothing.
It confirmed every figure against the stored results (86, 91, 79 and 92; 24,925 visits; 7,796 agreeing), and found two of the three new flags were stored and checked but read by nothing on the page.
TWO OF THE THREE D-d FLAGS ARE STORED, TWINNED, TYPED AND READ BY NOTHING ON THE PAGE.
An agent
audit:words
VerifiedThe data
Try to prove wrong seven claims in the overnight report about the page's words and structure, against the code and the rendered page.
It confirmed six claims, and caught a grammar fault that screen readers would hear in two test states and a test that guarded a cut phrase more narrowly than the report said.
A REAL GRAMMAR DEFECT, shipped by the same fix round that filled the close slot.
An agent
gates:build
Built
Run the production build read-only and confirm the claim that the dashboard page is prebuilt as a static page.
It ran the build, which succeeded, and confirmed the dashboard and lab pages come out as prebuilt static pages.
A check of the lead's
Found something real while looking: the page you saw at localhost:4180/v2 isn't quite the real page. Checking the judging room's proxy now to confirm and fix.
A check of the lead's
I'm caught up, and every gate the overnight session claimed green reproduces.
A message of Ben's
A message of Ben's
An agent
DATA
Pooled, also verifiedThe data
Store the staff, each branch's quarter timeline and the grade's halves as whole numbers, each with a second-program check worked out again from the database, and check the work order's sentences before using them.
It added 351 checked figures (8,147 now agreeing), planted eight errors in a copy of the data and saw the second program catch every one, and flagged where the work order was wrong.
TAMPER TEST: eight stored values perturbed in a copy of the dataset (a seat's name, a seat's role, a mark's pct by one hundredth, a dropped first_lock mark, a flipped lock_before_safe_date, a wrong label, a whole half, a weight).
An agent
SCAFFOLD
Built
Build the two review pages and six empty draft folders so six builders can each work in their own folder without touching a shared file.
It built the pages and placeholders, checked them in a browser, and noted three places where the work order described the existing code wrongly.
Three sentences in the work order did not survive the code.
A message of Ben's
An agent
CHECK:first-screen
VerifiedThe data
Try to prove false every plain-language rewrite of the first screen's sentences against the stored data, five test states and the rules research, and write a true sentence where one fails.
It checked 18 rows and found 11 not plainly true, including a date line that brought back 'day 44' and a half of the grade described as visits accepted the first time.
The projected finish is being written as today's grade in three places (header clock "Day 44", strip legend dropping "heading", blind spot "El Paso's grade is 93").
An agent
CHECK:breakdown
VerifiedThe data
Try to prove false every rewritten sentence about the hundred and the grade's two halves, checking each whole number against the stored figures.
It checked 12 rows and found 8 not plainly true, three of them the 40-point half described as visits again, and two lines promising that a fix undoes the loss, which the stored finishes contradict.
THE ONE PATTERN FAULT ACROSS MY SECTION: R24 is reintroduced in three places. Lead TYL, Lead SAT and the bar name all say the 40-point half counts VISITS. It counts visit transactions.
An agent
CHECK:why-and-week
VerifiedThe data
Try to prove false every rewritten sentence in the cause card and week-by-week sections, including 'the state has rejected every visit' and 'visits typed in by hand still get paid'.
It checked 17 rows, opening the modeled database read-only, and found 8 not plainly true, among them 1,308 visits called rejected when 15 were never sent.
An agent
CHECK:action-and-chrome
VerifiedThe data
Try to prove false every rewritten sentence about money, the action steps, the two dates and the page's frame, against stored figures, test states and the state's rules.
It checked 19 rows and found 11 false, including a step that would print 'these 0 attendants' in one test state, 'deadline' for a date nobody issues, and a billing line that left out the state's acceptance.
"Get these 0 attendants clocking in on the mobile app" is a sentence no stored value should be able to produce. The count belongs on the step's figure, where it already is.
An agent
BUILD:B2
Pooled, also verifiedThe data
Build draft B2 of the action section, 'the rail': the plan, the owners and a dated rail showing what each day is worth.
It built the rail with tests and corrected three claims in the work order: Tyler's eight-name list, the lock warning at both branches, and which lock date belongs to the plan.
Three things in the work order turned out false against the data or the pixels.
An agent
BUILD:B3
Pooled, also verifiedThe data
Build draft B3 of the action section, 'if, then': three plain rows (fixed today, fixed by the safe date, not fixed) with the plan and its owners.
It built the draft and corrected three work-order claims: the eight-name list, which lock date the plan is up against, and a per-step owner the data never stores.
Three things in the work order did not survive the data.
An agent
BUILD:A2
Pooled, also verifiedThe data
Build draft A2 of the first screen, 'the board': one ranked, clickable table of the seven branches, using only stored figures.
It built the board with 35 tests and reported five things in the work order that were false or unusable, including a suggested El Paso sentence built on a 23 the page would have had to subtract.
Five things in the work order turned out false or unusable against the code and the data.
An agent
BUILD:B1
Pooled, also verifiedThe data
Build draft B1 of the action section, 'the timeline': the plan's steps and the three finishes on one time axis, each step with its count, owner and date.
It built the timeline and found four work-order claims the data did not support, including a per-step date and a per-step owner that nothing stores.
Five things in the work order turned out false or incomplete against the code and the data.
An agent
BUILD:A1
Pooled, also verifiedThe data
Build draft A1 of the first screen, 'the track is the selector': the score track's rows become the control that opens each branch.
It built the draft, checking the work order's suggested sentences against the data first and replacing the ones that were false.
Three sentences the work order drafted turned out false against the code or the data.
An agent
BUILD:A3
Pooled, also verifiedThe data
Build draft A3 of the first screen, 'master and detail': the seven badges pick a branch and one panel walks its state, breakdown, why, insight and action.
It built the draft Ben later picked, used the fact-checked sentences instead of the retracted examples, and caught a 48-versus-47-days mismatch in a suggested line.
The work order's section 2 worked-example table was RETRACTED while I worked.
A check of the lead's
I checked my own rewrites before applying anything, and most of them failed. 38 of the 66 rewritten sentences said something that isn't quite true.
An agent
CRITIC:first
VerifiedThe data
As a fresh reader who must not read the code, shoot the three first-screen drafts and list defects only, including any number or sentence that disagrees with another.
It listed 55 defects across the three drafts, including a panel saying nothing in El Paso's grade needs fixing beside one saying about 1 in 4 authorized hours is not delivered, and figures that disagreed with their neighbours.
Nothing on the comp distinguishes them, so they read as the same figure stated twice and wrong once.
An agent
CRITIC:action
VerifiedThe data
As a fresh reader who must not read the code, shoot the three action drafts for San Antonio, Tyler and the too-late test state, and list defects only.
It listed 34 defects, including a lead-in line that was false of one of the rows beneath it and points that looked better for fixing later than fixing today.
Doing the work today scores worse on one half than doing it a month later.
An agent
FIX:action
Pooled, also verifiedThe data
Fix every real defect the critic listed in the three action drafts, checking each claim against the pixels and the code first.
It fixed the defects it received, showed four of the critic's claims false or overstated, and corrected a line in B1 that was false in the too-late test state.
The 164 against 183 at Tyler is not one count printed twice.
An agent
FIX:first
Pooled, also verifiedThe data
Fix every real defect the critic listed in the three first-screen drafts, checking each claim against the pixels and code first.
In four commits it fixed most of the defects it received and explained the ones it left, showed some of the critic's claims wrong against the stored figures and the pixels, and removed a 'then resubmit' step that does not exist.
That sentence was not on the critic's list and I corrected it anyway, since the work order says to print the true sentence.
An agent
TRAPS:A1
VerifiedThe data
Read-only: check every sentence draft A1 can print against the sixteen known wording traps, the stored data and every test state, and give a corrected sentence for each real problem.
It found 12 problems (2 false, 6 misleading, 4 about voice only), among them 'put 24 records back on MCO 1' where the clients moved to it, and 'Austin is under the line' in a state where Austin finishes at 81.
So the 24 records were on MCO 2, the clients MOVED to MCO 1, and MCO 1 is a new value for those records, not a restoration.
An agent
TRAPS:B2
VerifiedThe data
Read-only: check every sentence draft B2 can print, for San Antonio and Tyler and in every test state, against the sixteen known wording traps and the stored data, and give a corrected sentence for each real problem.
It found 21 problems (10 false, 7 misleading, 4 about voice only), re-running the payer query on the modeled database to confirm 15 of the 1,308 visits were never sent, and catching the safe date described as the last day a fix does anything.
Re-ran sql/15-payer-mismatch.sql against pecan-hollow.sqlite: of the 1,308, exactly 15 have exports = 0, so the state never saw them, let alone rejected them. 1,293 were sent and not accepted.
An agent
TRAPS:B1
VerifiedThe data
Read-only: check every sentence draft B1 can print, for San Antonio and Tyler and in every test state, against the sixteen known wording traps and the stored data, and give a corrected sentence for each real problem.
It found 20 problems (8 false, 9 misleading, 3 about voice only), including 'nothing needs fixing' in a state where money locks that day, 'no safe date on file' beside the printed safe date, and 1,308 visits called bounced.
What is true is that the GRADE is not at risk: projection.recovery is null, which is the only thing the null-plan branch actually tests.
An agent
TRAPS:A2
VerifiedThe data
Read-only: check every sentence draft A2 can print against the sixteen known wording traps, the stored data and every test state, and give a corrected sentence for each real problem.
It found 16 problems (3 false, 11 misleading, 2 about voice only), among them a typed-in 'five that pass' that is wrong in three test states and a claim that every branch loses most points in the same half.
It is also a comparison the page is making rather than reading: it claims a ranking across seven branches that no stored field states.
An agent
TRAPS:A3
VerifiedThe data
Read-only: check every sentence draft A3 can print against the sixteen known wording traps, the stored data and every test state, and give a corrected sentence for each real problem.
It found 13 problems (1 false, 5 misleading, 7 about voice only), including 'Above the 80 that passes' printed for a branch finishing at 77, and judged A3's wording the most trap-disciplined in the kit.
points_short is only stored on a branch marked under; every branch marked close stores null and gets this sentence regardless of its finish.
An agent
TRAPS:B3
VerifiedThe data
Read-only: check every sentence draft B3 can print, for San Antonio and Tyler and in every test state, against the sixteen known wording traps and the stored data, and give a corrected sentence for each real problem.
It found 18 problems (7 false, 8 misleading, 3 about voice only), re-running the payer query read-only to confirm 15 of the 1,308 were never sent, and catching 'left Jul 11' printed in states where that date is still ahead.
I ran that query read-only against pecan-hollow.sqlite: of San Antonio's 1,308, exactly 15 have zero exports, so the state has never seen them and never rejected them; 1,293 were actually rejected.
An agent
DATA
Pooled, also sourcedThis world
Fetch a public-domain outline of Texas, place the seven branch cities on it with stored positions and a second-program check, and verify the coordinates against a second source.
It downloaded Natural Earth's public-domain state and city files, drew Texas from them, checked each branch city against the city file (all within 3.4 km), and added 18 checked figures (8,165 agreeing).
What the twin can and cannot catch, so nobody over-trusts it: the coordinates are authored constants, so a wrong city would agree with itself in reconcile.py.
An agent
BUILD:A4
PooledThe data
Copy A3, apply every A3 finding from the trap check, and build a map of Texas and a ranked list as one control for picking a branch.
It built A4 in the separate build copy with 63 tests, applied the thirteen A3 findings (a later check found one only partly applied), and kept every position on the stored percents.
THE A3 TRAPCHECK. All thirteen findings are applied to my copy and each is asserted by the test.
An agent
BUILD:A5
PooledThe data
Copy A3, apply every A3 finding, and build a Texas map as the main picker with a callout card for the picked branch.
It built A5 with 72 tests, measured a callout card that moves to the pin and found it could not fit, and shipped a fixed card tied to the pin instead.
I built it, measured it, and it does not work at this size: all seven branches sit in the eastern half of Texas, so a card large enough to hold the packet's short form (292 x 477 at the 1440 canvas) covers two other pins wherever it lands
An agent
CRITIC
VerifiedThis world
As a fresh reader who must not read the code, shoot drafts A4 and A5, compare the outline and the city positions with your own knowledge of Texas, and list defects only.
It listed 35 defects, including chips that covered the cities with two sitting outside the state, a 40-point half named both visits and sends, and an outline credited in one draft but not the other.
No point marks the city. The rounded chip is centred on the city coordinate and covers it, and the name sits on a different side for every branch (right of the chip for El Paso, Austin, Tyler; left for Fort Worth, San Antonio, McAllen; above for Houston).
An agent
FIX
Built, also verified
Fix every real defect the critic listed in A4 and A5, checking each against the pixels first and listing any judged not real.
In A4 it made each city a point with a plate above it, in A5 it removed a leader line that crossed the map, it confirmed four wording defects it could not fix in the shared labels file, and it judged one critic claim not real and another not something it could act on.
THE ONE THING TO READ FIRST: four of the critic's A4 findings are real and I could not fix any of them, because all four sentences live in the kit's CENTRAL labels.ts, which the hard rules forbid me to edit and which A1, A2, A3, A5 and the live /v2 all print.
An agent
TRAPS
VerifiedThe data
Read-only: check every sentence A4 and A5 can print against the sixteen traps, the stored data and every test state, and confirm each A3 finding was applied in both.
It found 9 problems (1 false, 6 misleading, 2 about voice only), querying the modeled data to show a waiting count covers this quarter only, and found one A3 finding only two thirds applied in A4.
I queried the dataset: Austin has 79 delivered visits never entered inside SFY26 Q4 and 23 more before it, back to 2025-12-26 (Fort Worth 79 + 22, Houston 80 + 15, McAllen 61 + 12, Tyler 164 + 25, El Paso 49 + 11, San Antonio 61 + 10). So as written it is short by 23 at Austin.
A check of the lead's
Let me check one thing before I answer: what geography the modeled data actually holds below the branch, since that decides which map structures are honest.
A message of Ben's
A check of the lead's
Before I answer, let me look at how the current world is parameterised, so the reshaping proposal is about the real generator and not a guess.
An agent
find:reviewer
Sourced, also verifiedThis world
Read the modeled world as an EVV compliance reviewer at a Texas health plan or at the state would, and propose up to ten earned research questions plus what geography might confound, without researching the answers.
It read the world document, the data layout and the stored figures, named two things that would make a reviewer stop reading, and proposed ten questions and fourteen confounders on contracts, plan changes, reporting duties, workforce and the clock.
What I test a modeled dataset against is not the formula (the formula is published and this world already has it right) but the plumbing around it
An agent
find:director
Sourced, also verifiedThis world
Read the modeled world as the director of operations of a seven-branch Texas agency would, and propose up to ten earned research questions plus what geography might confound, without researching the answers.
It read the world document, the data layout and the stored figures, named what would make a director stop trusting them, and proposed ten questions and twelve confounders, from how authorized hours are lost to when the money arrives.
What makes me stop reading: a metric that would look the same at a well run rural branch and a badly run one, a dollar figure with no cost of the fix beside it, a weekly chart with no holidays or weather in it
An agent
research:01-service-areas
SourcedThe industry
Find how Texas decides where a home care agency may serve, how the health plans' service areas are drawn county by county, and which area each of the seven modeled cities sits in.
It read the state's licensing rule and the state's county lists for the health plans' service areas, and found that the seven cities sit in seven different service areas, each served by two or three plans, with no single plan covering all seven.
The seven modeled cities land in seven different service areas: San Antonio in Bexar, Tyler in MRSA Northeast, Austin in Travis, Fort Worth in Tarrant, El Paso in El Paso, Houston in Harris, McAllen in Hidalgo, and every county each city physically touches stays inside its own area
and 2 pages not shown
9 web searches
An agent
find:scientist
Sourced, also verifiedThis world
Read the modeled world and its data as a data scientist who audits synthetic data would, and propose up to ten earned research questions plus what geography might confound, without researching the answers.
It read the schema, the generator's settings and code and the stored figures, named the specific tells that the records were made up, and proposed ten questions and thirteen confounders.
Read against that checklist, Pecan Hollow's data layer is unusually honest about its rules and unusually thin about its randomness.
An agent
research:06-county-boundaries
SourcedThis world
Find the best public county boundary files for Texas, their size and licence, how much detail a 700-pixel map needs, and whether the Natural Earth state outline and the Census county lines agree closely enough to overlay.
It downloaded and measured the Census county files, found the Census state outline is exactly the union of its counties while the Natural Earth outline strays up to 21 pixels along the Gulf coast, and measured the trap a county map sets: the 10 largest counties hold 14 percent of the land and 1 percent of the people.
the 10 largest counties by area hold 14 percent of the land with 1 percent of the people while the 10 most populous hold 4 percent of the land with 59 percent of the people, so a 254 county choropleth would put most of its ink where there is almost no business.
and 1 page not shown
2 web searches
An agent
research:08-what-ops-teams-track
SourcedThe industry
Find which EVV reports and measures Texas agencies actually run and act on, which systems hold them, and which role owns each step.
It read the state's report specifications, vendor guides and job postings, found the state names every standard report and no system report scores an attendant, and found the one report with an authorized-hours column leaves it blank.
Texas names the report list itself, so this question has a documented answer rather than a folklore one
and 6 pages not shown
16 web searches
An agent
research:09-real-data-shape
Sourced, also verifiedThe industry
Find what a real Texas EVV visit record looks like field by field, the real rejection reasons and the everyday data quirks, and say what the project's own layout lacks or gets wrong.
It read the state's published file layouts and rejection guide and found the modeled visit row too thin: one clock-method field where the state has two, no bill times, invented rejection text, and a checker that forbids overlaps that really happen.
Against that, our schema is right about the world and thin about the row
7 web searches
An agent
research:04-compliance-ladder-and-recovery
SourcedThe industry
Find what happens to a Texas provider whose usage score falls under 80, on what clock, and how it climbs back, and how a request to unlock a locked visit works.
It read the state handbook's review, lock and unlock sections and worked out that, of the quarters in the world's data, only the second could have been reviewed by the world's today, so a recovering branch must have failed that quarter, with real notice and plan deadlines.
a quarter cannot be reviewed until 95 days after its last day, so at the world's today of 2026-07-15 the only quarter a branch can have been officially told about is SFY26 Q2 (reviewable from 2026-06-03)
8 web searches
An agent
research:02-score-distribution
SourcedThe industry
Find any published figures on how Texas providers' EVV usage scores are spread: the share under 80, how many sit near the line, and how scores move from quarter to quarter.
It read six sets of state workgroup minutes, a report to the Legislature and a 2018 state inspection, found no published spread of Texas scores in anything it reached, and found a payer mismatch is the top rejection in every month observed.
Texas publishes the line and the enforcement, never the spread, and this pass closes that question instead of leaving it open
and 1 page not shown
20 web searches
An agent
research:03-reason-codes
Sourced, also verifiedThe industry
Check the state's current EVV reason code list and how each code affects the usage score, and say exactly where the project's own copy of the list differs.
It confirmed the project's reason-code file matches the state's list code for code and that no reason code costs score points, and found the world document's alternative device rule wrong, the single clock-method field unable to hold a half hand-keyed visit, and the generator using only 17 of the 28 codes.
The load bearing finding is a negative: no reason code counts against the usage score
9 web searches
An agent
research:05-branch-scale-and-money
SourcedThis world
Find what a real Texas home care branch looks like in numbers: clients, attendants, hours per client, payer mix, the state's pay rates, the attendant wage and margins.
It read the state's rate sheets, budget and licensing directory and found the managed care rate is published, the modeled branches are about half the size of real ones, and the modeled clients get too few hours a week.
The state publishes almost everything this question needs, and two of the answers contradict the model.
16 web searches
An agent
research:07-clock-in-methods-and-geography
SourcedThe industry
Find how Texas visits are clocked in, the landline check and alternative device rules, and what landline ownership and cell coverage by rurality do to the method mix.
It read the state's clock-method rules and federal phone surveys, found a 75 percent cap on alternative devices already in force, and argued the modeled landline share at Tyler was higher than the population supports.
Texas allows exactly three clock methods and the rules, not preference, decide which one a member can use, so the method mix is not a free parameter.
11 web searches
An agent
research:10-time-and-rule-changes
Sourced, also verifiedThe industry
Close every open item in the world document's list of rules still to verify, and find the EVV rule changes of 2025 and 2026 and the review calendar.
It read the handbook's revision logs, answered every open item on the world document's list, found three 2025 and 2026 rule changes that bite, and showed the world document treated the third quarter as settled when its visits were still open.
Every item in section 7 of the world doc closes, and the rules that matter are current, but one finding changes the world's time logic.
12 web searches
An agent
merge:earned-questions
SourcedThe industry
Read the three question-finders' 30 proposals and pick at most eight earned questions for a second wave, dropping any the ten core questions already cover.
It deduplicated the 30 proposed questions, picked 8 for a second wave of research and dropped 12 with a reason each, and wrote them into the earned-questions document.
Thirty proposals, deduplicated and tested against the ten core questions already running, leave eight.
An agent
verify:03-reason-codes
Verified, also sourcedThe industry
Try to prove report 03 (reason codes and the score) wrong, claim by claim, against its cited sources and at least one other, saying holds only when a fetched page supports the claim.
It re-fetched every source, read the state's PDFs as text and checked the project's files; 19 claims held and 1 could not be checked, and it found the report had stated the manual half of the usage score backwards.
The handbook's numerator is the opposite: the total number of electronic (nonmanual) visit transactions divided by the total number of visit transactions accepted by the EVV Aggregator, multiplied by 60 percent
4 web searches
An agent
verify:07-clock-in-methods-and-geography
Verified, also sourcedThe industry
Try to prove report 07 (clock-in methods and geography) wrong, claim by claim, against its cited sources and at least one other.
It re-read the clock-method rules and the federal phone survey tables; 22 claims held, 6 were refuted and 5 could not be checked, and it found the report had left out adults who keep a landline but take most calls on a cell phone: about 16 in 100 Texas adults have a landline, not 5, and about 41 in 100 adults 65 and over nationally, not 24.
The policy half of the report is solid: 14 of the 15 load-bearing rules claims hold against the handbook and against a second HHSC document that was fetched separately. The population half has a real error.
3 web searches
An agent
verify:04-compliance-ladder-and-recovery
Verified, also sourcedThe industry
Try to prove report 04 (what happens under the line, and unlocks) wrong, claim by claim, against its cited sources and at least one other.
It re-read the handbook's review, appeal and unlock sections and the managed care manual; 19 claims held and 4 could not be checked, and it found the appeal clocks, an unlock corroboration, what an unlock buys and the scored unit overstated.
The appeal clocks are HHSC clocks and the report presents them as general.
3 web searches
An agent
verify:09-real-data-shape
Verified, also sourcedThe industry
Try to prove report 09 (what real EVV data looks like at the row level) wrong, claim by claim, against its cited sources and at least one other.
It found the state's index of the rules in force and re-read the current layouts; 20 claims held, 3 were refuted and 2 could not be checked, and it found the report had cited superseded editions, so it gave clock times as captured in central time where the edition in force says local time.
The report's field-level substance survives the check. Its citations do not.
4 web searches
An agent
research:13-gap-reporting-duties
SourcedThe industry
Find what a Texas provider is already required to report when authorized hours go undelivered, and who sees it.
It read a health plan's provider manual, the caseworker handbook, the managed care contract and the state's report specifications and found the duty to report runs one client at a time and never adds up, so the page's blind-spot claim had to narrow by one notch.
A Texas PAS provider does already owe a report when it cannot deliver authorized hours, but the duty runs one member at a time and never adds up, so the page's claim narrows by exactly one notch and gets stronger for it.
and 1 page not shown
10 web searches
An agent
research:12-authorization-lifecycle
SourcedThe industry
Find how a Texas client's authorized hours are set and change over time: the assessment, reassessment, suspension during a hospital stay, program transfers, gaps, and escalation when a provider cannot staff the hours.
It read the state's caseworker and managed care handbooks and forms and found hours are reassessed yearly and suspended once a hospital or nursing home stay is confirmed to run past 30 days, and that the state's own review of the health plans does check whether services start on time.
so the honest claim is not that under-delivery is invisible but that no report compares one branch's delivered hours with its authorized hours, and that report hands the page a real external benchmark in exchange.
11 web searches
An agent
research:14-contracts-and-branch-identity
SourcedThe industry
Find how managed care contracts are set up across service areas, what the usage score is computed against, and whether a branch exists as an identifier in the state's data.
It read the handbook, the state's file layouts and the vendor's FAQ for Texas providers and found the state scores a contract's provider number, or a provider ID and tax ID pair, never a branch, and that the visit record's fields for an office's location are all retired.
So a branch level score is computable from data the agency really holds and is a management view the agency builds, not a number the state publishes
and 2 pages not shown
10 web searches
An agent
verify:02-score-distribution
Verified, also sourcedThe industry
Try to prove report 02 (the spread of usage scores) wrong, claim by claim, against its cited sources and at least one other.
It extracted each cited PDF and searched it for the exact figures; 17 claims held and 3 were refuted, among them that Texas had long denied claims without a matching visit record, which the state's own 2018 inspection contradicts.
Rules since 2015, enforcement much later. The conclusion survives, but the reason must become current enforcement: the 2024 handbook deny-or-recoup policy plus grace periods ending 2026-01-30.
and 1 page not shown
1 web search
An agent
research:18-money-after-the-claim
SourcedThe industry
Find what happens to an EVV claim after the match: when it is paid, how a denied line is reworked, how the claim and visit clocks interact, and what can be taken back later.
It read the handbook, the state's claims manual, the managed care contract and a public training module and found a claim is a cycle of rework, payment and possible recoupment, where the model kept only one pass.
A claim is a cycle, and the model currently keeps only its first minute: one claim per visit, adjudicated once on the Monday after acceptance, never corrected.
and 2 pages not shown
10 web searches
An agent
research:16-workforce-and-tenure
SourcedThis world
Find the real workforce numbers behind a Texas branch (turnover, vacancy, caseload, wages by metro) and what a new attendant's first eight weeks look like.
It read the state's workforce report built on cost reports, federal wage tables and licensing rules and found turnover and vacancy far calmer than the model, a cashier out-earning an aide in all seven metros, and no required TB screening.
Texas HHSC collects the exact numbers the model asserts, and they are far calmer than the model.
14 web searches
An agent
research:15-plan-change-churn
SourcedThe industry
Find how a Texas Medicaid client changes health plan, when the change takes effect, and how the provider agency learns of it.
It read the state's managed care handbook, enrollment pages and the vendor's guide and found that a plan change made through the state's enrollment line takes effect on the first of a month and that the state's notice goes to the receiving plan, with no notice to the agency found, so San Antonio's fault sentence and its cohort date had to change.
The state notifies only the receiving MCO, on the 834 daily enrollment file; the provider agency is told nothing and has to look
10 web searches
An agent
research:17-relative-and-live-in-attendants
SourcedThe industry
Find how often the attendant is a relative or lives in the same home, which programs allow it, and whether it changes how visits are clocked.
It read the state's rules and handbooks and national surveys and found Texas bars only a few relatives and grants no live-in exemption from clocking in, found no published count of family attendants in Texas, and found Family Care is funded outside Medicaid.
Texas puts almost no relation test on who may be a paid attendant, so a relative attendant is ordinary rather than exceptional
and 1 page not shown
12 web searches
An agent
research:11-delivered-vs-authorized
SourcedThe industry
Find what the gap between authorized and delivered hours normally looks like in Texas attendant care, what it is made of, and whether unstaffed hours can ever be made up.
It found the federal government collects no figure for hours delivered against hours authorized (the reason its rule now requires states to report one) and found no Texas figure, used the only agency-level range it found, from an out-of-state study, and found the state's rules split a gap the rules allow from hours nobody staffed.
The page's strongest finding does not invert, but its numbers need labeling.
and 2 pages not shown
13 web searches
An agent
verify:06-county-boundaries
Verified, also sourcedThis world
Try to prove report 06 (the county boundary source) wrong, claim by claim, against its cited sources and at least one other.
It downloaded the Census and Natural Earth files again and re-measured them with a reader it wrote for the check; nothing was refuted, and it corrected several figures and wordings.
Nothing was refuted. Nine corrections are listed below, all of them about sourcing, wording or reproducibility, none of them about a conclusion.
and 1 page not shown
1 web search
An agent
verify:08-what-ops-teams-track
Verified, also sourcedThe industry
Try to prove report 08 (what operations teams track) wrong, claim by claim, against its cited sources and at least one other.
It re-read the state's report specifications and the cited vendor and job pages; 23 claims held, 1 was refuted and 5 could not be checked, and it found one claim's cited page did not contain it.
Worse than unverified: the cited page does not contain it.
and 3 pages not shown
1 web search
An agent
verify:10-time-and-rule-changes
Verified, also sourcedThe industry
Try to prove report 10 (the rule changes of 2025 and 2026) wrong, claim by claim, against its cited sources and at least one other.
It read the revision logs, the job aid and the bill codes table as text; 23 claims held, 1 was refuted and 3 could not be checked, and it found only 15 of the 24 managed care bill-code rows carry the attendant code the report gave for all.
The per 15 min half and the per hour half hold; the S5125 half does not.
1 web search
An agent
verify:01-service-areas
Verified, also sourcedThe industry
Try to prove report 01 (service areas and licensing by county) wrong, claim by claim, against its cited sources and at least one other.
It recomputed the county arithmetic from the state's pages; nothing was refuted and 3 claims could not be checked, and it pointed out that a misspelled county name the report had already noticed on the state's lists would break a map join.
County arithmetic was recomputed from the fetched page rather than taken from the report. No claim was refuted. Three claims cannot be checked, and all three are the ones the report already tags unverified, so the report's tagging is honest.
and 1 page not shown
1 web search
An agent
verify:05-branch-scale-and-money
Verified, also sourcedThis world
Try to prove report 05 (what a branch looks like in numbers) wrong, claim by claim, against its cited sources and at least one other.
The two spreadsheets were downloaded and recomputed from the raw rows rather than taken on trust.
1 web search
An agent
verify:12-authorization-lifecycle
Verified, also sourcedThe industry
Try to prove report 12 (how authorized hours are set, changed and suspended) wrong, claim by claim, against its cited sources and at least one other.
It re-read the handbook sections and assessment forms; 21 claims held, 3 were refuted and 7 could not be checked, and it found the report's proposed monthly measure would let a month of missed visits hide.
So 4.33 times adjusted weekly hours is a monthly ceiling on what may be delivered and billed, not a band inside which missed visits are compliant.
2 web searches
An agent
verify:17-relative-and-live-in-attendants
Verified, also sourcedThe industry
Try to prove report 17 (how often the attendant is family) wrong, claim by claim, against its cited sources and at least one other.
It re-read the rules and handbooks the report cited; 12 claims held, 2 were refuted and 2 could not be checked, and it found Family Care is inside EVV after all.
Family Care visits are inside EVV despite Title XX funding, so the one program where a spouse may be the attendant can appear in the data and open question 1 should close.
and 1 page not shown
2 web searches
An agent
verify:15-plan-change-churn
Verified, also sourcedThe industry
Try to prove report 15 (how a client changes plan and how the agency finds out) wrong, claim by claim, against its cited sources and at least one other.
It fetched the report's sources and checked its quotes from the state's rejection documents word for word; 18 claims held and 4 could not be checked, and it found the report's central claim that nothing rechecks a stale payer went beyond its source.
2 web searches
An agent
verify:14-contracts-and-branch-identity
Verified, also sourcedThe industry
Try to prove report 14 (what a contract is, and where a branch comes from in EVV data) wrong, claim by claim, against its cited sources and at least one other.
It re-read the state's file layouts and plan lists by direct fetch; 25 claims held, 3 were refuted in part and 6 could not be checked.
Three claims are refuted in part. Five stay unverifiable, which is what the report already says about them.
1 web search
An agent
verify:18-money-after-the-claim
Verified, also sourcedThe industry
Try to prove report 18 (remittance, rework and recoupment) wrong, claim by claim, against its cited sources and at least one other.
It read the federal rule, the state's claims manual and the managed care contract as text; 18 claims held, 3 were refuted and 6 could not be checked, and it found the federal payment floor and a plan's recovery notice misread.
The federal floor under this world is twelve months, not thirty days.
and 1 page not shown
1 web search
An agent
verify:16-workforce-and-tenure
Verified, also sourcedThis world
Try to prove report 16 (turnover, vacancy, wages and a new hire's first weeks) wrong, claim by claim, against its cited sources and at least one other.
It re-read the state's workforce plan, licensing letters and federal wage tables; 25 claims held and 6 could not be checked, among them three load-bearing figures with no reachable source, one of them a 150 percent turnover figure the report's answer leaned on.
None could be checked here because the session's web search budget was already spent, and the answer section leans on the 150 percent figure beside the state's 21 to 37 percent.
and 1 page not shown
3 web searches
An agent
verify:11-delivered-vs-authorized
Verified, also sourcedThe industry
Try to prove report 11 (the normal gap between authorized and delivered hours) wrong, claim by claim, against its cited sources and at least one other.
It downloaded the federal rule and the Arkansas evaluation and searched them for the exact passages; 25 claims held, 1 was refuted and 3 could not be checked, and it found the report's 68 percent covered only clients who got any care.
The 68 percent covers only those who received any care, not agency served clients as a group.
and 2 pages not shown
1 web search
An agent
verify:13-gap-reporting-duties
Verified, also sourcedThe industry
Try to prove report 13 (what a provider must already report when hours go undelivered) wrong, claim by claim, against its cited sources and at least one other.
It extracted the managed care contract, the managed care manual and the state's report specifications as text; 21 claims held, 2 were refuted and 2 could not be checked, and both refutations made the report's answer stronger.
The two refutations both cut the same way, and both make the report's own answer stronger rather than weaker.
and 3 pages not shown
1 web search
An agent
critic:completeness
Verified, also sourcedThe industry
Read all eighteen reports and their verifications and say what is missing, what contradicts, which load-bearing claims are still unproven, and what a next round should ask.
It read the eighteen reports and their checks together and listed 11 gaps, 19 contradictions and 66 unproven, corrected or refuted load-bearing claims, and wrote ten questions for a second round.
The pass is strong on rules and weak on people. Almost every sentence in the EVV Policy Handbook that touches this world has now been read against the primary source, and the verifications caught real errors in the reading of it.
An agent
research:32-public-evv-training-modules
SourcedThe industry
Read the state's public training modules and job aids on electronic visit verification (EVV) and find what the state portal's usage report and other reports show, including whether they can be filtered by office.
It found the state's own portal screens pasted into public state PDFs, which showed the state's usage report has no office filter and none of the ten specified system reports takes one, and caught round 1 using an outdated wording of the rejection for the wrong payer.
Reports 08 and 14 concluded the EVV Portal screens are not public because they stopped at the TMHP learning system login. That is wrong, and the way in is not the learning system: it is the screenshots pasted into public HHSC and TMHP PDFs.
7 web searches
An agent
research:23-six-plans-provider-manuals
SourcedThe industry
Read the provider manuals of the six real health plans serving the seven modeled areas and record what each asks of an agency when authorized hours go undelivered.
It downloaded all six manuals and found the notification duty round 1 had built on appears in none of them and belongs to a plan that is not on the list for this kind of care in any of the seven areas.
All six provider manuals were fetched and read as full text this run, and the notification duty report 13 built on is in none of them.
10 web searches
An agent
research:21-payer-mismatch-counts
SourcedThe industry
Find out whether a rejection for a member's payer not matching the state's record (edit Ex00031C) counts against the rejected part of the score, from the handbook, the state's edit list and any state notice.
It read the handbook, two editions of the state's edit list and five sets of the state workgroup's minutes, and found the rejections count, though no state document it read calls this edit a provider error.
The handbook does not send the reader into Appendix M to find a label; it declares the appendix to be that list. Ex00031C and Ex00043C are live Business Edits in the edition in force
and 1 page not shown
6 web searches
An agent
research:27-units-matching-and-span-billing
SourcedThe industry
Find whether billed units for attendant care are matched against the visit record, for which bill codes, and how often attendant care is billed.
It read the state's bill code table and crosswalk and found billed hours are matched for exactly the codes this agency bills, and that same-day visits combine into one claim line, so the claim design had to change.
Yes, units are matched, and for exactly the codes this world bills. The question's either or was the mistake
10 web searches
An agent
research:26-client-demographics-and-season
SourcedThis world
Find who Texas attendant-care clients are (age, whether they also have Medicare, living alone, priority status) and what seasons and dated weather events touch the data window.
It read the state's member survey, rate-setting data and health reports, found the clients the state surveys younger and less alone than assumed, and argued the modeled winter dip belongs on Tyler and McAllen rather than Austin.
The clients are younger, sicker and less alone than the phrase "home care for the elderly" implies, and Texas publishes enough to prove it.
and 3 pages not shown
15 web searches
An agent
research:22-evv-system-daily-validation
SourcedThe industry
Find out what the state-supplied visit-clocking system (EVV) checks about a member each day, and when an agency learns that a member changed health plan.
It read the vendor's documentation and the handbook and found no standard check in the agency's own system that rechecks a member's plan, while each night's export to the state returns a rejection into the branch's own queue by morning.
14 web searches
An agent
research:31-schedules-and-authorizations-loaded
SourcedThe industry
Find whether schedules and health-plan authorizations must be loaded into the state's visit-clocking system, how they are entered, and whether the planned Houston story (schedules stopped being loaded) holds.
It read the handbook and the vendor's guides and found schedules are optional and authorizations are not, so the middle of Houston's planned story did not hold and was replaced by lapsed hand-keyed authorizations.
Schedules are optional and authorizations are not, and that asymmetry answers the whole question.
5 web searches
An agent
research:28-enforcement-record
SourcedThe industry
Find what real audits, penalties and settlements against Texas attendant-care agencies look like, with amounts and methods, and what a director actually fears.
It read two real 2022 state audits of attendant-care branches, the state's penalty table and the inspector general's reports, and found both audits ended in small exact repayments for paperwork defects, with neither projected to the whole client population.
So what a director fears is a documentation defect turned into a repayment demand on visits nobody claims did not happen, a public table with his agency's name in it, and an attendant whose false clock in becomes the agency's case
14 web searches
An agent
research:19-one-sided-punch
SourcedThe industry
Find out from the state's handbook and data layouts whether a visit with one electronic punch and one punch keyed in by the office counts as a manual visit in the state's score.
It read the state's January 2026 handbook revision, its data layouts and its code lists, and found that one hand-keyed punch makes the whole visit count as manual, once, so the Fort Worth story stands.
On the EVV Usage Report such a row is one accepted transaction, counted in Total Accepted and Total Manual and absent from Total Electronic, so it costs a full transaction of the 60 percent manual term exactly as a fully keyed visit does, and it costs it once
and 5 pages not shown
10 web searches
An agent
research:24-quality-reporting-service-delivery
SourcedThe industry
Find any Texas measure of attendant services delivered against the plan, what the 84 against 95 percent figure measures, and whether a staffing gap is ever a reportable critical incident.
It found the state's external quality review, which measured how many members got all their approved attendant care (18 to 90 in 100, by plan), so round 1's claim that Texas publishes nothing to compare against had to narrow.
Texas does publish a measure of attendant services delivered against the amount authorized, and round 1 missed it
12 web searches
An agent
research:30-mountain-time
SourcedThe industry
Find how the state's electronic visit verification (EVV, the clock-in system) handles El Paso's Mountain Time: which clock the times are kept in, which day a late visit belongs to, and whether any document addresses El Paso.
It read the state's business rules, the handbook and the federal time zone boundary, and found the state never converts times and that all of Hudspeth County, not part of it, is on Mountain Time.
Texas EVV has no conversion problem because Texas EVV does not convert. Business rule VSS-28P says the EVV system must use the Local Time Zone
12 web searches
An agent
research:20-filing-clock-ltc-ffs
SourcedThe industry
Find the claim filing deadline for the state's own long term care attendant claims and at the six health plans, and say whether it or the 95-day lock on correcting a visit comes first.
It found three sources for a 12-month filing window on the state's own attendant claims and 95 days at all six plans, which showed three earlier reports partly wrong and left one clock on the page: the 95-day lock.
Neither 95 days nor 365. A long term care fee for service claim for Primary Home Care, Community Attendant Services or Family Care must reach TMHP within 12 months, running from the end of the month in which the service was provided
and 3 pages not shown
17 web searches
An agent
research:25-cost-report-turnover
SourcedThe industry
Find how the state's cost report defines and counts attendant turnover and vacancy, and whether any newer workforce data exists.
It read the state's cost report forms and instructions and found the state never asked providers for a turnover rate, only for counts (staff on hand at year end and staff who left), and found no turnover or vacancy figure published since fiscal 2019.
The state never asked a provider for a turnover rate: it asked for counts and did the division itself.
and 4 pages not shown
11 web searches
An agent
verify:19-one-sided-punch
Verified, also sourcedThe industry
Try to prove report 19 wrong: fetch its sources and at least one independent one, and check each load-bearing claim's wording, numbers, revision and dates.
It re-read every source behind report 19 and found each link held; the one claim it overturned, a count that missed a third state edit, made the finding stronger rather than weaker.
the one refuted claim was an absence whose refuting passage supports the conclusion more strongly, and the named schema change stands with the addition that the whole-visit derived fields (records 65 and 126) must also go null for a one-sided row.
and 2 pages not shown
2 web searches
An agent
verify:21-payer-mismatch-counts
Verified, also sourcedThe industry
Try to prove report 21 wrong: fetch its sources and at least one independent one, and check each load-bearing claim's wording, numbers, revision and dates.
It confirmed every link of report 21's chain in the sources, but showed the report's reading proved too much, since the same state list also tests errors the state blames on the payer.
Holds, with one sharpening: every link in the chain is verified verbatim and nothing in any fetched state or plan document excuses Ex00031C
and 1 page not shown
4 web searches
An agent
verify:31-schedules-and-authorizations-loaded
Verified, also sourcedThe industry
Try to prove report 31 wrong: fetch its sources and at least one independent one, and check each load-bearing claim's wording, numbers, revision and dates.
It confirmed the rewritten Houston story but removed one leg of it, and overturned five overstatements, among them a report column called blank that is in fact filled in.
so Houston's damage should run through the 60 percent manual term and the prebilling hold rather than the 40 percent rejection term.
2 web searches
An agent
verify:22-evv-system-daily-validation
Verified, also sourcedThe industry
Try to prove report 22 wrong: fetch its sources and at least one independent one, and check each load-bearing claim's wording, numbers, revision and dates.
It found the story held and got stronger from the state's rejection guide, and caught three overstatements, including a sentence saying the state warned the branch when the rejection simply waited in its queue.
the one wording change owed is that the state did not push an alert, the rejection landed in a queue the branch owned.
3 web searches
An agent
verify:28-enforcement-record
Verified, also sourcedThe industry
Try to prove report 28 wrong: fetch its sources and at least one independent one, and check each load-bearing claim's wording, numbers, revision and dates.
It read every amount in the state's penalty table, caught a misread total and a too-narrow range, and found a 2012 State Auditor audit of home health services in the Texas Health Steps program saying the state did not verify that attendants delivered those services.
That is a State Auditor's Office report on personal attendant services, and it is the state's own auditor saying in 2012 that delivery went unverified.
and 1 page not shown
4 web searches
An agent
verify:25-cost-report-turnover
Verified, also sourcedThe industry
Try to prove report 25 wrong: fetch its sources and at least one independent one, and check each load-bearing claim's wording, numbers, revision and dates.
It re-read every source and found the report right, but caught it naming two different denominators for one turnover figure, and found that the state's base-wage requirement had ended on 2025-08-31.
the only repairs needed are that the report contradicts itself on whether the denominator is the average headcount or just the headcount
5 web searches
An agent
verify:32-public-evv-training-modules
Verified, also sourcedThe industry
Try to prove report 32 wrong: fetch its sources and at least one independent one, and check each load-bearing claim's wording, numbers, revision and dates.
It fetched one cited document twice and showed the address the research cited serves an older edition, and read a printed screenshot that showed a report working differently from the research's description.
the guide's own resolution path (read the member's eligibility record, then fix the payer on the authorization in the EVV system) leaves the San Antonio mechanism intact, so the sentence changes and the story does not.
1 web search
An agent
verify:27-units-matching-and-span-billing
Verified, also sourcedThe industry
Try to prove report 27 wrong: fetch its sources and at least one independent one, and check each load-bearing claim's wording, numbers, revision and dates.
It re-fetched the bill code table and crosswalk and confirmed both of report 27's changes, while correcting its claim that an earlier report had worked from a wrong premise.
13030 states the claim line aggregation and span rules in three separate places, so both named changes are forced by the sources rather than chosen.
3 web searches
An agent
verify:23-six-plans-provider-manuals
Verified, also sourcedThe industry
Try to prove report 23 wrong: fetch its sources and at least one independent one, and check each load-bearing claim's wording, numbers, revision and dates.
It searched the same six manuals for words the report never tried and found one plan's duty to give notice within a business day of suspending a member's services, so the report's replacement sentence fell too.
prints a provider side duty to notify no later than the first business day after it suspends a member's services, so what stands in place of the old duty is not the four member side triggers at all six plans.
1 web search
An agent
verify:20-filing-clock-ltc-ffs
Verified, also sourcedThe industry
Try to prove report 20 wrong: fetch its sources and at least one independent one, and check each load-bearing claim's wording, numbers, revision and dates.
It found and read the rule report 20 could not reach, which confirms the 95-day filing deadline does not apply to long term care, and corrected the report's reading of the federal 365-day rule.
The rule is readable at Cornell LII, the same site the report used for three 26 TAC rules, and its text settles the report's open question 2 in the report's own favour
1 web search
An agent
verify:26-client-demographics-and-season
Verified, also sourcedThis world
Try to prove report 26 wrong: fetch its sources and at least one independent one, and check each load-bearing claim's wording, numbers, revision and dates.
It reproduced the state's regional death rates and member figures, corrected the dates given for the January winter storm, and noted that the July flood declaration came only after the data window had closed.
the proxy carries the direction of the correction (Austin is not the seasonal branch, Tyler's region is highest) but not a magnitude for any weight put on Tyler and McAllen.
and 3 pages not shown
4 web searches
An agent
verify:24-quality-reporting-service-delivery
Verified, also sourcedThe industry
Try to prove report 24 wrong: fetch its sources and at least one independent one, and check each load-bearing claim's wording, numbers, revision and dates.
It confirmed the 17.9 to 89.7 percent spread and the critical-incident finding, and narrowed the report: the state counts members, not hours, and the report quoted the wrong version of the neglect definition.
Holds: both halves survive, since the EQRO's 17.9 to 89.7 percent plan spread is real and a staffing gap is provably not an enumerated critical event
1 web search
An agent
verify:30-mountain-time
Verified, also sourcedThe industry
Try to prove report 30 wrong: fetch its sources and at least one independent one, and check each load-bearing claim's wording, numbers, revision and dates.
It attacked every leg of the time zone ruling at its source and found all of them held, adding small corrections, among them a caution that part of one more county keeps Mountain Time unofficially.
Holds. Every load-bearing leg survived a direct attack on its own source
3 web searches
An agent
critic:round-2
Verified, also pooledThe industry
Read the fourteen round 2 reports and their checks; say which questions closed, where round 2 contradicts the lead's round 1 rulings, and whether a third round is earned.
It found four places where round 2 contradicted the lead's round 1 rulings, two of them resting on premises round 2 had disproved, said how to resolve each, and earned three last questions for a third round.
The outcome survives and the reasoning does not. Keep about 35 and about 70 percent, stop calling the cost report the winner of a three-way comparison
An agent
research:33-span-billing-and-cadence
SourcedThe industry
Search the six plans' manuals and the state's claims manuals for span billing and any required claim cadence for attendant care.
It searched all six plan manuals word by word and found none allows span billing for attendant care and two forbid it, so the world bills one claim line per day and the 95-day lock stays the one clock.
So the money clock and the data clock do not diverge anywhere in this world, provided the generator writes one claim line per client, per payer, per code, per date of service.
11 web searches
An agent
research:35-usage-report-quarter-in-progress
SourcedThe industry
Find whether the state's usage report can be run for the quarter still in progress, by reading the state's numbered public training modules and notices.
It recovered the state's retired training modules from the Internet Archive and found the state's own 2019 training said the report would be available for up-to-date monitoring, so the page had to stop implying an agency is blind inside the quarter.
The enumeration worked and it moved the answer. HHSC published a numbered EVV Tool Kit of at least 17 module PDFs in the same folder where report 32 found module 15; today only module 15 is live on hhs.texas.gov and the rest survive in the Internet Archive.
and 5 pages not shown
5 web searches
An agent
research:34-five-plans-suspension-duty
SourcedThe industry
Search the other five plans' manuals and bulletins for any duty to report a suspension or interruption of a member's services, and whether any ties it to a shift nobody staffed.
It found none of the other five plans carries such a duty, and read the one plan's suspension notice as a copy of a state rule whose ten reasons all happen to the member, never a shift nobody worked.
No plan and no HHSC rule found in this run ties a notification duty to a shift the agency failed to staff
10 web searches
An agent
verify:34-five-plans-suspension-duty
Verified, also sourcedThe industry
Try to prove report 34 wrong: fetch its sources and at least one independent one, and check each load-bearing claim's wording, numbers, revision and dates.
It recounted every hit and ruled that report 34's key reading, that one plan's suspension clause copies a state rule, is an inference no source states, so it must not ship as a fact.
The narrow reading is the better one but is a reading, not a fetched fact.
2 web searches
An agent
verify:33-span-billing-and-cadence
Verified, also sourcedThe industry
Try to prove report 33 wrong: fetch its sources and at least one independent one, check each load-bearing claim, and for each claim that something is absent, search the same documents for the report's words and for others it did not try.
It recounted every word search line by line, found the report's zero-hit claims wrong for two words, and found a Texas plan outside the rosters that allows span billing, so billing by the day became a stated choice for this world.
The per date grain is a decision about this world's six payers, not a statewide fact.
2 web searches
An agent
verify:35-usage-report-quarter-in-progress
Verified, also sourcedThe industry
Try to prove report 35 wrong: fetch its sources and at least one independent one, and check each load-bearing claim's wording, numbers, revision and dates.
It re-downloaded the 2019 module and confirmed the quotes, but ruled the monthly reading an inference no source prints, and a 2019 sentence not to be quoted as a description of today's portal.
Every component is verified, but the conclusion is printed by no source I reached, so it stays an inference and the report's unverified tag is correct.
4 web searches
An agent
critic:round-3
Verified, also pooledThe industry
Read the three round 3 reports and their checks; say what closed, what is still open, and whether a fourth round is earned.
It ruled the three questions closed as far as the page needs, named eight items left open as labelled choices or wording limits, and ruled no fourth round earned.
It should not be re-asked without a new kind of source, meaning a claims data set rather than a manual. Three failed attempts is as settled as an absence gets.
An agent
draft:E
PooledThe data
Draft appendix E of the world brief (the calendar and the measures) from the lead's synthesis and the research reports with their checks, citing a source for every figure and marking anything the research does not supply.
It wrote the calendar (quarters, deadlines, dated events) and one definition for each measure it carried, and listed eleven things no source gives, such as how long a payer takes to send a notice.
How long a payer takes between running a usage review and sending the notice of noncompliance. Nothing states it, so the notice date is a modeling choice (E4, E5)
An agent
draft:B
PooledThe industry
Draft appendix B of the world brief (the rules table) from the lead's synthesis and the research reports with their checks, citing a source for every figure and marking anything the research does not supply.
It wrote the rules table from 22 research reports, each rule with its section and edition, plus what the page may never say and fifteen questions no source answers.
No published claim cadence for long term care fee for service attendant services, or at any of the six plans on the rosters; recorded as an absence three times across two rounds, so the model picks a cadence and the method page names it as a choice
An agent
draft:G
PooledThe industry
Draft appendix G of the world brief (what the page may say) from the lead's synthesis and the research reports with their checks, citing a source for every figure and marking anything the research does not supply.
It wrote the page's surviving sentences, 28 fallen ones each with the reason it fell, the state quotes and reports to name, and 23 prohibitions.
Any real Texas distribution of EVV usage scores: a share of providers below 80, an average, a percentile or a band count. None exists, which is itself the finding and the basis of prohibition 1.
An agent
draft:F
PooledThis world
Draft appendix F of the world brief (the geography) from the lead's synthesis and the research reports with their checks, citing a source for every figure and marking anything the research does not supply.
It wrote the geography reference: the county map source, each branch's licensed counties, the Mountain Time line, and what the map may and may never claim.
This is why the 52 counties are a design choice under a published rule rather than a found fact, and why no county list may ever be attributed to a real agency.
An agent
draft:A
PooledThis world
Draft appendix A of the world brief (the agency) from the lead's synthesis and the research reports with their checks, citing a source for every figure and marking anything the research does not supply.
It wrote the reference document on the modeled agency from nine research reports, recorded ten things no source supplies as labelled choices or open gaps, and flagged a false sentence in the lead's own synthesis.
The share of client and attendant pairs who are relatives. No Texas count exists and over two thirds of states could not report their own.
An agent
draft:C
PooledThis world
Draft appendix C of the world brief (the database design) from the lead's synthesis and the research reports with their checks, citing a source for every figure and marking anything the research does not supply.
It designed the database table by table, with a clock method for each punch, claim lines, holds and rejections, but missed round 3's billing findings, which its reviewer caught.
11. visit (clock method per punch, bill times, transaction fields, schedule optional, service and units, overnight split)
An agent
review:E
VerifiedThe data
Check every figure, quote, section and date in appendix E (the calendar and the measures) against the research reports and their checks, and list each defect with its correction.
It found 16 defects in the calendar and measures, among them a county count off by three, a state rule misread about when a suspension starts, and a deadline counted one holiday short.
Both halves contradict the spine, which is senior to the appendices.
An agent
review:F
VerifiedThis world
Check every figure, quote, section and date in appendix F (the geography) against the research reports and their checks, and list each defect with its correction.
It checked the map document against the research and found ten defects, the worst two plan labels mapped to the wrong real plans, the wrong way round.
The note also contradicts the appendix's own roster table three paragraphs above it, which follows the spine correctly
An agent
review:A
VerifiedThis world
Check every figure, quote, section and date in appendix A (the agency) against the research reports and their checks, and list each defect with its correction.
It found 18 defects in the agency document, among them a claim that Tyler ranks last on every measure (false on three of five) and a rank slip that had got past a research report and its checker.
The recommendation survives; the supporting sentence does not.
An agent
review:G
VerifiedThe industry
Check every figure, quote, section and date in appendix G (what the page may say) against the research reports and their checks, and list each defect with its correction.
It caught a sentence drafted for the page's section on the scoring lag that went beyond every source the research reached, a state notice dated four years late, and a sentence about a batch size that San Antonio's own client count made false.
Section 3's own rule is that the page prints a quote whole or not at all, and this is the one entry that cannot be printed whole
An agent
review:C
VerifiedThis world
Check every figure, quote, section and date in appendix C (the database design) against the research reports and their checks, and list each defect with its correction.
It found the database design still followed round 2's span billing, which round 3 had reversed, misread a state rule as two steps where it is one or the other, and had a column rule that would bar Fort Worth's own case.
The appendix recommends the one cadence the synthesis ruled out, and files a closed question as an open absence.
An agent
review:B
VerifiedThe industry
Check every figure, quote, section and date in appendix B (the rules table) against the research reports and their checks, and list each defect with its correction.
It found 13 defects in the rules table, among them a prohibition that reversed the round 3 critic's ruling and quotes credited to the handbook that come from a withdrawn policy; two of its own calls, one of them about style, were themselves wrong.
The word four belongs to UMCM 8.1 item 63 alone; one plan prints a fifth trigger and five of the six print nothing like the list.
An agent
fix:F
Pooled, also verifiedThis world
Apply the reviewer's corrections to appendix F (the geography), checking each against the research first, and say which were declined and why.
It applied all ten corrections to the map document after checking each at its source, the reversed plan labels among them.
The note now agrees with the roster table above it.
An agent
fix:E
Pooled, also verifiedThe data
Apply the reviewer's corrections to appendix E (the calendar and the measures), checking each against the research first, and say which were declined and why.
It applied all sixteen corrections, re-walking the deadline count by hand and defining three measures the draft had left out.
Re-walked the count from Thursday 2026-06-18: the tenth business day is Friday 2026-07-03 unless the observed holiday is skipped too, so 2026-07-06 requires both
An agent
fix:A
Pooled, also verifiedThis world
Apply the reviewer's corrections to appendix A (the agency), checking each against the research first, and say which were declined and why.
It applied all 18 corrections to the agency document, each checked against the research first, and reworked the attendant count so the roster and the weekly hours agree.
The recommendation is kept and the four grounds are now a list rather than one run-on sentence.
An agent
fix:G
Pooled, also verifiedThe industry
Apply the reviewer's corrections to appendix G (what the page may say), checking each against the research first, and say which were declined and why.
It applied every correction to the page sentences, rewrote the sentence on the scoring lag to the instruction the state's job aid actually gives, and showed the reviewer was wrong that the rules appendix did not exist yet.
the fallen February to May drafting is recorded with the lock arithmetic that kills it
An agent
fix:C
Pooled, also verifiedThis world
Apply the reviewer's corrections to appendix C (the database design), checking each against the research first, and say which were declined and why.
It rewrote the database design to round 3's one claim line per date, made the two compliance steps alternatives, and declined one correction after showing the reviewer had two documents the wrong way round.
Defect 1, the span billing track rewritten to round 3.
An agent
fix:B
Pooled, also verifiedThe industry
Apply the reviewer's corrections to appendix B (the rules table), checking each against the research first, and say which were declined and why.
It applied the reviewer's corrections to the rules table but kept one claim the reviewer wanted cut, after finding the reviewer had misread the research (the checker's own correction said most penalties sit between 500 and 700 dollars), and declined one style flag raised in error.
Removing it would discard the verify's own correction.
A message of Ben's
A check of the lead's
Gearing up. I'll prove the current pipeline is green in the isolated checkout, cut the build branch there, and write the two documents the next session starts from: the build plan and the carryover.
A message of Ben's
A message of Ben's
An agent
B1 contract
PooledThis world
Build the base of the new modeled world from the brief: the settings file, the 31-table database, the frame of the simulation with places for the branch stories to plug in, and the clients, attendants and staff.
It built the settings for the seven branches and payers, the 31-table database and the clients and attendants, citing an appendix row for every published figure and marking every guess ASSUMED.
Every published figure cites its appendix row; every open figure is marked ASSUMED in the code and listed in the handoff.
An agent
B2 hours and visits
PooledThis world
Build the authorized hours and the visits day by day, deriving every consequence by rule from causes already in the tables.
It built the authorizations, absences and about 252,000 visits, each hand-keyed or missed visit following from a cause in the tables, and flagged that the device rule another builder wrote was narrower than the state's test.
Nothing in either module names a branch: El Paso's shortfall (86.7 against 89.2) arrives from its own replacement lag and turnover in params.ts.
An agent
B3 pipeline claims events
PooledThis world
Build the prebilling hold, the nightly export, the claim cycle and the dated events on top of the hours and visits.
It built the export, claims filed one line per day with the filing deadline beside the lock, a small recoupment sample never projected, and the storm, flood and plan-switch events, and flagged the baseline score as too high.
a small recoupment sample priced to the line and never extrapolated
An agent
B4 validator
Verified, also pooledThis world
Write the validator, the program that checks the generated records: its 23 numbered checks, six time-zone checks, a file for the stories' checks, and a branch summary table printed first.
It wrote the checks the world must pass, added a third verdict for things that cannot be asserted, and reported that the Mountain Time check had nothing to test because El Paso had no overnight shifts.
a PASS that sometimes means a query proved this and sometimes means nobody checked is worse than a third word.
An agent
critic
VerifiedThis world
Read-only: find defects in the four builders' world against the brief's appendices, running the generator and validator and querying the tables yourself.
It queried the generated tables itself and filed 33 defects, six blocking, among them weather absences merged into half-year spells and checks that passed with nothing to check.
A four-day storm in January and a three-day flood in July are two closures, not a six-month one.
An agent
fixer
PooledThis world
Fix every blocking and should-fix defect the critic filed, then re-run the checks.
It worked through the critic's defects, reported all six blocking ones and every should-fix one it counted as fixed, re-ran the checks (72 pass, none fail), and left the never-list change to the lead because that ruling was his.
the fourth (splitting the never list into whole-label and substring halves) is not taken because it changes what the re-identification guard checks and appendix A section 6 is the lead's ruling, not a builder's
A check of the lead's
W1 finished: six agents, about three hours, no errors. Before I trust any of it, I'll pull the critic's verdict and the fixer's gate from the result, then run the gate myself in the build worktree and read the branch table.
An agent
W2a integrator
Pooled, also verifiedThis world
Prepare the modeled agency so five story builders could work at once: tune each branch's normal clock-in level to the brief, add the lead's first four rulings, and set up empty story files with their checks.
It tuned all seven branches to within two points of the brief's figures, built the story files, and flagged four places where the brief or the lead's rulings could not all be true.
Two things need the lead: Austin's "Plan C contract at 87" cannot be tuned by any lever L1 created and reads to me as a Q4 pace figure rather than a stored cell (it sits at 91.7), and the spouse question is the one place ruling L3 and appendix A 3.3 point in opposite directions
An agent
Probe isolated worktree layout
Built
Run ten git and shell commands in an isolated copy of the project and report the raw output, including whether a folder link can be made without administrator rights.
It showed the isolated copy started from an old commit with no installed packages, and that a folder link to the packages works without administrator rights.
[Result: Junction creation and removal works without admin rights]
An agent
story houston
Pooled, also verifiedThis world
Build Houston's story in its own copy: the scheduler leaves, nobody keys new authorizations for nine weeks, and visits pile up held.
It built the empty seat and the held visits exactly as written, and showed the brief's failing score could not come from that cause in this world, where only 42 authorizations fell due in the nine weeks against the brief's 140, instead of turning a dial to fake it.
NOT landed, and this is the headline: Q2 86.6 against a target of 76, Q3 89.9 against 81, Q4 91.8 against 84, and the Plan A and Plan B Q2 cells at 84.8 and 86.2 against "under 80". That is not a tuning miss.
An agent
story el-paso
Pooled, also verifiedThis world
Build El Paso's story in its own copy: homes left without an attendant, so the branch passes its grade while delivering less care.
It built the unstaffed hours and landed the hours and money on target, and reported two sizes it could not reach without a change to shared code, rather than fudging them.
Two appendix D sizes cannot be landed without a shared-module change and are reported rather than fudged.
An agent
story fort-worth
Pooled, also verifiedThis world
Build Fort Worth's story in its own copy: hiring classes of new attendants who clock in on the app and forget to clock out.
It built the hiring classes and the missed clock-outs, and measured that the brief's two scores for the branch could not both come from that cause.
So appendix D's Q4 81 sits BELOW the ceiling of its own mechanism and its Q3 87 sits ABOVE it: there is no setting of any new-hire cause, mine or anybody's, that lands both
An agent
story tyler
Pooled, also verifiedThis world
Build Tyler's story in its own copy: a spring review removes landlines from 20 homes, nobody orders the devices that would replace them, and the attendants there have no way to clock in.
It built the twenty homes and the unordered devices, and showed the brief's eight attendants could not cover twenty homes without booking one person in two homes at once, so it used one attendant per home.
THE BIG ONE. The cohort is twenty attendants, one per home, not appendix D's eight. This world cannot hold eight for twenty
An agent
story san-antonio
Pooled, also verifiedThis world
Build San Antonio's story in its own copy of the project: 26 clients switch plans, the office's EVV setup keeps the old plan, and every visit sent for them is rejected. Set causes only, never results.
It built the cohort, its switch dates and its rejections as the brief described, and showed that 26 clients out of 300 could not pull the branch's score down as far as the brief said, so it refused to enlarge the group to hit the number.
The cohort was not grown to chase the score targets. Appendix D's own preamble says targets are targets, and the spine's R3 rules the cohort 'a trickle'
An agent
integrator
PooledThis world
Merge the five branch stories into the shared build, decide which shared-code requests to build, and run the full gate.
It merged all five stories without conflict, built three shared rules and declined four, and reported that three builders had independently found the brief's scores could not come from its own causes.
The six that do not are one finding, reached independently by three builders: the mechanism appendix D states cannot carry the score appendix D states at the population appendix D states.
An agent
critic
VerifiedThis world
Read-only: check each merged story against the brief, its checks and the other branches, running the generator and its own database queries but editing nothing.
With its own database queries it found 24 defects, among them a brief sentence that overstated El Paso's recognised interruptions about sevenfold and a Monday queue that ran every day of the week.
The check asserts only the unstaffed share and prints the member-side hours in its detail, so it passes at 56.5 while the sentence appendix D authorises is wrong by a factor of seven.
An agent
fixer
PooledThis world
Fix every serious defect the critic found, in the story or in shared code as a rule and never as a tuning dial, re-run the gate, and record each fix or refusal.
It handled 21 of the critic's 24 findings with three new shared rules, brief amendments and notes for the lead, and declined three: two that were the lead's to rule and one small one left for later.
El Paso's honest split is rebuilt as four exhaustive terms the check asserts add back to the shortfall (56.5 percent nobody staffed, 22.3 percent lost inside visits that happened, 14.6 percent never put on a calendar, 6.6 percent called off)
A check of the lead's
Appendix D's score targets for three branches (Houston 76 / 81 / 84, San Antonio 86 / 78, Fort Worth 81) were the lead's own dramatisation, written before the mechanisms had been run at the ruled sizes.
An agent
W2c integrator
Pooled, also verifiedThis world
Build the lead's rulings: one graded contract per plan per service area, the rule that makes Houston's failed quarter reachable, and hiring that starts warm; then amend the brief to the landed figures.
It made the state's 24 contracts computable from the tables, reached Houston's failed quarter on Plan A (77.2), closed El Paso's first-quarter miss, and rewrote a Houston check that no longer tested anything.
L12 turned scored_identity into 24 rows, one per payer per service delivery area, and gave visit_transaction and claim_attempt the contract each row went out under, so the state's 24 contract cells are computable from the tables with no branch anywhere in them
A message of Ben's
A message of Ben's
An agent
critic
VerifiedThis world
Read-only: check the three new rules and every sentence the brief amended, against the tables.
It recomputed the brief's figures from the tables and found a dozen wrong or stale ones, among them El Paso's quarter shortfall given as the whole window's.
3,301 is El Paso's SFY26 Q4 shortfall, not the window's. I recomputed the measure the appendix itself defines (authorized with suspension days removed, minus delivered bill hours): window authorized 66,635.5, window delivered 53,543.3, window shortfall 13,092.3 hours.
An agent
fixer
Pooled, also verifiedThis world
Fix the critic's findings, re-measuring every figure before writing it, and re-run the gate.
It re-measured every figure the critic asserted (all reproduced), fixed 14 of 15 findings, refused to widen a tolerance the lead had set, and found one more stale summary figure itself.
I re-measured every figure the critic asserted before writing it, and all five reproduce exactly, including a re-run of the Houston story-off probe
A check of the lead's
The gate is green in my own hands (typecheck, site tsc, the same hash, validate all passing, clean tree).
An agent
W3a analyze port
PooledThe data
Split the program that works out every figure into seven groups and move every old figure onto the rebuilt database under the same name.
It moved every old figure across with none lost, and left the lead three questions, among them two honest delivered-care figures that differ and ten rejection codes with no published text.
A key tree diff against the committed v1 analysis.json lists 0 keys lost and 69 added, so nothing types.ts reaches is missing and nothing had to be stubbed for W4.
An agent
W3a reconcile port
VerifiedThe data
Rebuild the second program (a Python program that works every figure out again without reading the first one) on the rebuilt database, without reading the first program's code.
It rebuilt the second program without reading the first program's code; all 9,416 figures agreed, and it pointed out that the written rule rounds money's hours down to the quarter hour where the old code rounded to the nearest, a difference that now shows on the rebuilt data.
The gate is green: 9,416 figures compared, every one agrees, exit 0, against v1's 8,165 on v1's world. There are no disagreements between the analysis and the twin.
An agent
critic
VerifiedThe data
Read-only: check both ported programs, running them on a scratch copy of the data.
It worked out seven key figures from the raw tables with its own queries (all matched) and found a gap that would have dropped figures, money that had silently changed basis, and a check that could not fail.
So in place of adjudicating disputes I recomputed seven load-bearing figures independently, from the RAW v2 tables only (visit, visit_transaction, rejection_code, client, payer, attendant, quarter, meta), writing my own SQL from the handbook rules
An agent
fixer
Pooled, also verifiedThe data
Fix the critic's findings on both programs and settle any disagreements, then re-run everything.
It fixed the one blocking gap and all five of the critic's should-fix findings, re-measured the money split on its own, and ended with 9,623 figures agreeing.
Measured independently of the critic and of the analysis: of 3,270 on-the-clock visits and $186,341, 1,579 visits and $89,066 (48 percent of the money, 5,094.5 of 10,635 hours) are priced from planned_hours
An agent
ANALYZE g-map
Pooled, also verifiedThe data
Build the map group in its own copy: write the figure list first, then store each branch's county, service area and clock.
It stored 407 map figures and found that the shared rule for turning shares into words would call 'all' of something 'nearly all'.
One real defect was found and worked around inside the group: spoken in shared.ts flags a share of exactly 1 as nearly_all, and four branches are licensed in EVERY county of their area, so the page would have said "nearly all" of all four
An agent
ANALYZE a-score
PooledThe data
Build the score group of figures in its own copy: write the figure list first, then store the 24 graded contracts, the pace, the finish for every possible fix day and the notice state.
It stored 12,270 figures and measured, for the first time, that Tyler's two graded contracts must be fixed by Aug 3 and Aug 4 while the branch's own blend says Aug 13.
Ruling L19 now has its measurement: at the grain the state grades, Tyler's two contracts carry fix-by dates of 2026-08-03 and 2026-08-04 where the branch blend says 2026-08-13, and San Antonio's Plan B contract in Bexar has no day left at all
An agent
ANALYZE c-money
Pooled, also verifiedThe data
Build the money group in its own copy: write the figure list first, then store the four kinds of money, the deadlines, the rejections and the audit exposure.
It stored 4,108 money figures and warned that the brief's audit rule, applied word for word, would put about $1.79 million at risk where the narrow reading gives zero.
the lock binds first or ties on all 206,011 claim lines and is later than the filing deadline on none, which is E14's sentence measured rather than asserted
An agent
ANALYZE f-calendar
Pooled, also verifiedThe data
Build the calendar group in its own copy: write the figure list first, then store the fiscal year, the data window, the quarters and their lock dates, the dated events and the respiratory season.
It stored the calendar and found that the season's branch figures match the project's own brief, on which branches the respiratory season hits hardest, only as a rate per 100 clients; as raw counts they say the opposite.
the season footprint carries clients in care and a per hundred clients twin, because read as raw counts it says the opposite of E8 and read as a rate it says what E8 says
An agent
ANALYZE d-workforce
Pooled, also verifiedThe data
Build the workforce group in its own copy: write the figure list first, then store turnover, vacancy, the replacement lag, the tenure panel, the pair types and the concentration of hand-typed work.
It stored 3,559 workforce figures and warned that the model's part-time and full-time turnover runs opposite to the state's published gap.
The part time and full time turnover split does not reproduce the published gap, and the page must not set the two side by side as though it did.
An agent
ANALYZE e-stories
Pooled, also verifiedThe data
Build the stories group in its own copy: write the figure list first, then store every figure the brief's five stories print.
It stored 3,682 story figures, each defined by its cause rather than by a branch name, and found brief figures that the tables do not give.
The shaping decision: no group module may name a branch id, so not one figure is defined as San Antonio's or Tyler's.
An agent
ANALYZE b-delivered
PooledThe data
Build the delivered-care group in its own copy: write the figure list first, then store delivered against authorized hours by week, month and quarter, and the four-part shortfall.
It stored delivered care against the authorization, reproducing the brief's El Paso figures exactly, with the typing-in backlog kept beside the hours rather than inside them.
Appendix D section 5's four-part split is stored at branch and quarter, and El Paso's quarter in progress reads 53.7, 24.9, 14.5 and 6.9 in 100 with the four adding back to the 3,300.93 hour shortfall exactly
An agent
RECONCILE g-map
VerifiedThe data
Work out every map figure again in Python from the figure list, the brief and the tables, without reading the first program's code, and write up every disagreement.
It agreed on all 327 map figures and then proved its checks could fail by breaking 35 stored figures one at a time; it caught all 35.
Because a twin that agrees on its first run proves nothing, the report was made falsifiable: 35 stored figures were broken one at a time in a copy of the analysis and the twin caught all 35
An agent
RECONCILE f-calendar
VerifiedThe data
Work out every calendar figure again in Python from the brief and the tables, without reading the first program's code, and write up every disagreement.
It agreed on all 1,028 calendar figures on the first run and flagged that the word rule and one stored figure had no written definition to check against.
Nothing disagrees, and no comparison failed on the way: every one passed on its first run.
An agent
RECONCILE d-workforce
VerifiedThe data
Work out every workforce figure again in Python, without reading the first program's code, and write up every disagreement.
It found 57 disagreements in two definitions and traced one to a date quirk that put 20 visits in the wrong tenure band.
Every visit that moves belongs to an attendant hired on the 29th, 30th or 31st of a month, and every one falls in the first days of the month the anniversary lands in.
An agent
RECONCILE b-delivered
VerifiedThe data
Work out every delivered-care figure again in Python, without reading the first program's code, and write up every disagreement.
It agreed on all 5,502 figures and reproduced the brief's El Paso figures with no sight of the first program.
The twin reproduced appendix D's own figures with no sight of the analysis: El Paso 82.2 / 80.7 / 76.2, shortfall 3,300.93 hours, the four parts 53.7 / 24.9 / 14.5 / 6.9
An agent
RECONCILE a-score
VerifiedThe data
Work out every score and projection figure again in Python from the written definitions, without reading the first program's code, and write up every disagreement.
It found 627 disagreements, all in the quarter's projection, and traced 567 of them to one unclear sentence in the brief and the other 60 to a second unclear clause; the gap left Tyler's fix-by date a day apart between the two programs.
Every one of the 627 disagreements is inside the forward walk, and they are five readings of one sentence, not 627 separate defects
An agent
RECONCILE c-money
VerifiedThe data
Work out every money figure again in Python, without reading the first program's code, and write up every disagreement.
It agreed on all 3,636 money figures, then changed ten stored figures on purpose to prove its checks could fail; all ten were caught.
Because an all-PASS first run is exactly what a toothless comparison also looks like, ten stored figures were altered one at a time in a scratch copy of analysis.json, one per family, and all ten produced a FAIL row naming the figure.
An agent
RECONCILE e-stories
VerifiedThe data
Work out every story figure again in Python from the figure list and the brief, without reading the first program's code, and write up every disagreement.
It wrote the second program's checks for 3,431 new story figures (8,935 in its group, counting the 5,504 checked earlier) and found the first program storing a wage index of 0 at all seven branches, plus a Tyler total added up from figures already rounded.
as stored it says every branch in Texas pays its aides nothing. A reader who knows the DFW market would stop there.
An agent
settle d-workforce
Verified, also pooledThe data
Re-read the written definitions and settle the twin's disagreements with the first program, fixing the first program where it was wrong; send the rest to the lead.
It agreed the twin was right about the date quirk, fixed the first program so exactly 20 visits changed band, and left the undefined headcount for the lead.
Measured before and after, exactly 20 delivered visits of 242,134 change band, 9 in SFY26 Q2, 10 in Q3 and 1 in Q4, every one from under_6_months to 6_to_12_months, and none at the twelve month boundary.
An agent
settle e-stories
Verified, also pooledThe data
Re-read the written definitions and settle the twin's disagreements with the first program, fixing the first program where it was wrong; send the rest to the lead.
It confirmed both of the twin's findings, found the true cause of the zero wage index, fixed both, and warned that the same silent-zero trap was open to every group.
The wage line, 21 figures: the twin's diagnosis of an integer coercion was wrong, and the finding itself was exactly right.
An agent
settle a-score
VerifiedThe data
Re-read the written definitions and settle the twin's disagreements with the first program, fixing the first program where it was wrong; send the rest to the lead.
It argued, clause by clause, that the twin had misread the brief's projection sentence, and narrowed the rest to one question for the lead.
The twin's kept reading divides a week's posted rejections by an agency wide posting curve from sql/a-score/02-settle-time.sql, which is not a repeat count, is not per visit, is not the branch's own
An agent
W3c integrator
PooledThe data
Merge the seven groups' work, run the whole gate, apply the shared changes the lead had ruled, and record what stayed in dispute.
It merged all seven groups and was part way through the shared changes when a Windows Update restart killed it about twenty minutes in.
Now strip the local constants and add the imports.
A message of Ben's
A check of the lead's
I'll check the repo state, the build worktree, the carryover doc, and the Windows event log for what happened overnight.
A check of the lead's
Last check: the unattended-runs memory claims "the machine never sleeps." That was true, but it missed update restarts. I'm reading it so I can correct it.
A check of the lead's
The failing counts differ from what W3b reported (627 in a, 7 in d), so treat the half-finished edits as unverified.
An agent
W3c integrator (resumed)
Pooled, also verifiedThe data
Pick up exactly where the killed run stopped: check its unfinished edits line by line, run the gate, apply the size budget, and commit the data files.
It checked the killed run's edits line by line, finished two that were half done, and landed the merged figures (27,757 of 28,390 agreeing, the rest waiting on the lead's rulings).
I checked the killed run's uncommitted shared edits line by line. Two were only half applied and are now finished: groups a and f still had their own copy of the spoken ladder, and no twin outside group g compared the new all flag.
An agent
rulings fixer
Pooled, also verifiedThe data
Apply each of the lead's rulings exactly as written, rewrite the brief's projection definition first, and move the twin to its new words.
It applied or confirmed every ruling, rewrote the brief's projection definition, and brought all 28,507 figures into agreement.
L45 was the big one. Appendix E9 was amended first, marked as the lead's ruling L45.
An agent
critic
VerifiedThe data
Read-only: check every stored figure against its written definition and the brief's list of printed figures, the honesty rule, the size cap, repeatability and the second-program checks.
It changed stored figures on purpose and found whole families that no second-program check covered, although the report said every figure agreed, plus figures the page needs that were cut or never stored.
Mutation test run by the critic: each family was perturbed and the whole twin rerun, and every one of these passed unchanged, so no twin compares them.
An agent
fixer
Pooled, also verifiedThe data
Fix the critic's findings, add the missing second-program checks and stored figures, and re-run everything.
It fixed all five blocking findings and 15 of the 18 should-fix ones (declining three, with reasons), added checks until 37,419 figures agreed, and proved the new checks could fail by changing 8 values on purpose.
A mutation run on a scratch copy changed 8 values across the new twin families, and all 8 were caught.
An agent
W4 fixtures
Pooled, also verifiedThe data
Rebuild the test states (dated rewinds of the data) on the new figures, add a state for a notice past its deadline, and make the script check its own output.
It rebuilt six test states and found that 29 page tests had been passing only because stale old files were on disk.
The other 29 are v1 expectations in the kit tests running against the new fixtures; they passed before only because stale v1 fixture files were on disk.
A check of the lead's
My gate at landing: both typechecks and validate green, analyze twice byte-identical, reconcile agreeing on every one of 37,419 figures
An agent
W4 kit
Pooled, also verifiedThe data
Make the page kit read the new stored figures, rewording only sentences whose meaning changed, with no redesign.
It moved the kit onto the new figures with every test passing and caught a page sentence that said every contract was above the line when three were projected under.
M17 ContractLine's sentence now reads the contract cells the state grades. Its row still shows the payer blends. The old sentence said every contract was above the line in close.json while three cells were projected under.
An agent
critic
VerifiedThe data
Read-only: check the gates, the honesty rule in the kit, and that labels changed only where a figure's meaning changed.
It found the first screen still telling readers that 26 clients had been on the wrong plan since Jan 1, when only 2 switched that day.
On the version 2 world that date is only the first of four switch dates: 2 clients on Jan 1, then 12, 8 and 4 (story.stale_payer.by_switch_date).
An agent
fixer
PooledThe data
Fix the critic's findings in the kit and re-run the gate.
It fixed all four of the critic's main findings and its seven small ones, rewording every sentence that treated the first switch date as everyone's.
Every remaining sentence that treated enrollment_changed_on as the day all clients switched now says "the first of them since", "since their switch" or "since each client's plan change".
An agent
W5 map
Pooled, also sourcedThis world
Redraw the Texas map from the Census Bureau's 2025 county boundary file with one projection, and store and double-check every map figure.
It downloaded the Census county file, redrew the map's 254 counties, confirmed that all 52 county codes typed by hand earlier matched the file, and tightened a loose check.
branch_county's 52 codes, the home counties and the four D-SNP counties are read from the file by name. Every one equals the value typed by hand before W5, so the world's hash did not change.
An agent
critic
Verified, also sourcedThis world
Read-only: check that every map figure has a second-program check, the frame matches the counties, no county belongs to two branches, and no client sits outside its branch's counties.
It downloaded the Census file itself, read it with its own reader, and found every outline point within about half a pixel; nothing rose above a small nit.
Downloaded cb_2025_us_county_500k.zip myself: 11,758,981 bytes, sha256 aa976c00...777c, equal to the header. Read it with my own struct-based shp and dbf reader, not the generator's and not pyshp.
A message of Ben's
A check of the lead's
No, it wasn't. That page was frozen at 2026-09-21: it still asked you to judge Questions A and B, which you ruled that day
A message of Ben's
An agent
W3d builder
Pooled, also verifiedThe data
Store and double-check the last figures the contract-first screen needs, fix two findings from the test states, clear the earlier nits, and make every stored figure name the query it comes from.
It stored the new figures, fixed the passed-deadline wording and the quarter-end crash, and recorded where each figure comes from, though its critic found 28 dataset figures pointing at no query a reader could audit and six pointing at the wrong one.
every path in analysis-keys.json now says where it comes from. 2,804 of 3,787 paths carry a query id, and three new checks hold the record to queries.json and to itself.
An agent
critic
VerifiedThe data
Read-only: check every new figure against its definition and its twin, the honesty rule, the size cap, repeatability, and the record of where each figure comes from.
It walked every path in the source record and found 560 stored values that name neither a query nor a computation: 522 are stated constants, sentences and labels, and 28 are dataset values read by statements outside the query list, which no reader could audit.
They are read by inline statements that are not in queries.json, so the provenance record names no query a reader can audit.
An agent
fixer
PooledThe data
Fix the critic's findings and re-run everything.
It gave the unaudited figures real queries, corrected the wrong sources, and ended with 39,050 figures agreeing.
Blocking: the 38 paths that came straight from the dataset now name a query.
An agent
S5f data
Pooled, also verifiedThe data
Store the few contract-by-branch figures the new first screen needs, with second-program checks, and nothing else.
It stored 23 new keys, checked them a second time, proved that check could fail with six deliberate edits, and warned that a planned sentence would be false for passing contracts.
A mutation test with six edits failed five comparisons (16 figures), and the file was restored.
An agent
A6 builder
PooledThe data
Build comp A6 in the lab: the county-true territory map two ways and alone, with the contracts the state grades leading.
It built the map comp with the graded contracts leading and every date read from stored figures, and kept the blend's date off the map card because it argued with the contracts' dates.
The callout lists the contracts the state grades at the picked branch. Each contract's day comes from its stored date, or from the cell's stored recovery word read by id ('No fix day clears it'). The branch's own blend sits on its own line, labelled as ours.
An agent
critic
VerifiedThe data
Read-only: check comp A6's shots, words and data against the work order and the stored file, as Ben would read them.
It read the shots as Ben would and caught Tyler's action step printing the blend's last safe day (Aug 13) where the graded contracts' days are Aug 4 and Aug 3.
Aug 13 is the blend's stored projection.must_be_fixed_by. The state-graded contracts' stored days are Aug 4 (Plan A) and Aug 3 (Plan B).
An agent
fixer
Pooled, also verifiedThe data
Fix the critic's findings on comp A6, re-shoot, and record each fix or refusal.
It relabelled the blend's date and fixed most findings, and itself spotted a second step with the same wrong-date problem that the critic had missed.
The page still prints the stored date and chooses no minimum date.
An agent
build:strip-and-tyler
PooledThe data
Build two of the lead's rulings: the strip under the map follows each branch's stored contract status with contract words, and Tyler's name plate moves above its pin.
It made the strip follow the stored contract status with new contract words, proved the old strip was unchanged byte for byte, and moved Tyler's name plate.
Each badge's status is the stored contract_status. The mark and colours come from the same data-status and STATUS_SHAPE path A6's pins use, so there is no second mapping.
An agent
critic:strip-and-tyler
VerifiedThe data
Read-only: check the round's code, tests and shots for correctness and honesty.
It found the round clean apart from a test that claimed more than it checked and a code comment that was untrue in one browser case.
The round is clean apart from two nits: a stale comment in A6's view.tsx and a test whose name claims more than it checks.
An agent
fix:strip-and-tyler
Built
Fix the critic's two nits and re-run the gate.
It widened the test to every mount in every file, proved it fails on a planted mount, and corrected the comment.
As a check, I added a grain to the workbench's second mount and the test failed. I then reverted that edit with git checkout.
A message of Ben's
A check of the lead's
Yes. I checked the stored file rather than assume it, and the pool is the same for all seven.
A message of Ben's
An agent
mine:FTW
Pooled, also verifiedThe data
Using only stored figures, write what Fort Worth's page should say: where it stands, why, the insight, and actions with owners and worth, every claim beside its figure.
It wrote Fort Worth's content and warned against blaming new hires without the count of settled staff, and against fix-day finishes that would print a worse number for fixing.
Fort Worth passes all three of its contracts, and the points it loses go to visits typed in by hand, 70 in 100 of them because the attendant clocked in on the app and never clocked out, while its attendants in their first six months type in more of their visits every quarter.
An agent
mine:AUS
Pooled, also verifiedThe data
Using only stored figures, write what Austin's page should say: where it stands, why, the insight, and actions with owners and worth, every claim beside its figure.
It wrote Austin's content and found the brief undersold Austin: 116 recoupment notices on its winter claims, which the grade never shows.
The brief undersells Austin. "Passing with depth, nothing needs fixing" is true of the grade and false of the branch.
An agent
mine:COMPANY
Pooled, also verifiedThe data
Using only stored figures, write what the company-wide page should say: where it stands, why, the insight, and actions with owners and worth, every claim beside its figure.
It wrote the company-wide content and warned of places where a stored field would make a sentence false.
21 of our 24 payer contracts will pass this quarter, Tyler's two can still be saved if the fix lands by Aug 3, San Antonio's Plan B cannot be saved this quarter, and Houston owes its notice training by Jul 20.
An agent
mine:TYL
Pooled, also verifiedThe data
Using only stored figures, write what Tyler's page should say: where it stands, why, the insight, and actions with owners and worth, every claim beside its figure.
It wrote Tyler's content and showed that a manager working to the blend's later date would pass the blend and fail both contracts the state grades.
Tyler will fail two of the three payer contracts the state grades there unless the 20 alternative devices nobody ordered after the spring landline review are in use by Aug 3.
An agent
mine:HOU
Pooled, also verifiedThe data
Using only stored figures, write what Houston's page should say: where it stands, why, the insight, and actions with owners and worth, every claim beside its figure.
It wrote Houston's content and found the brief wrong to say Houston failed its winter quarter: the branch's own blend passed while three of its four contracts failed.
The branch blend's stored Q2 score is 79.67, rounded 80, and the pass mark before rounding is 79.5, so by our own blend Houston PASSED Q2. The failure belongs to three contracts (Plan A 77, Plan B 76, Plan E 77).
An agent
mine:SAT
Pooled, also verifiedThe data
Using only stored figures, write what San Antonio's page should say: where it stands, why, the insight, and actions with owners and worth, every claim beside its figure.
It wrote San Antonio's content and found three older claims the stored figures contradict: the 'since Jan 1' date, the blend still passing, and the kit's first lock date.
payer_mismatch.enrollment_changed_on is 2026-01-01, the EARLIEST switch date, so "26 clients on the wrong plan since Jan 1" is false for 24 of them.
An agent
mine:MCA
Pooled, also verifiedThe data
Using only stored figures, write what McAllen's page should say: where it stands, why, the insight, and actions with owners and worth, every claim beside its figure.
It wrote McAllen's content and flagged two hour measures that disagree by 4 points and a brief claim about McAllen's wages that the stored figures contradict.
Two hour denominators live side by side at McAllen and disagree by 4 points.
An agent
mine:ELP
Pooled, also verifiedThe data
Using only stored figures, write what El Paso's page should say: where it stands, why, the insight, and actions with owners and worth, every claim beside its figure.
It wrote El Paso's content: the branch passes its grades while delivering 76 of every 100 authorized hours, and 13 of every 100 had no attendant to send.
El Paso passes all three contracts the state grades, yet it delivered 76 of every 100 authorized hours this quarter, and 13 of every 100 had no attendant to send.
An agent
verify:FTW
VerifiedThe data
Refute what you can in the Fort Worth miner's content, check every figure against the stored file by its path without doing arithmetic, and return the corrected content.
It found every figure matched its stored value and refuted the meaning of several claims, among them an attendant wrongly called a spring hire and a misquoted rule on when a payer may approve reopening a locked visit.
Every value the miner cited matched its path exactly. The refutations are about meaning, wording and derived claims.
An agent
verify:TYL
VerifiedThe data
Refute what you can in the Tyler miner's content, check every figure against the stored file by its path without doing arithmetic, and return the corrected content.
It found every figure matched and refuted claims built on the miner's own sums, a wrong year, and unstored roles and rankings.
Every path in the miner's ladder resolves in analysis.json with the value it gave.
An agent
verify:MCA
VerifiedThe data
Refute what you can in the McAllen miner's content, check every figure against the stored file by its path without doing arithmetic, and return the corrected content.
It found every value matched and refuted claims the figures did not bear out, among them a Plan C slide and a comparison of two different measures.
All paths were resolved against data/home-care-ops/analysis.json at world-v2 f789735 and their how_we_know read from analysis-keys.json. Every value the miner cited exists and matches.
An agent
verify:HOU
VerifiedThe data
Refute what you can in the Houston miner's content, check every figure against the stored file by its path without doing arithmetic, and return the corrected content.
It found most values exact and refuted a claim about 83 authorizations, a wrong path and a misleading tenure claim, among others.
Verified every path against analysis.json at world-v2 f789735 and every how_we_know against analysis-keys.json; most of the miner's values were exact.
An agent
verify:SAT
VerifiedThe data
Refute what you can in the San Antonio miner's content, check every figure against the stored file by its path without doing arithmetic, and return the corrected content.
It found every figure matched and refuted claims the figures did not support, including a rule sentence that exists nowhere in the stored rules.
Every path exists and every value matches. The defects are in what the figures were made to say, not in the figures.
An agent
verify:AUS
VerifiedThe data
Refute what you can in the Austin miner's content, check every figure against the stored file by its path without doing arithmetic, and return the corrected content.
It found the figures matched, apart from two rule sentences the miner had paraphrased, and refuted several claims, including a new-hire trend that rises only by construction and a visit counted twice.
Verified every figure path the miner cited against analysis.json at world-v2 f789735 with a resolver, and each how_we_know against analysis-keys.json.
An agent
verify:COMPANY
VerifiedThe data
Refute what you can in the company miner's content, check every figure against the stored file by its path without doing arithmetic, and return the corrected content.
It found all 218 figures matched and refuted a false claim about new attendants, a wrong base for El Paso's action, and several overreaches.
All 218 figures the miner cited resolve in analysis.json at the stated path with the exact value, and every how_we_know now uses analysis-keys.json's spelling.
An agent
verify:ELP
VerifiedThe data
Refute what you can in the El Paso miner's content, check every figure against the stored file by its path without doing arithmetic, and return the corrected content.
It found all 207 figures matched and corrected wording the stored figures do not support, including three 'about's on exact figures.
All 207 of the miner's figures resolve at their cited paths in data/home-care-ops/analysis.json (world-v2, f789735) with exact values and a how_we_know that matches analysis-keys.json; none was invented.
An agent
write:spine
Pooled, also verifiedThe data
Assemble the eight checked ladders into one content spine, as a document and a text-only page in the judging room, without retyping any figure.
It assembled the spine from the skeptics' versions as returned, corrected three sentences where two skeptics had misread a stored measure, and merged 135 requests for missing figures into 47.
I used the eight verified ladders exactly as the skeptics returned them. I pulled them from the workflow journal (wf_98b1d2b2-224), so I did not retype any figure.
An agent
build:a7-rows
Built
Build comp A7 (a trial design of the dashboard's first screen): all seven branches as rows with the same five cells, the map as the selector, every cell a stored value, and shoot it at 2560 wide.
It built the rows design, wired the map to the rows, added 39 tests and fixed its layout over four rounds of screenshots, committing nothing.
The five column heads are printed once over them. Every row is the same five cells, in the same type, at one height.
A message of Ben's
An agent
build:s6c-keys
Pooled, also verifiedThe data
Store the ten new figures the first screen needs, each with a twin in the second program (the separate Python program that works every figure out again), committing each item.
It stored and twinned five of the ten items (display strings, direction words, each branch's place among the seven, first unpaid locks, Tyler's order-by date) and was part way through the sixth when the session ended.
The rule is stated once at method.rank_rule: 1 is the largest, ties go to the branch earlier in headline.branch_order. The twins recount each figure and rank it themselves: 758 figures, every one agrees.
An agent
build:a7b-cards
Built
Build comp A7b: the same seven-branch content as A7, drawn as cards instead of rows, so Ben can compare the two.
It built the cards design from A7's own rows and words, added a notice line and 35 tests, and listed seven ways the cards read differently from the rows.
HOW IT USES A7. Each card is A7's own row object for its branch (the test checks it is the same object). The words and figures come from A7's labels and slice by import.
A message of Ben's
An agent
map:bens-looks
Built
Read-only: list every time Ben looked at project 3's pages, his words, his verdict and what followed, and say whether the work had drifted from the site's purpose.
It counted about 13 looks since Sep 19, found a repeating pattern of praise for the direction and rejection of the form, and noted the direction had reset four times in five days.
Ben has looked at pixels on project 3 about 13 times since 2026-09-19 (listed below). The pattern holds every time.
An agent
map:deploy-readiness
Built
Read-only: map what shipping would involve: the routes that would go public, what reaches the browser, the site rules at risk, and the safe deploy sequence.
It found the merge would be clean but would publish eight home-care pages with no gate, and that the homepage had not been touched.
An agent
map:content-and-rulings
Pooled, also verifiedThe data
Read-only: lay out the first screen's anatomy, each branch's insight and action as the content spine (the lead's reference document of what each branch has to say) states them, the rulings, and what is still open, noting where code and documents disagree.
It laid out the anatomy and each branch's insight and action, and found three places where what had been built and the documents disagreed, among them Tyler's worth and Houston's.
The content layer is ruled but not yet stored, and nothing on the first screen can go to Ben until it is.
An agent
map:build-state
VerifiedThe data
Read-only: match each item of the new-figures work order to what was committed or left uncommitted, judge whether the uncommitted item 6 was complete, run only the two type checks, and summarize the staged next step.
It confirmed items 1 to 5 were committed, judged the dead builder's uncommitted item 6, and found it departed from the work order at Houston and that its check by the second program had never been shown to pass.
S6c is half landed, which contradicts the carryover's description. Items 1-5 (display strings, direction words, ranks, first locks by cause, Tyler's order-by date and binding safe day) are committed by the S6c builder as 1034d28, 6525a79, dbdfc50, 8ab8e00 and 06748f0 (09-23 23:39 to 23:55).
A check of the lead's
While the readers work, I'll look at the actual pixels. You haven't seen A7 or A7b yet, and I shouldn't hold a view on "ready" without seeing them.
A message of Ben's
A check of the lead's
You're right, it landed well, and my first take this morning was off: I criticized A7 and A7b for problems that session had already caught and ruled on.
An agent
critic
VerifiedThe data
Check the five readers' maps of the project's state for what they missed or got wrong.
It checked the five maps against git, the code and the stored file, found several claims out of date or wrong, and confirmed one map's finding that Tyler's device order was priced 1 to 2 points too high, adding that the running chain would probably print it that way.
So the running S6e will probably price Tyler's device action at the fixed-now figures, 1 to 2 points above what ordering the devices can deliver. Nothing in the chain catches this.
An agent
build:s6c-rest
Pooled, also verifiedThe data
Finish items 6 to 10 of the new-figures work order, each stored and twinned in the second program, then write the handoff for all ten.
It read the work order, the uncommitted item 6 and the code for five minutes and changed nothing before the lead stopped the chain.
Search spine for worth in points and Houston Plan A finish
A check of the lead's
The grounding map is in, and three of its findings affect the chain that's running now. Before deciding whether to intervene, I'll check how far the S6c builder has got.
An agent
build:s6c-rest
Pooled, also verifiedThe data
Review and finish item 6 from the uncommitted work, store items 7 to 10 with twins in the second program, run the full check and write the handoff.
It stored and twinned the last five items, fixed the worth rule's reasons and a test broken since item 4, and the second program agreed on 40,815 figures.
All ten keys are stored, each has a basis in analysis-keys.json, and each is twinned in reconcile.py.
A message of Ben's
A message of Ben's
An agent
read:embed
Built
Read-only: work out how a live piece of the dashboard could sit on the homepage without sending the 1.4 MB data file to the browser; its task stated that comp A7 would send it.
It showed its task's premise was wrong: comp A7 runs on the server and the live pages carry none of the data file's field names, so no new data layer was needed.
The task's premise about A7 is wrong, and that matters more than anything else here. comps/first/a7-seven/index.tsx is a server component.
An agent
read:homepage-visual
Built
Once the Northline Roasters coffee-store demo finishes fading in, it captures the mouse wheel. In a real wheel test at 1440, about 9 notches brought the page to the embedded store.
An agent
read:receipts
Verified, also pooledThe data
Read-only: list the process numbers the showcase might print, where each lives and whether code can read it, and propose the smallest honest file to hold them.
It rebuilt the modeled world from its seed and reran the analysis, getting files identical to the committed ones, and sorted the process claims into true, stale and not yet stored.
The strongest new receipt I found, and verified: regenerating from the seed into scratch (7 s) and re-running the analysis (25 s, 80 cross-checks) gives git blobs identical to HEAD for manifest.json, analysis.json, queries.json and analysis-keys.json.
A message of Ben's
An agent
read:rules-that-hide
Pooled, also sourcedThe industry
Build an inventory of the real Texas rules the modeled world runs on, above all those that hide a problem, each exact and sourced, confirming or refuting seven candidate rules.
It wrote a rules inventory of 19 rules and, on its own initiative, re-read six live sections of the state's handbooks, finding every quoted sentence matched and four candidate wordings needing to be made exact.
I re-read six sections of the state's own handbooks today (2026-09-24, raw page text via curl, not a summarizing fetch): EVV Policy Handbook 7000, 9000, 11000, 12000 and 14000, and caseworker handbook section 4674. Every quoted sentence in them matched the project word for word.
and 1 page not shown
An agent
read:orchestration
Built
Read-only: inventory the workflow orchestration as it actually ran, with counts from the journals, the working patterns used, where Ben ruled, and what went wrong and was caught.
It counted 49 runs and 484 agent launches from the journals, described each working pattern with a real example, and listed 15 failures and what caught each.
49 workflow runs were launched between 2026-09-18 16:56 CDT and today, with 484 agent launches. 43 completed, 1 was stopped by the lead and relaunched, 2 died (one to the 02:42 Windows Update restart, one when the session hit the usage limit) and 3 were running while I wrote.
An agent
build:sh0-receipts
Pooled, also verifiedThe data
Make every process number the showcase prints come from a file written by code: the second program's result, a receipts file, a planted-error check, and fixes to the stale typed counts.
It made the second program write its result to a file on every run, made the site refuse a stale or disagreeing result, and planted one wrong figure in each of seven groups: 7 planted, 7 caught.
planted-errors.ts (npm run homecare:planted) changes one stored figure by one unit in each of the seven groups, in a scratch copy. reconcile exited 1. Result: 7 planted, 7 caught, and nothing failed outside the planted groups.
An agent
critic:s6c
VerifiedThe data
Read-only: check every new stored figure against its rule, check the second program's twins are independent and that only new values were added, and rerun the whole check.
It reran the whole check, planted 22 mistakes that each failed by name, worked the worth rule by hand at all seven branches, and found one fix needed and three small ones.
Every mutation lands in its own named FAIL. Clean re-run: 40815, every one agrees
An agent
fix:s6c
Pooled, also verifiedThe data
Fix every defect the new-figures critic found, keeping the ground rules, and rerun the whole check.
It fixed the main defect and two small ones, proved each guard with throwaway planted mistakes, and left the fourth for the lead to rule.
The should (fff2fe9): the close-hand-made fixture now rebuilds headline.first_locks_by_branch from the branch order it rewrites.
A message of Ben's
An agent
build:sh1-comps
PooledThe data
Build three trial versions of the homepage showcase section (payoff first, the journey, compact plus a lesson page) as real pages, and shoot them.
It read the brief and inventories for under a minute before the lead stopped it and handed the three versions to three builders working at once.
Check showcase worktree state and disk space
An agent
build:sh1-c3
Pooled, also verifiedThe data
Build trial version C3: a compact homepage section with one way into a full lesson page, and build the lesson page too.
It built the compact section and the lesson page with every number read from the stored files, held by a test, and flagged figures the brief used that no file stores.
Every number is read from manifest.json, analysis.json, reconcile.json or receipts.json through a typed slice that only selects and renames.
An agent
build:sh1-c2
Pooled, also verifiedThe data
Build trial version C2, the journey: the three lessons in order, each ending in the piece of the dashboard it makes possible, landing on the live first screen and Tyler's worked example.
It built the journey with each rule drawn in the branch where it hides a problem and every figure showing its stored path, and flagged places where the brief or the dashboard disagreed with the file.
Every figure shows its stored path on hover or focus. Every number comes from manifest, analysis, reconcile or receipts.
An agent
build:sh1-c1
Pooled, also verifiedThe data
Build trial version C1 of the showcase, payoff first: the thesis and the live first screen at the top, then the three lessons with their receipts, then Tyler's worked example.
It built C1 with every figure read from a stored file and a test that fails on any typed digit, and printed the stored figures where the brief's examples differed.
It checks every figure against its file and path. It also removes every stored figure from the text and fails on any digit left over
An agent
build:s6e
Pooled, also verifiedThe data
Store the content spine's sentences in the data file as templates over stored figures (each branch's cause, insight and two actions, and the company headline), twin them, and reprint comps A7 and A7b.
It stored templated sentences for all seven branches and the headline, twinned every sentence, planted five errors that the second program caught, and reprinted the designs; the second program agreed on 41,235 figures.
The twin (reconcile/groups/e_ladders.py) resolves every slot itself, re-renders every sentence string for string, re-tests the stored conditions and the worth rule, and picks the company pair from its own recounts.
A message of Ben's
A message of Ben's
An agent
build:sh2-home
Built
Put the showcase on the real homepage in Lead Intake's slot, in C3's compact form, fix the hero's wheel trap and renumber the sections.
It read the brief, comp C3 and the homepage's hero code for three minutes before the lead stopped it when Ben proposed keeping Lead Intake.
Read AppShell render for preview vs embedded
An agent
build:sh2-case-study
PooledThe data
Build the case study as its own page from C3's lesson page and C2's rule panels, with every number read from stored files.
It cut the trial versions' screenshots into readable pieces and read C3's code for three minutes before the lead stopped the run.
Let me look at the C3 lesson shots and C2 shots first to see what I'm building from.
A message of Ben's
An agent
build:sh2-case-study
PooledThe data
Build only the case-study page from C3's lesson page and C2's rule panels, with Tyler's device order priced on the day it can land.
It read the brief and the earlier handoffs and was looking at the trial versions' screenshots when the run was stopped a minute in, after Ben asked to pause building the site until the vision was decided.
Let me look at the shots first, then read the C3 lesson code.
A message of Ben's
An agent
critic:s6e
VerifiedThe data
Read-only: check every figure in every stored sentence against the file in all seven states, the voice rules, and the reshot designs.
It re-rendered every sentence with its own resolver in all seven files, confirmed the numbers, and found 12 defects in what the sentences claimed, one of them blocking.
Everything mechanical in the slice holds: my own resolver re-renders all sentences in all 7 files exactly as stored, the leaf diff adds leaves only, analyze is byte-identical over two runs, reconcile agrees on all 41,235 figures, and the twin fails on a planted slot change.
An agent
hunt:method
Built
Find, shoot and annotate 6 to 8 pages where 'how we know' is itself interesting, as design references for the case-study page.
It opened and shot method and proof pages (engineering write-ups, a forecast's how-it-works page, methodology hubs), but was cut off by the billing stop before returning its notes.
Is the full article publicly readable? List its section headings and describe any figures/diagrams (what they show). Does it have a section on what went wrong / lessons / evaluation?
8 web searches
An agent
hunt:journalism
Built
Find, shoot and annotate 6 to 8 great data-journalism explainer pages as design references for the case-study page.
It searched award lists and news graphics desks and shot more than a dozen explainer pages, but was cut off by the billing stop before returning its notes.
List every winning project with its outlet, its URL if given, and one line on its format or visual device (explainer, scrollytelling, annotated charts, interactive tool, methodology shown).
15 web searches
An agent
hunt:civic-health
Built
Find, shoot and annotate 6 to 8 pages that make government rules or health data plain, as design references for the case-study page.
It opened eight civic and health-policy pages (benefits case studies, a Medicaid explainer, a page grading states' data) for how they explain rules plainly, but was cut off by the billing stop.
Describe the page structure: headings verbatim in order, what figures/charts/annotations it uses, and how it explains why a letter grade hid problems. Any interactive device?
20 web searches
An agent
hunt:explorables
Built
Find, shoot and annotate 6 to 8 instructive interactive explainers as design references for the case-study page.
It shot eight interactive explainers and wrote what each does and what the case study could take from it, the only hunter to return before the billing stop.
Beat 1, the rules that hide the problem. The mechanism is the same as ours: one metric looks fine overall and hides a problem in one group.
3 web searches
An agent
hunt:case-studies
Built
Find, shoot and annotate 6 to 8 great case-study pages as templates for the case-study page.
It opened 23 pages from design studios and product teams (portfolio lists, case studies and write-ups) and shot several, but was cut off by the billing stop before returning its notes.
Describe the case study page structure in detail: headings, sections, how questions are framed, visuals (screenshots, videos, charts), stat callouts, captions, and pacing.
11 web searches
A message of Ben's
An agent
fix:s6e
PooledThe data
Fix every defect the sentence critic found, and rerun the whole check.
It studied the defects and the data for twelve minutes, including checking whether Fort Worth's accepted transactions are distinct visits, and was part way through editing the sentence templates when the billing stop cut it off.
Now I have what I need. Let me write the ladder module changes.
An agent
curate
Built
Remove duplicates from the hunters' finds, pick the best and build the reference board in the judging room for Ben to keep or kill.
It never started: the billing stop failed it the moment it launched.
A message of Ben's
A message of Ben's
A check of the lead's
You're right to push on that. I called the data layer sound on the strength of Gemini's ledger entry and a stored reconcile file, without checking it myself.
A check of the lead's
You're right, and I was wrong to call the data layer sound. The numbers hold. The meaning built on them was left half-finished, and nothing on the live site prints even the finished part.
A message of Ben's
A check of the lead's
Pulling your own words from the session transcripts, so the line is drawn from what you said and not from anyone's reading of it.
A message of Ben's
A message of Ben's
An agent
mine:tyler-sanantonio
Pooled, also verifiedThe data
Read-only: for Tyler and San Antonio, say from stored values what is most worth discussing, judge each link of the stored cause-to-consequence chain, and list facts stored but not printed.
It named each branch's kind of problem, judged the six links of each stored chain, found a wrong date in Tyler's stored timeline, and noted that San Antonio's failing finish still printed as if it were a save, a defect an earlier check had already found.
Tyler delivers more of its planned care than any other branch, yet both STAR+PLUS contracts are heading to 74 because 20 client home phones were removed after the landline review and no alternative device was ordered.
An agent
read:the-package
Pooled, also verifiedThe data
Read-only: list what each branch's hand-off email needs, what is stored and what is owed, every list that could be attached, and a stored shape for the email package.
It found the file stores the core of all seven emails but no sender, subject or most attachments, found the one possible sender in the raw staff data, and flagged stored worths that would mislead in an email.
Both what the email says and how it looks have to be settled before anyone builds it. The file already stores the core of all seven emails: the recipient, the finding, the cause, two asks each with an owner and a last day, and what each ask is worth.
An agent
read:visual-track
Built, also verified
The case-study board is 26 usable references out of 27, and building it is four small writes plus one room card.
An agent
mine:houston-fortworth
Pooled, also verifiedThe data
Read-only: for Houston and Fort Worth, say from stored values what is most worth discussing, judge each link of the stored cause-to-consequence chain, and list facts stored but not printed.
It named Houston's kind of problem as a past failure under a payer notice, judged each stored link, and found Houston's stored worth implying that on-time training avoids a corrective action plan.
Before the quarter that decides Plan A's notice starts on Sep 1 (our reading), the meeting should decide who keys managed care authorizations when Houston's one scheduler is out.
An agent
audit:s6e-defects
VerifiedThe data
Read-only: audit which of the sentence critic's defects are fixed, partly fixed or open today, and write a smallest-first work order for the fix round.
It re-rendered every stored sentence (no differences), queried the raw data to settle one doubt, found the carryover overstated a fix, and wrote the work order the Sep 28 fix round followed step by step.
Of the S6e critic's 12 defects plus its outside-the-slice note, 2 are fixed, 3 are partly fixed and 8 are open.
An agent
read:world-and-dials
Pooled, also verifiedThis world
Read-only: explain how a contract comes to read 'close', give the honest options for putting a contract on the line, and cost each builder finding about the world.
It found no contract is close today for a structural reason, cited an earlier builder's measurement that Fort Worth cannot reach the line through its new hires, and proposed a relapse at Houston as the honest candidate.
The W2b builder measured the limit. Even if every visit by every new hire were typed in by hand, the branch would still finish Q4 at 82.8. At the rate its hires actually type in, they cost it about one point.
An agent
mine:austin-elpaso-mcallen
Pooled, also verifiedThe data
Read-only: for Austin, El Paso and McAllen, say from stored values what is most worth discussing, judge each link of the stored chain, and list facts stored but not printed.
It named each branch's kind of problem (Austin: passed, then paid it back) and found Austin's stored first action answers neither its finding nor its cause.
Payers want about $7,082 back from Austin's Dec to Feb claims, a quarter that passed on every contract.
A message of Ben's
A message of Ben's
A message of Ben's
A message of Ben's
An agent
build:f1-fix-round
Pooled, also verifiedThe data
Fix the open sentence defects in the defect audit's order, make the stated rules true, and make the second program catch a sentence chosen by the wrong rule.
It committed four fixes (the 'on the app' label, the stored cause at every branch, guarded 'passed' and 'failed', and failing finishes that say so). It planted a mistake for each of the last three, which the second program caught, and it was stopped when the machine ran out of memory.
Every branch's own cause variants now carry the condition on branches[].cause.key, and every branch falls back to the shared sentences method.ladder.cause_fallbacks names for its key.
An agent
write:hou-ftw
PooledThe data
Write the hand-off emails for Houston and Fort Worth as sentences whose every figure is read from the stored file, changing no file.
It drafted both emails from stored figures, led Houston with its late-keyed authorizations and the payer's notice, and refused a superlative that rested on a near tie.
Houston and Fort Worth are 1st and 2nd on one-sided cost and rate; led by the missed punch they read as one kind. I led Houston with the authorization keying and the notice, and kept the one-sided cost for Fort Worth alone.
An agent
write:aus-elp-mca
PooledThe data
Write the hand-off emails for Austin, El Paso and McAllen as sentences whose every figure is read from the stored file, changing no file.
It drafted three emails, moved Austin's ask to the biller and the recoupment and El Paso's to the scheduler, and flagged money priced from planned hours that needs 'by our estimate'.
Estimate wording: El Paso's never-staffed money is priced at the payer's rate on the hours planned (queries.json, visits-not-delivered), and McAllen's waiting money is partly priced at planned hours (priced_from.plan.visits 43), so both sentences say 'by our estimate'.
An agent
write:tyl-sat
PooledThe data
Write the hand-off emails for Tyler and San Antonio as sentences whose every figure is read from the stored file, changing no file.
It drafted both emails sentence by sentence with every figure resolving from the stored file, and listed what it left out and what it could not verify.
Every slot in both emails was resolved through scripts/dev/kit/ladder-render.ts (resolve, say, render) against data/home-care-ops/analysis.json at dataset hash df5b334b in the main tree.
An agent
build:v1-board
Built
Build the case-study reference board in the judging room from the dead hunters' 53 shots, one plain sentence per reference, for Ben to keep or kill.
It looked at all 53 shots, wrote one plain sentence per reference, built and tested the board, and corrected notes the pictures did not bear out.
Looked at all 53 shots before writing about any of them.
A check of the lead's
I've done section 8 and launched both workflows. They're running now, and the gate passed on every line before any change was made.
An agent
refute:tyl-sat
VerifiedThe data
Read-only: resolve every figure in the Tyler and San Antonio drafts and try to refute every claim; where unsure, the sentence does not hold.
It re-rendered all 14 sentences exactly and reworded eight, most importantly moving Tyler's device order from a modelled Jul 27 with a 0.15-point margin to 'today'.
The finding that cuts across both emails: method.worth_rule.models states that every usable finish 'models a return to the slice's own best finished quarter from the fix day (method.fix_days.basis), a labelled modelling choice wherever it prints'.
An agent
refute:hou-ftw
VerifiedThe data
Read-only: resolve every figure in the Houston and Fort Worth drafts and try to refute every claim.
It re-rendered all 14 sentences exactly, found no figure wrong, and reworded six for claims the file does not support, such as a 'likely' range and a superlative resting on a near tie.
Nothing is refuted in its figures. Each fixed template above was rendered with the resolver and checked.
An agent
refute:aus-elp-mca
VerifiedThe data
Read-only: resolve every figure in the Austin, El Paso and McAllen drafts and try to refute every claim.
It resolved every draft sentence against the stored file and was drafting fixes when the machine ran out of memory and killed it.
Resolve and render every draft sentence against analysis.json
An agent
build:f1-fix-round
Pooled, also verifiedThe data
Continue the sentence fix round after the crash, told what its predecessor had committed and left.
Restarted by the lead's resume, it checked memory and began verifying its predecessor's four commits before the lead stopped the run within a minute.
Now I'll verify each of the four committed steps against its message.
An agent
build:v1-board
Built
Build the case-study reference board (the same order as its predecessor).
Restarted by the lead's resume, it found the board already committed and was checking those commits when the lead stopped the run within a minute.
The earlier run committed the board, the mount line and the front-page card. Only the board shot is left untracked.
An agent
refute:aus-elp-mca
VerifiedThe data
Read-only: resolve every figure in the Austin, El Paso and McAllen drafts and try to refute every claim, checking doubts against the raw modeled data.
It refuted McAllen's second ask, found the stored 'waiting to be keyed' label overcounts at all seven branches, and struck Austin's promise of appeal deadlines the file does not hold.
'Type in the 97' sends two coordinators to key 54 visits the attendants clocked on the app.
An agent
build:f1-fix-round
Pooled, also verifiedThe data
Continue the fix round after the crash: check the dead builder's four commits and its uncommitted step 5, then carry on through the rest of the work order from step 6.
It confirmed the four earlier commits against the stored file, committed step 5 (no action borrows another's dollars) and step 6 (four new checks in the second program), and was on step 7 when the machine froze again.
Each mutation is caught by exactly the check it targets and by nothing else, which also shows the twin was blind to all five before.
An agent
assemble:e1
Pooled, also verifiedThe data
Apply the skeptics' verdicts, print the seven emails from the stored file by script, test them and commit, for Ben to read.
It applied every verdict, dropped the two sentences the skeptics struck, printed the seven emails by script from the stored file, and showed each test guard fails on a planted mistake.
The script fills every slot with the kit's resolve, say and render, and it computes and compares nothing.
A check of the lead's
Verified: the rerun reproduces both printed files byte for byte, 1,387 of 1,387 tests pass, and there are no dashes.
A message of Ben's
An agent
build:f1-fix-round
Pooled, also verifiedThe data
Finish F1, the fix round for the stored sentences: make the sentence rules true, make the second program (the one that works every figure out again) catch a wrongly chosen sentence, and close the open audit defects.
It took over a fix round from two builders lost when the machine froze, checked each of their commits, re-ran their 11 planted mistakes and committed the rest, 17 commits in all, with one step left for the lead to rule.
Steps 1 to 6 were committed by the two builders who died. I checked each commit's diff against its message and against the final code, and re-ran all 11 of their planted mutations on the final file; each still fails reconcile by name.
An agent
critic:f1
VerifiedThe data
Read the fix round's work without changing anything, rerun its gate, and try to break the second program's new checks with planted mistakes; find defects, not direction.
It reran the fix round's gate, wrote 25 planted mistakes for the second program and had run three (two caught, one not) when the run was stopped, so it returned no answer.
Gate lines 1 to 15 reproduce. Now the mutations on the second program. First I'll check the stored values my mutations will lean on.
An agent
facts:f1
VerifiedThe data
Check the claims, not the figures, in every stored sentence of the live file and six test states, against the verified rules and the data.
It read every stored sentence in the seven files and was testing counts in the raw modeled world when it was stopped, so it returned no answer.
I have now read every sentence in all seven files. Before judging wording I need the ruled voice for product sentences. Let me read the voice pass.
A message of Ben's
A message of Ben's
An agent
Build email module comp C
PooledThe data
Build comp C of the email module as a static page, printing the seven emails word for word from the stored files, each sentence opening onto the stored figures behind it.
It built the letter page with every sentence opening onto small pictures of its stored figures, matched all seven emails word for word, and fixed pictures that misstated the data.
each branch's email reads as a real letter, and every sentence opens in place onto small pictures of the stored figures behind it.
An agent
Build email module comp B
PooledThe data
Build comp B of the email module as a static page, proving each email sentence with a row of small pictures of its stored figures.
It built the page with each email sentence proved by small panels of stored figures, and fixed drawing bugs that hid some of the figures.
A small ring marks where each contract stands with nothing done. A stem runs up to its finish on each fix day, so the email's "83 instead of 74" is drawn at every setting and the stems shrink week by week until the dots fall under the line.
An agent
Build email module comp A
PooledThe data
Build comp A of the email module as a static page where one branch picker drives rows of the email's sentences, each figure opening onto where it is stored and how it was checked.
It built the field-guide page with every figure opening onto its stored place, query and recount, and relabelled a McAllen column that did not match the email's count.
Any bold figure opens a note showing where it's stored, which query counts it, and how the second program rechecked its group.
A message of Ben's
A message of Ben's
An agent
critic:f1
VerifiedThe data
Read the fix round's work without changing anything, rerun its gate, and try to break the second program's new checks with planted mistakes; find defects, not direction.
It reran the gate itself and planted 30 mistakes of its own; 5 got past the second program, through four of its checks. It also ruled blocking a Houston sentence the builder had already flagged.
F1 needs one more fix round, and one defect blocks it: Houston's two notice worths read as the claim the brief's section 5 forbids (that doing the training on time avoids the corrective action plan). F1's own rewrite of worth_rule now says the opposite of that sentence.
An agent
facts:f1
VerifiedThe data
Check the claims, not the figures, in every stored sentence of the live file and six test states: does each say what its figures measure, state only verified rules, and not mislead a branch manager?
It read 460 stored sentences against the project's rules notes and the data and reported 21 defects, three of them blocking; four of the 21 had already been raised by an earlier audit or by the round's builder.
Verdict: fix round needed. Three blocking defects, all in sentences an email would print as asks: Houston's notice worth says what the brief's section 5 forbids (both variants), Houston's training action names one of the four trainings the notice requires as if it were the whole requirement
A check of the lead's
Comp A is back. Its builder reports a finished page, 14 screenshots and its own self-checks passing. A report isn't a check, so I'll verify it myself
An agent
fix:f1
Pooled, also verifiedThe data
Fix every blocking and should-level defect the critic and the fact-checker found, prove each new check with planted mistakes, and rerun the whole gate.
It fixed the four blocking defects and most of the rest in 17 commits, sent three to the lead, and proved its new checks with 13 planted mistakes that each failed by name.
13 planted mutations were each run on a scratch copy against the final file. All 13 make reconcile exit 1 and name their check.
A message of Ben's
A message of Ben's
A message of Ben's
An agent
read:dashboard-today
Built
Read-only: map the dashboard as it stood (first screen, map, modules, each branch's anatomy) for the structure spec.
It had read the brief's map section and Ben's rulings and started on comp A6 when the lead stopped the run a minute in.
The live /v2 renders comp A6 with the branch tint. Let me read comp A6 in full.
An agent
read:rules-and-data
PooledThe industry
Read-only: list the rules the page may and may not state, explain the grade plainly, and list the data facts it may state with their stored receipts.
It read the rules appendix in full and began on the checked-claims list when the lead stopped the run a minute in.
B-rules-v2 is read in full. Now the MAY SAY findings in the grounding JSON; I'll inspect its structure first.
An agent
read:page-today
Built
It read the live page's code, the earlier checks of its claims and its look, the showcase brief and Ben's rulings when the lead stopped the run a minute in.
I'll start by mapping the files named in the task.
An agent
read:chain-content
PooledThe data
Read-only: walk Tyler end to end as one chain of causes from stored values, link by link, and say what exists for all seven branches.
It read the email files and was writing a helper to read stored figures the way the email script does when the lead stopped the run a minute in.
I'll write a small read-only helper in the scratchpad that reproduces the kit's resolver exactly, so every path I name is resolved the same way the emails script resolves it.
An agent
read:state-and-order
Built
Read-only: map the state of the work, the rulings, the open pieces, what depends on what, and the day's lessons.
It had read the carryover, the brief and the email handoff and was checking the figures still owed when the lead stopped the run a minute in.
Sections 5 to 7 of E1's handoff are read. Let me check whether section 4's keys list grew past 49 after the lead's changes, then read V15 to V26.
A message of Ben's
A message of Ben's
An agent
read:link3-data
Pooled, also verifiedThe data
Ground link 3, the data: the modeled agency, what each of the two programs does, the second program's record and its fingerprint, and every process record, each read from the stored files.
It read the stored records and the two programs' own headers, recomputed the data file's fingerprint, and found the handoff's word 'recounts' claimed more than the second program does.
The agency is modeled, with no real people in it. It is built to follow the real rules, and it rebuilds exactly from one seed (about.note, entries[id=rebuild_from_seed].said_as).
An agent
read:link5-effect
Pooled, also verifiedThe data
Ground link 5, the effect: Tyler's hand-off email and the switch to all seven branches' letters, every sentence tied to stored figures.
It laid out Tyler's letter and the seven letters by stored path, and found two letters stating rules that were not in the verified rules list, one of which (Austin's) the emails' handoff had already flagged.
Tyler's letter reads straight through as a real email a branch manager could act on today: one order (20 alternative devices), one owner (Alicia Blackwell, EVV coordinator), a day (today, with our model's last day to order Jul 27, worked back from Plan B's last safe day Aug 3)
An agent
read:link2-rules
Pooled, also verifiedThe industry
Ground link 2, the rules: what Texas grades and where the line is, when a quarter counts, what failing costs, and the landline rule Tyler's story turns on, each tied to a verified rule.
It set each rule the page may state beside the verified rules list and the stored values, and flagged stored and drafted claims that said more than the rules allow, among them a stored sentence that gives a modeling choice as the state's rule.
What is counted and the line: the state sets one grade per contract, never per branch. It gives up to 60 points for the share of accepted visits clocked in and out electronically and up to 40 for the share of sends not rejected. 80 passes after rounding, so 79.5 passes.
An agent
read:dash-first-screen
Pooled, also verifiedThe data
Ground the dashboard's first screen with nothing picked: what the director sees across seven branches and which to open first, every value read by path.
It specified the first screen from the stored company finding, the map and each branch's finding, and warned that the file mixes two kinds of figure: the blended kind calls San Antonio 'close' while one of its contracts heads to 66.
The company's insight leads the screen (headline.insight, stored and twinned, L100), with the contracts' verdict as one supporting line under it and a stated fallback when the mirror does not hold.
An agent
read:dash-states
Pooled, also verifiedThe data
Ground the dashboard's states, interactions and hand-offs: what 'by when' and 'what it is worth' mean in each branch's state, every value read by path.
It worked out each branch's state from its contracts and warned, as earlier notes had, that the stored blended fields call San Antonio 'close' and not at risk while its Plan B contract is already too late.
A branch's state is read at the contract grain, never the blend, and it must be ONE stored key per branch (owed O1, stored by analyze, recomputed by reconcile).
An agent
read:code-as-it-stands
Pooled, also verifiedThe data
Read the code as it stands and say whether the two agreed pages can be built from the code and stored files without a dev server, and what each choice costs.
An agent
read:link4-diagnosis
Pooled, also verifiedThe data
Ground link 4, the diagnosis: the dashboard opened on Tyler, where each contract stands and why, every value read by path.
It read Tyler's contracts, levers, weekly steps and actions by path and warned which stored figures would contradict each other if printed side by side.
Where it stands: the state grades Tyler's three contracts, not Tyler. Both STAR+PLUS plans were passing (Plan A 94 then 86, Plan B 92 then 87) and are now heading to 74, under the line of 80. Fee for service is slipping too (92, 91) but holds, heading to 82.
An agent
read:dash-anatomy
Pooled, also verifiedThe data
Ground the branch view, one layout for all seven branches: contract rows, cause, levers, finding and the one thing to do, every value read by path.
It specified one six-part view for all seven branches from stored paths and flagged stored ranks that print a near tie as a clear place.
One anatomy, read by path, for all seven: the contract rows (branches[].contracts_graded[] plus each contract's three quarters from contract_cells[scored_identity_id].quarters[].rounded), the cause (cause.label over ladder.cause.said)
An agent
read:link1-problem
Pooled, also verifiedThe industry
Ground link 1 of the breakdown page, the problem at Tyler, in the stored files: what the state's score is, what it hides at Tyler, and Tyler's contract scores, every value read by path.
It read every value for link 1 by path and wrote the link's exact section, correcting the handoff's claim that Texas grades agencies: it grades each contract.
Texas does not grade Tyler's office. It grades each contract the office works under, one score each, on how its visits were clocked and sent, and a contract passes at 80.
An agent
write:01-breakdown-page
PooledThe data
Write the breakdown page's exact spec, 01-breakdown-page.md, from the nine readers' groundings, re-reading every figure by path.
It wrote the 75 KB spec for the breakdown page's five links, re-read every figure by path, and recorded the disagreements it settled between the readers and the handoff.
An agent
write:02-dashboard
PooledThe data
Write the dashboard's exact spec, 02-dashboard.md, from the readers' groundings, re-reading every figure by path.
It wrote the 57 KB dashboard spec (the first screen, the branch view, the states, the hand-offs and what is owed) and corrected one of its own draft counts against the stored file.
An agent
critic:feasibility
VerifiedThe data
Check whether both specs can be built as written from the stored files and the code, without a dev server; find contradictions and gaps.
It re-read about 150 stored paths, checked 12 code citations, ran the honesty lint on the public page itself, and found one blocking contradiction and 14 major gaps.
Not ready for slice 4 as written, but close. The facts are sound: every value re-read by path matches the stored file, and the code citations hold.
An agent
critic:honesty
VerifiedThe data
Check that the two specs say nothing the stored data or the verified rules do not: re-read every cited figure and compare every rule sentence with the verified rules list.
It re-read 1,177 cited values with its own script and found none wrong, then raised two blocking and seven major problems, among them rule sentences stronger than the verified rules allow.
Every figure checks out, with no wrong value in 1,177 re-read claims, but the specs are not ready for slice 4 until two blocking items and seven major items are fixed.
An agent
critic:rulings
VerifiedThe data
Check both specs against Ben's rulings and the structure he agreed: find where they break or drift from a ruling, not matters of taste.
It read both specs against Ben's rulings and found 15 major gaps, among them a wrong count of where the view and the email disagree and a question to Ben that hid that he had already cut the story.
Both files follow the structure Ben agreed. The chain is walked with Tyler, the email is the one element at the end in comp C's flow, and the dashboard is map first with the insight on top and one six-field anatomy.
An agent
revise:02
Pooled, also verifiedThe data
Revise the dashboard spec against every finding of the three critics, checking each at its source first.
It checked all 36 findings at their sources, applied them to the dashboard spec, and corrected one critic's count.
I revised docs/project3-structure/02-dashboard.md against all 36 findings, and it is now 79,690 bytes. I checked every finding myself first
A message of Ben's
An agent
revise:01
Pooled, also verifiedThe data
Revise the breakdown page spec against every finding of the three critics, checking each at its source first.
It re-read each finding at its source and applied all 49 on its side to the breakdown spec, including both blocking ones.
I revised docs/project3-structure/01-breakdown-page.md against every finding. All 49 hold on my side of the file, and I applied them
A check of the lead's
The reconcile passes: 41,669 figures compared, every one agrees, with 401 checks and 0 failed, exactly as the fixer reported.
A message of Ben's
A message of Ben's
An agent
register:03-owed
PooledThe data
Write one owed register, 03-owed.md, merging every owed item from both specs and the email handoff under one set of numbers, checking by path what is already stored.
It merged every owed item into one register of 47 entries and found five items the fix round had already stored.
I wrote C:/Users/Benod/Projects/bendixon-dev/docs/project3-structure/03-owed.md. It is the single owed register L127 asks for, with 47 entries.
An agent
consolidate:01
PooledThe data
Consolidate the breakdown page spec: cite the one register, use the fix round's final stored words, and shorten it, re-reading every value.
It rewrote the breakdown spec onto the register's numbers and the stored words, re-read every value, and cut the body to about 3,500 words.
An agent
consolidate:02
PooledThe data
Consolidate the dashboard spec: cite the one register, quote the fix round's final stored words, and shorten it, re-reading every value.
It moved the dashboard spec onto the fix round's stored words and the register's numbers, checked every quoted sentence against the stored files, and cut the body to about 4,000 words.
An automated cell-by-cell check against analysis.json and E1-emails.json found no mismatched sentence.
An agent
verify:agreement
VerifiedThe data
Check that the three files agree with each other, with the lead's rulings and with Ben's, and that nothing the shortening dropped was lost.
It compared the three files with each other, their earlier versions and the rulings, and found 2 blocking and 4 major disagreements.
Not ready for slice 4: 2 blocking, 4 major, 10 minor findings.
An agent
verify:honesty
VerifiedThe data
Check the three consolidated files for honesty: every stored value by path, every figure in the prose, every rule against the verified list.
It re-read 1,571 stated values and 2,125 numbers in the prose with its own scripts, found none wrong, and caught two sentences that claimed more than their sources.
Nothing blocking and nothing major. Every stated value (1,571 by path) agrees with the merged files. Every owed key is absent from them.
An agent
fix:03
Pooled, also verifiedThe data
Fix the register against the verifiers' blocking and major findings, checking each at its source.
It checked the stored kind lines itself and fixed the register so all five that break the grading rule carry the flag.
BLOCKING (O6/O7, the kind lines): checked it myself. E1-emails.json stores these kind lines: TYL 'Failing, and fixable in time', SAT 'The grade is lost, the money is not', HOU 'A past failure, under a payer's notice'.
An agent
fix:02
Pooled, also verifiedThe data
Fix the dashboard spec against the verifiers' blocking and major findings, checking each at its source.
It checked each finding at its source and fixed the dashboard spec, including a build note that broke the ruling to print every rank as stored, and a mark in Tyler's view that read a whole list holding only Houston's contracts.
Finding 2 (major, holds): the helper shows notice_state.contracts_failed_without_a_notice holds only SI-PLE-HARRIS and SI-PLB-HARRIS, both branch_id HOU, and [branch_id=TYL] reaches nothing.
An agent
fix:01
Pooled, also verifiedThe data
Fix the breakdown spec against the verifiers' findings, checking each at its source.
It re-read Austin's take-back reasons and searched the rules sources, then rewrote the take-back note, the summary's claim and the question that had been filed as settled.
Finding 1 (F.17): re-read Austin's by_reason (records_missing 111, visit_changed_after_match 2, payer_instruction_not_followed 3; 116 notices) and grepped the sources: F00110 is in neither rules-inventory.md nor the MAY SAY file.
A check of the lead's
Both summaries read right. Now my own spot-check: I'm pulling eight of the headline figures the specs cite for Tyler and reading each by path.
An agent
foundation
PooledThe data
Build the shared base the four wireframes use: code that reads and prints every stored figure with its path, the branch view as one shared piece, and the file of drafted words.
It built the shared code that reads and prints the stored figures, the branch view and 57 drafted titles and sentences, and checked the view on the live file and all six test states.
The shared foundation is built and working: the smoke page renders every part from the stored files, and the builder stopped on no path.
A message of Ben's
A check of the lead's
Before I answer, one quick check that matters if you're away for a few hours: whether Windows has an update restart pending.
A message of Ben's
A message of Ben's
An agent
build:dashboard-a
PooledThe data
Build direction A of the dashboard, the picked branch opening in the map's empty ground, as static pages from the stored files, and shoot each state.
It built direction A of the dashboard as four static pages, including a test-state page and the link-4 mount, and shot each state.
Direction A on the live file. Eight radios with All seven checked on load. Picking a branch lights its counties, pin and plate, and its view wraps around the map
An agent
build:breakdown-a
PooledThe data
Build direction A of the breakdown page as a static page in the judging room, every figure and sentence printed from the stored files, and shoot it at 2560 wide.
It built direction A, five chapters in one column with a side rail, printed from the stored files, and shot it at 2560 wide.
The breakdown page, direction A (Chapters), on the live file. It opens on the frame and five chapters in one column, with the five-step path in a sticky rail at the side.
An agent
build:dashboard-b
PooledThe data
Build direction B of the dashboard, the picked branch taking the canvas with the map shrunk to a strip, as static pages from the stored files.
It built direction B of the dashboard as four static pages, where a picked branch fills the canvas under a small locator map.
Picking by territory, pin, finding row, tab or a name in the insight hides the insight, the map and the findings.
An agent
build:breakdown-b
PooledThe data
Build direction B of the breakdown page (one object read five times) as a static page from the stored files, and shoot it at 2560 wide.
It built direction B, with Tyler's contract rows pinned beside the words through links 1 to 3, and shot it at 2560 wide.
Direction B, one object read five times. Tyler is checked at link 5, the 'What it is worth' proof is open, and every fold is closed.
An agent
design
PooledThis world
Design the smallest real-world changes to the modeled agency's settings that make Fort Worth and Austin near perfect, each with its mechanism, a verified rule and a stored test for 'near perfect'.
It proposed a stored test for 'near perfect' and candidate setting changes for Fort Worth and Austin, measured six trial worlds, and warned that any set that works rewrites Houston's story.
Fort Worth's settled attendants clock in and out at the agency's own rate. The slip is only in its newest: hired in large classes, they miss the clock out at today's rate for their first 13 weeks, and the office keys it.
An agent
critic:dashboard-b
VerifiedThe data
Check dashboard wireframe B for honesty, coverage of the spec and its shots; find defects, not direction.
It traced 1,341 printed values back to the stored files with its own checker (none wrong) and found one sentence on the test-state page printing the project's own blend of Tyler's contracts as Tyler's figure, with no flag.
Direction B is honest and complete for Ben to judge, with one defect to fix before he opens the fallback page. Every figure and sentence on the live pages traces by path (1,341 spans, 0 mismatches).
An agent
critic:breakdown-a
VerifiedThe data
Check breakdown wireframe A for honesty, coverage of the spec and its shots.
It re-read 1,193 printed values in a headless browser (none wrong) and found two sentences the page wrote itself outside any draft frame, one of them a rule claim.
1,193 traced values re-read with 0 mismatches, and no typed figure.
An agent
critic:dashboard-a
VerifiedThe data
Check dashboard wireframe A for honesty, coverage of the spec and its shots.
It rebuilt the pages itself, re-resolved 1,260 printed values (none wrong), clicked every pick in Chrome, and found one blocking defect: on the test-state page, Tyler's worth sentence printed the project's own blend of its contracts with no flag.
Every one of 1,260 traced values re-resolves.
An agent
fix:dashboard-b
Pooled, also verifiedThe data
Fix dashboard wireframe B against its critic's findings, check each fix, and reshoot it.
It flagged the test state's blended sentence, tightened the branch strip, and reshot every state at 2560 wide.
A wrap in my file puts the flag on that exact stored worth template, the same lookup common.ts uses, and does nothing once common.ts's OWN_FIGURE list holds it.
An agent
measure
Pooled, also verifiedThis world
Build each candidate change in a scratch copy of the world, measure it against the six tests, then the combined set, and record every run.
It built and measured 27 scratch worlds, chose a set that makes both branches near perfect, and reported that the set flips Houston's story and needs a missing rule entered.
I measured 27 worlds in the build tree (C:/Users/Benod/Projects/bendixon-dev-build, branch w1 at 27587f3). The baseline reproduces the live dataset df5b334b, and its analysis.json matches the live file byte for byte.
An agent
fix:dashboard-a
Pooled, also verifiedThe data
Fix dashboard wireframe A against its critic's findings, check each fix, and reshoot it.
It confirmed the blocking and major findings, added a local flag list for the test-state sentences, and retook all 13 shots.
All four pages are rebuilt and all 13 shots retaken at 2560, and I looked at every one. The blocking finding and all three majors held
An agent
fix:breakdown-a
Pooled, also verifiedThe data
Fix breakdown wireframe A against its critic's findings and reshoot it.
It removed the unframed sentences, drew pictures for the six other letters, and re-read 1,855 printed values against their stored paths (none wrong).
1,855 printed values re-read against their stored paths with 0 mismatches (909 distinct paths).
An agent
critic:breakdown-b
VerifiedThe data
Check breakdown wireframe B for honesty, coverage of the spec and its shots.
It wrote its own checker, re-read 1,121 printed values (none wrong), and caught two unframed labels claiming more about the two programs than any record says.
It re-read every traced leaf span in the page: 1,121 spans across 545 distinct paths, with 0 mismatches and 0 paths that reach nothing.
An agent
fix:breakdown-b
Pooled, also verifiedThe data
Fix breakdown wireframe B against its critic's findings and reshoot it.
It removed the unframed process claims, fixed the pinned pictures that slid out of view, and reshot the page.
The two lanes are now named only 'The first program' and 'The second program', with no process claims.
An agent
challenge:six-tests
VerifiedThis world
Try to refute the chosen change against the six tests: only Fort Worth and Austin change, each rule is real and verified, the effect is measured, and it is the smallest change that lands.
It refuted the chosen set: it rewrites Houston's stored story, moves figures promised not to move, is not the smallest change, and rests Austin's take-backs on a rule missing from the verified list.
Verdict: refuted as it stands. The measurement reproduces, but the chosen set (FTW-A1 plus AUS-B1) fails O2 and test 4 of L124, and it fails test 2 for Austin under this task's standard (the rule must be in the rules inventory or the MAY SAY findings).
An agent
challenge:re-measure
VerifiedThis world
Rebuild the chosen change independently from the report's description and re-measure everything it claims.
It wrote the change itself in a clean copy, reproduced the chosen world byte for byte, and found two figures that move on the page and one miscopied number in the record.
Not refuted: the chosen set (FTW-A1 plus AUS-B1) reproduces byte for byte, and every figure in the report's before and after table and untouched-branch diff checks out.
An agent
revise
Pooled, also verifiedThis world
Revise the chosen change against the challengers' findings that hold, re-measuring each change, and rewrite the record.
It checked each challenge itself, measured 13 more worlds, chose lighter settings, and found that only a knife-edge setting keeps Houston's story and that no set leaves every figure on the page unchanged.
Verdict: the revised set passes L124's six tests on their letter, but no set can meet two further things the scope asks.
An agent
assemble
Built
Put the four wireframes side by side on one page in the judging room, with a key and the question for Ben.
It made the wireframes page and a card on the room's front page, checked that all 12 pages answer, and shot the page.
All 12 pages returned HTTP 200 (checked with curl at 23:18 to 23:30 CDT, Sep 28).
An agent
apply
Pooled, also verifiedThis world
Apply the chosen change in one regeneration on the side branch w1, one commit per step, run the full gate, and write the before and after.
It regenerated the world with the change on a side branch in 8 commits, ran the gate, and stopped short of green because 20 tests caught a wrong sentence at Houston.
W1 is applied on branch w1 in the build tree, but the gate is not fully green. Line 11 (npm run test:kit) shows 1387 tests, 1367 pass, 20 fail, all at Houston.
A message of Ben's
A check of the lead's
Before the internet dropped, my last command (recording rulings L142 to L145 in the handoff) was blocked, so it probably never ran. Checking the actual state first.
A check of the lead's
The seven emails came out word for word, with no diff at all. The three comp pages changed in 50 lines. Checking whether that's only the fingerprint, or Houston's cause showing up on the page
A message of Ben's
A message of Ben's
A message of Ben's
A message of Ben's
An agent
Build Tyler's round-two view
PooledThe data
Rebuild Tyler's branch view as two static pages from its build order, block by block, with every sentence checked against the stored file.
It rebuilt the view in the new order, printed the lead's verdict unchanged while flagging that it broke a rule, and rewrote a headline that needed a count nothing stores.
The drafted sentences print in the new block order, and every claim they make is checked against the stored file before the page is written.
A message of Ben's
A message of Ben's
A check of the lead's
The builder is done. Before you see it, I'm looking at it myself and checking its claims.
A message of Ben's
A message of Ben's
A message of Ben's
A message of Ben's
A message of Ben's
An agent
Gather look references with shots
Built
Collect candidate pages and screenshots for a page look that works light and dark and for pictures of a pipeline, for Ben to judge a few.
It kept 8 page-look and 9 pipeline references with 40 screenshots and noted the pages it could not capture.
I've sent the lead my report: 8 references kept for the page-look need and 9 for the pipeline need.
21 web searches
A message of Ben's
An agent
read:the-next-lead
VerifiedThe data
As a fresh reader in the next lead's seat, start work from the handoff and build orders alone, read-only, and report where you had to guess, what is missing and what contradicts Ben or the record.
It ran every check the handoff names, which passed, and found two orders that would print false claims about who did the work.
Every check the handoff names ran as written and passed.
An agent
read:a-builder-of-the-branch-views
VerifiedThe data
As a fresh reader in the seat of the builder of the branch views, start from the documents alone, read-only, and report guesses, gaps and contradictions.
It checked every condition in the order against the stored data, and a sample of them against the other version, and found two San Antonio pictures that would come out wrong without any error.
I could start P1 from the documents alone. check-paths.ts passes: 222 paths, 0 miss.
An agent
read:a-builder-of-the-orchestration-links
VerifiedThe data
As a fresh reader in the seat of the builder of links 2 and 3 (how the work was orchestrated), start from the documents alone, read-only, and report guesses, gaps and anything false.
It found every stored value right but eight claims in the orders that the project's own record contradicts.
A builder could start P4 and P5 from these documents, but several facts would come out wrong. Every stored path resolves (check-paths: 222 tried, 0 miss; dry run all-built passes), and every stored value in the P4 and P5 tables matches the file on 2026-09-29.
An agent
read:the-owner-s-words
VerifiedThe data
As a fresh reader holding Ben's own words, read the handoff and the orders and report every drafted sentence that is false, contradicts him, or uses words he forbade.
It checked the drafted sentences against Ben's rulings and the record and found seven that were false or put words in his mouth.
A fresh session can follow these documents: the order of reading, the launch steps and the rules are clear. The risk is in the drafted words.
A message of Ben's
A check of the lead's
Checking the readers' data claims (axis ranges, McAllen's cause, Austin's appeal count and others) with a small read-only script against the stored file.
An agent
apply:p3
Pooled, also verifiedThe data
Apply the cold read's findings to link 1's build order under the lead's rulings, checking each at its source.
It applied 4 findings to link 1's order, checked 31 conditions in both versions of the data, and noted a chart label that misstates what it counts.
All 31 conditions hold in live and w1-world (tmp/lead/p3/conditions.ts).
An agent
apply:p4
Pooled, also verifiedThe industry
Apply the cold read's findings to link 2's build order (the rules research story) under the lead's rulings, checking each at its source.
It applied 25 findings to link 2's order, removed every claim that a person did the lead agent's work, and found one of the lead's own rulings false at its source.
Box 4 is written up by the lead agent, "What I decided" holds only D17's two items, a separate "What the lead agent decided" row holds the lead's four, and the headline is the lead's (L159).
An agent
apply:d1-pattern
Built
Apply the cold read's findings to the pattern page, Tyler's build order and the dark-and-light order under the lead's rulings, checking each at its source.
It applied 8 findings, among them marking the page's light and dark switch as not yet ruled, and corrected where the lead said a rule is written down.
Only its part on 01-tyler-view-build-order.md line 57: the Verdict row's From cell now reads
An agent
apply:p1
Pooled, also verifiedThe data
Apply the cold read's findings to the branch views' build order under the lead's rulings, checking each at its source and every condition in both versions of the data.
It applied 28 findings to the branch views order, checked 59 conditions in both versions of the data, and rewrote lines, some of them the lead's own drafts, that went further than the data.
Every condition now written in the order holds in both worlds: 59 checked by tmp/lead/p1/conds.ts.
An agent
apply:p6
Pooled, also verifiedThe data
Apply the cold read's findings to the build order for the letters' proof pictures under the lead's rulings, checking each at its source.
It applied 6 findings to the proofs order and found that one reader's claim about Plan B's last safe day was wrong.
The axis is now laid out by date, so the weekly stops are evenly spaced, and each contract's last safe day is placed by its own stored date between the two stops it falls between.
An agent
apply:p5
Pooled, also verifiedThe data
Apply the cold read's findings to link 3's build order (the data and how it was checked) under the lead's rulings, checking each at its source.
It applied 20 findings to link 3's order and rewrote three lines that said more than the record: the starting number, the landing gate, and 'one lead agent'.
load({fixture: 'w1-world'}) swaps only analysis.json. So I also checked every receipts: and reconcile: condition against the w1 branch's own files (git show w1:..., read only).
An agent
apply:p2
Pooled, also verifiedThe data
Apply the cold read's findings to the first screen's build order under the lead's rulings, checking each at its source.
It applied 8 findings to the first screen order, so each branch's label speaks of its contracts, and checked all 41 paths in both versions of the data.
Also run: all 41 P2 paths resolve against tmp/home-care-ops-fixtures/w1-world.json, 0 misses, via tmp/lead/p2/w1paths.ts.
An agent
build:d1-tyler-dark-and-light
Built
Dress Tyler's approved view twice, dark and light, as the breakdown page would show it, so Ben can rule on a switch by eye.
It built Tyler's view in the site's dark colours and the kit's light ones side by side and measured every colour pair against the contrast floors.
Side by side, the two pages say the same thing at the same speed: the verdict in heavy Bricolage is the first thing seen, the white or ink line at 80 is the strongest line in each picture, and red squares and teal circles carry the grades in both.
An agent
build:p2-first-screen
PooledThe data
Rebuild the words of the dashboard's first screen (the headline, the words on the map, the money and how to read it) from a one-page build order.
It rebuilt the first screen's words from stored figures, added three checks the order lacked, and flagged one sentence that said more than the state's rule.
An agent
build:p1-branch-views
PooledThe data
Build the dashboard view for the six other branches on the pattern Ben approved for Tyler, from a one-page build order.
It built the six other branches' views from the stored figures, drafting 60 sentences (51 of them tied to conditions the page checks), and flagged where the order or the words did not fit the data.
Each verdict, block headline and why picture reads in five seconds at all six branches, and every view builds cleanly against both the live analysis.json and the w1-world fixture (Tyler through the same door too).
An agent
build:p5-link3
PooledThe data
Rebuild link 3 of the breakdown page, the data and how it was checked, from a one-page build order.
It rebuilt the data section's words and drawing, printing each figure copied from the records with its file and line, and making the script check, in the code itself, two lines no stored receipt covered.
No stored receipt covers the build gate or no-hidden-inputs lines (O20). In place of one, the script opens receipt-checks.ts, analyze/index.ts and reconcile/shared.py and stops unless each holds the lines the sentence describes.
An agent
build:p6-link5-proofs
PooledThe data
Rebuild the pictures under the sentences of the seven emails (link 5) from a one-page build order.
It rebuilt Tyler's email pictures from stored dates and figures and found a small chart that read as contradicting the chart above it.
Tyler's pictures now read in five seconds, one headline over each: the grids, the dates to scale with the people and devices labelled on the marks, and the time left to order.
An agent
build:p3-link1
PooledThe industry
Rebuild link 1 of the breakdown page, the problem, from a one-page build order.
It rebuilt the problem section's words and pictures from stored figures and asked for the model's last safe days on its date line, which the lead then added (L183).
The top works in five seconds. The answer is the largest thing on the page, block 1 is Tyler's approved picture, and in block 2 one chart rises while the other stays flat.
An agent
build:p4-link2
PooledThe industry
Rebuild link 2 of the breakdown page, the rules and how Ben learned them by directing AI agents, from a one-page build order.
It rebuilt the rules section and the drawing of the agent work, and added one row so a failing contract's grade could not be read as a pass.
Its idea reads: my decisions sit on top, the lead agent's under them, and the agents' four steps sit in boxes below.
A check of the lead's
P2's script runs clean with the fix. Section 6 says to run each builder's script myself rather than trust its word, so here is P5's:
An agent
assemble
Built
Mount the built pieces as the two round-two pages, the dashboard and the breakdown page, and take their pictures.
It joined the seven pieces into the two pages, fixed the seams between them, and listed what still read badly.
Slice 4's rail printed the old chapter questions from the draft file, and those are the ones Ben's diagnosis called two questions in one. The rail now copies each piece's own question from its HTML
A message of Ben's
A message of Ben's
A check of the lead's
One thing to check: in that script I opened the handoff for writing but never wrote to it, which would have emptied it.
An agent
read:data-w1
Pooled, also verifiedThe data
It compared the two branches' stored data, found the merge had no text conflict, and set out the order of re-runs needed after it.
An agent
read:base-breakdown
Pooled, also verifiedThe data
Map the approved round-two breakdown page section by section: its headlines, what each block reads, figures copied from records, Ben's first-person lines, and defects.
An agent
read:kit-modules
Pooled, also verifiedThe data
For each picture on the approved pages, say what the component kit holds today and how far it is from the approved picture, checking the brief's table against the code.
It checked the kit's modules against the approved pictures, found the brief's table wrong on two rows, and ran the kit's tests (1,387 pass).
For 4 of the 10 pictures the kit already has working, tested parts: the hundred squares, the weekly line, the money pair and the Texas map.
An agent
read:base-dashboard
Pooled, also verifiedThe data
Map the approved round-two dashboard block by block: what each block shows, what it reads from the stored data, and any defect.
It mapped the dashboard's 94 drafted sentences and every stored path, and found places where the docs and the built page disagree.
Every figure on the face is read by path from analysis.json. No figure is marked "recorded, not stored".
An agent
read:owed-register
Pooled, also verifiedThe data
List every item on the owed register, check at the code and data whether each is already paid, and say what each waits on.
It checked the 42 owed items against the code and data on both branches, confirmed the ruling counter still misses the lead's rulings after L111, and found a ruling that calls Tyler's 110 clients 'in care' when the database has 99 in care that day.
An agent
read:look
Built
Say what is decided about the look of the real pages, what of it exists in code, and what is still open.
It listed the colours, typefaces and open design questions, and ran the token check (108 pairs, no failures).
Decided and in code: the dashboard keeps the Field guide skin (V3, V34).
An agent
read:orchestration-record
VerifiedThe data
Find the true record of how the work was orchestrated, check every count about it that the breakdown page prints, and say how stale it is.
It counted the workflow records itself and found the page's counts a day stale, 'agents' used for launches, two runs marked completed whose agents never returned, and a line in the lead's own message that the record contradicts.
Today they hold 68 runs and 607 agent launches (58 completed, 6 killed, 4 with no run record).
An agent
read:v37-reach
Verified, also pooledThe industry
Find where the pages frame the problem as Tyler's, and what the repo can truthfully support about the problem as Ben now states it (V37).
It found that Ben's new wording and the approved page's biggest sentence both say something the project's research forbids, and listed what the state's published rules do support.
What the repo supports from published sources is the report side. The state scores each contract and never an office.
A check of the lead's
Both reports are strong, and the V37 reader found something that matters. Before I rely on it, I'm checking its three load-bearing claims at the source.
A message of Ben's
A message of Ben's
A message of Ben's
An agent
build:s1-purpose
PooledThe industry
Build section 1 of the breakdown page on Ben's purpose: his reason in his words, three small pictures of what he learned about the industry, and what the page does.
It built the section from stored figures and the verified rules, and caught a heading that credited the state with a review an outside body ran.
The opening passes the five-second test. The largest words are Ben's reason, and the bold line under them names the dashboard and the email.
A message of Ben's
A message of Ben's
An agent
research:stored-and-unshown
Pooled, also verifiedThis world
Find what the modeled world already holds about McAllen that its view does not show, and which of it is a light issue worth a conversation.
It queried the modeled database and the stored figures, found a real McAllen pattern (the attendant who covers most of its fill-in visits is not on the phone app), and showed that the approved view still overcounts the visits waiting to be typed in, a fault an earlier skeptic had raised.
the attendant who covers most of its fill-in visits was never set up on the phone app, so 70 in 100 of McAllen's fill-in visits since March 1 were not clocked electronically, against 12 to 33 in 100 at the other six.
2 web searches
An agent
research:the-projects-own-research
PooledThis world
Search the project's own research for verified facts that could pad out a well-run branch, and McAllen in particular.
It searched the research reports and the database and traced last winter's $4,150 loss to one client's change of health plan on Jan 1.
The best light issue is already in the modeled world. Last winter's $4,150 loss began on Jan 1, 2026, the day the state's new combined Medicare and Medicaid plans started in Hidalgo County (verified).
An agent
research:published-sources
Sourced, also verifiedThis world
Find the light, everyday issues a well-run home care branch in South Texas really has, according to published sources.
It read the Weather Service, state letters and rules, and Census figures, and found the modeled January storm coded McAllen's visits as ice when no ice fell in the Valley, on days before the real freeze, and a stored label calling $13 a legal wage floor.
Warning 1: the modeled storm is on the wrong days. It is coded as ice (130 C), and the Weather Service says no ice fell in the Valley. It was a dry, record freeze on January 26 and 27.
and 2 pages not shown
9 web searches
A message of Ben's
A check of the lead's
Reading the two published sources myself (the Valley freeze and the state's wage letter) before I rely on them.
A message of Ben's
An agent
build:m1-mcallen-plan-change
Pooled, also verifiedThe data
Rebuild McAllen's view around one client's change of health plan on Jan 1, from the lead's build order and its checked figures.
It drew the four clients' plan change and, with three read-only queries of the database, found two of the lead's sentences slightly untrue and one line of the picture running past the client's last visit.
I ran three read-only queries on data/home-care-ops/pecan-hollow.sqlite, only to check the three points above. Nothing from them prints on the page.
A check of the lead's
The view is built and reads well. The builder caught three places where my wording was slightly untrue, and I've confirmed them in the database.
A message of Ben's
A message of Ben's
An agent
build:mcallen-out-of-the-storm
PooledThis world
Take McAllen out of the modeled January winter storm, rebuild the data once and run every check, without committing anything.
It narrowed the storm to five branches and rebuilt the world: 41,640 figures agreed and 7 of 7 planted errors were caught, though the receipts could not be written until the new data was committed.
The note names the five branches and says El Paso is not in it, and McAllen is not in it because the Rio Grande Valley had a dry freeze and no ice.
An agent
check:mcallen-out-of-the-storm
VerifiedThis world
Test eight claims about the storm change with your own runs, trusting neither the builder nor the order.
It rebuilt the world itself, compared 1,269,260 rows of the six other branches (none differ), reran the second program and the planted errors, and found all eight claims held.
1,269,260 rows of HOU SAT FTW AUS ELP TYL compared, 0 differ.
A message of Ben's
A message of Ben's
A message of Ben's
A check of the lead's
The state matches the record. Dataset 54185a2c, 41,640 figures agreed, kit 1,387 of 1,387, lint and tokens green, room up, main untouched.
A check of the lead's
Everything the orders cite now resolves (19 stored paths, 42 files, 0 misses).
A message of Ben's
A check of the lead's
Nearly there: the plan is redrawn without the Tyler step, and I'm checking the script that would prepare the build before I report.
A message of Ben's
A message of Ben's
An agent
sheet:letters-six
Pooled, also verifiedThe data
Write the sheet for the six other emails and their pictures: every sentence and figure they print, where each is stored, what is owed and their known faults, building nothing.
It traced the six other letters (401 paths, 56 pictures) and found moved figures missing from the merge note and lines under the branch names that grade a branch.
The order (section 4) and L170 pick San Antonio's axis by branch id; L234 and SL1 5a make it by the file's rule. Same answer today (50 to 100 at San Antonio, 70 to 100 elsewhere).
An agent
sheet:view-ftw
Pooled, also verifiedThe data
Write the sheet for Fort Worth's view: every sentence and figure it prints, where each is stored, what is still owed and its known faults, building nothing.
It traced the view against the built page and found a sentence that could read as an industry average and moved figures missing from the merge note.
I relied on the built page at e6d14c6 and on git diff e6d14c6 HEAD being empty over the script, the page, the shot and data/home-care-ops/.
An agent
sheet:first-screen
Pooled, also verifiedThe data
Write the sheet for the dashboard's first screen: every sentence and figure it prints, where each is stored, what is still owed and its known faults, building nothing.
It traced all 115 stored paths the screen reads (none reach nothing) and found a fold that gives Tyler a fail, a present-tense claim on a past condition, and moved figures missing from the merge note.
None of the stored paths the sheet names: 74 written, 115 resolved one by one through common.ts's live() loader (X expanded to the seven branches), 0 reach nothing.
An agent
sheet:package
Pooled, also verifiedThe data
Write the sheet for the emails' package: every sentence and figure it prints, where each is stored, what is still owed and its known faults, building nothing.
It checked the letters' stored words and conditions, recounting one figure in the database, and found the page checked only part of the approved emails' conditions.
my read-only recount from Jul 1 to Jul 14 reproduces the stored 14 clients, 136 visits and 498.5 hours, so since is the 14's window start
An agent
sheet:view-elp
Pooled, also verifiedThe data
Write the sheet for El Paso's view: every sentence and figure it prints, where each is stored, what is still owed and its known faults, building nothing.
It traced the view's stored paths and found two different stored totals both called El Paso's authorized hours, one printed as 'planned' and one read in the fold.
Whether 'counts only the visits somebody worked' is exact against the state's formula. The research (G-page-sentences 1.4 and 1.8) counts transactions and says a clocked hour nobody worked is counted.
An agent
sheet:section-3
Pooled, also verifiedThe data
Write the sheet for section 3 of the breakdown page, the data and how it was made and checked: every sentence and figure it prints, what each stands on, what is owed and its faults, building nothing.
It traced all 170 paths the section prints and found a sentence about the lead agent that a later record contradicts, a claim stronger than its source, and 'agents' used for launches.
None of the 170 paths the built page traces in #link-3 reaches nothing: all 170 resolved today with the kit's resolver.
An agent
sheet:view-sat
Pooled, also verifiedThe data
Write the sheet for San Antonio's view: every sentence and figure it prints, where each is stored, what is still owed and its known faults, building nothing.
It traced the view's 137 stored paths and found the build order's week-by-week line could not draw San Antonio's approved picture.
None of the 137 distinct paths the built San Antonio view traces reaches nothing.
An agent
sheet:section-2
Pooled, also verifiedThe industry
Write the sheet for section 2 of the breakdown page, the rules and how they were learned: every sentence and figure it prints, what each stands on, what is owed and its faults, building nothing.
It checked the section's rule sentences against the verified rules list and found four that said more or less than their sources.
The lead agent's definition: the P5 order and L159 say it prints once. The joined page prints it twice, in sections 2 and 3 (also in base-breakdown).
An agent
sheet:view-mca
Pooled, also verifiedThe data
Write the sheet for McAllen's view: every sentence and figure it prints, where each is stored, what is still owed and its known faults, building nothing.
It checked the rebuilt McAllen view against the database and its sources and found a label claiming more than it checks, a rule worded wider than its source, and a fold that contradicts the face.
Since L228 the dashboard mounts M1: 23 drafted sentences, 23 recorded keys, 42 data-recorded marks on the dashboard page
An agent
sheet:section-1
Pooled, also verifiedThe industry
Write the sheet for section 1 of the breakdown page: every sentence and figure it prints, what each stands on, what is owed and its faults, building nothing.
It checked section 1 against its order, the rules and the stored dates and found the date line's shading wrong about when visits can be corrected.
The fold's heading for 1b: the order says 'what the state itself measured'; the page says 'what the state's own quality review measured' (L198).
An agent
sheet:view-hou
Pooled, also verifiedThe data
Write the sheet for Houston's view: every sentence and figure it prints, where each is stored, what is still owed and its known faults, building nothing.
It resolved the sheet's stored paths and found a notice sentence that dropped the rule's 'temporarily' and a moved rank missing from the merge note.
I resolved every backticked stored path in the sheet with load() from scripts/dev/wireframes/common.ts, read only (101 resolve, 4 reach nothing), and checked 12 repo file paths on disk (all 12 exist)
An agent
sheet:view-aus
Pooled, also verifiedThe data
Write the sheet for Austin's view: every sentence and figure it prints, where each is stored, what is still owed and its known faults, building nothing.
It resolved all 75 stored paths and found Austin's verdict partly false about when the money was paid, and the fold naming a report that may not list Austin's take-back notices.
None of the 75 stored paths the sheet cites in full: check-orders.ts over the sheet gave 75 tried, 15 files looked for, 0 reach nothing.
An agent
sheet:letter-tyl
Pooled, also verifiedThe data
Write the sheet for Tyler's email and its pictures: every sentence and figure it prints, where each is stored, what is owed and its known faults, building nothing.
It traced Tyler's letter and its pictures and found a picture headline standing on no check that its claim holds.
Order P6's path list names branches[id=TYL].story.landlines.attendants; the script never reads it (it counts the roster instead)
A message of Ben's
An agent
cold:sl2-sl3
VerifiedThe data
Read the orders for Tyler's view and for its checker as they would on the first morning, check everything they name, and write nothing.
It found 20 faults, two that would have stopped a build (a finish mark set past the edge of its picture, and point labels whose places nothing stored), and warned that corrections made after the builders' copy was cut would not reach them unless carried over.
Ran a node read of drilldown.events at every branch: TYL has 2 (clocked_electronically_share, landlines_removed), SAT has 4 (first_export_clean_share, members_changed_plans), the other five have 0.
An agent
cold:sl1
VerifiedThe data
Read the build order for the views file as its builder would on the first morning, check that every file, key and figure it names exists and is what it says, and write nothing.
It checked the order against the code and stored data and found 11 faults, one that would have stopped the build and a ruling it misquoted.
Read heading_to for every graded contract in data/home-care-ops/analysis.json with a node one-liner: TYL 74, 74, 82; HOU 84, 85, 85, 87; ELP 90, 90, 93; MCA 95, 96, 96, 96.
A check of the lead's
Four writers report they changed something besides their sheet. The tree shows only the new sheets, so I'm reading their reports whole to find out what and where.
An agent
store:views-file
Pooled, also verifiedThe data
Store Tyler's nine sentences and every chart place for the seven branch views in a file of their own, have the second program make it all again, and prove the check can fail.
It built the views file and the second program's new group: 44,415 figures agreed, and both errors it planted on purpose (a point moved one unit, one word changed) were caught.
The word plant: 'care' changed to 'work' in TYL's verdict, named in group h's words check.
A check of the lead's
That command went wrong: the shell read the backticks in my text as commands, and one word (label) is a real Windows command. Checking that it changed nothing before I go on.
An agent
cold:shared-rows
VerifiedThe industry
Read the orders for the shared rows and the shapes as their builder would, check everything they name, and write nothing.
It found 17 faults, two that would stop a build, and that the order asked for a source on one rule whose only source is a document the project forbids citing.
Before starting, the builder of the shared rows needs a few things no order gives.
An agent
cold:the-end
VerifiedThe data
Read the orders for the assembler and the checks, and the night's plan, as the assembler and then a checker would, check everything they name, and write nothing.
It found 17 faults, four that would stop a build, chief among them that the chain deletes files the stored sentences read from.
What no order gives the assembler is the contract between the pieces and the page they join into.
An agent
cold:pictures
VerifiedThe data
Read the orders for the shared pictures and the shapes as a builder of each of the three sets would, check everything they name, and write nothing.
It found 17 faults, one that would stop a build: no order let a label carry the path of the figure it prints.
A builder of the shared pictures needs things that no order gives.
An agent
cold:a-view
VerifiedThe data
Read the orders for a piece, and the sheets for San Antonio's and El Paso's views, as their builders would, check everything they name, and write nothing.
It found 18 faults, two that would stop a build: El Paso's approved headline could not be stored under the rules, and no order gave a branch its own pictures.
What a view builder needs and no order gives is a plain statement of how its data gets into views.json.
An agent
cold:a-section
VerifiedThe data
Read the orders for a piece, and the sheets for section 2 and the first screen, as their builders would, check everything they name, and write nothing.
It found 19 faults, one that would stop a build: nobody was told how the first screen places the branch views in its own grid.
No order sets the joins between a piece and the things around it.
An agent
cold:a-letter
VerifiedThe data
Read the orders for a piece, and the sheets for the emails' package and Houston's letter, as their builders would, check everything they name, and write nothing.
It found 23 faults, four that would stop a build, among them that nobody was set to build the email itself.
No order gives a builder the email itself as one shared thing
A check of the lead's
The machinery reads cleanly, and its changes to existing files are small and exactly what the order allowed. Running the whole gate myself now.
A message of Ben's
A message of Ben's
A message of Ben's
An agent
machinery:sl1b
Pooled, also verifiedThe data
Grow the views machinery so the night's builders can add their pieces safely, each made again by a twin in the second program, without changing one stored value.
It extended both programs, showed each stops on four kinds of bad input, and left every stored value unchanged, with 44,437 figures agreed.
2,916 of 2,916 start leaves hold the same value, 0 changed, 0 missing, 37 added.
An agent
view:tylers-view
Built
Build Tyler's branch view for real, as components of the kit written out as plain pages, from the stored words and places; it becomes the pattern the other views copy.
It built the view and five pages to check it and matched the approved picture within a few pixels, except at the finish, where it flagged that the stored places hid both range brackets behind the finish marks.
Everything else matches the picture within a few pixels, in the real type and colours.
An agent
cold:re-end
VerifiedThe data
Read the rewritten night orders again as the package builder, the letters builder, the assembler and a checker would, and say what they would still have to guess.
It found 18 things still unsettled, three that would stop a build, and that the emails' approved opening line would promise figures behind sentences that had none that night.
No order says who may change the receipts door so that a page_words entry can pass it.
An agent
cold:re-piece
VerifiedThe data
Read the rewritten night orders again as four builders would (the shared rows, section 2, El Paso's view, the first screen) and say what they would still have to guess.
It found 21 things still unsettled, two that would stop a build, among them that the receipts' checking door refuses the new kind of receipt the orders asked for.
A builder needs an owner for the one file every stage after B2 reads through a checked door
A message of Ben's
A check of the lead's
Windows Update is confirmed paused in the registry (from 21:14 tonight to Oct 1, no restart pending).
An agent
check:tylers-view
VerifiedThe data
Read every sentence, figure and mark on Tyler's built view against the stored files, the database and the verified rules, trusting no one, and say what you would not sign.
It remade Tyler's grades from the visit records with its own query, found every printed figure equal to its stored value, and would not sign three of Ben's approved sentences: two say more than the state's rules allow, and one states our model's projection without saying it is the model's.
I would sign the figures, the places, the second program's agreement and the untouched analysis.json.
An agent
shared:receipts-words
PooledThe data
Store the sentences of sections 2 and 3 that stand on the receipts or the second program's record, and the two bars of counts, in the receipts file at the chain's end.
It stored 38 sentences and labels with the four word changes the lead had ruled, and found that the day of the counts was read off the universal clock (UTC), so a Texas evening printed as the next day.
The receipts are process facts checked by receipts.ts's own gate; the second program does not make them or these words again
An agent
shared:shared-rows
Pooled, also verifiedThe industry
Store the rows every piece shares (the state's rules with their sources, the note on dollars, the steps of a notice and more) and the makers of chart places, in both programs.
It stored 16 rule rows with their document, revision and date, refused three sources it could not stand behind, and had the two programs agree on 45,817 figures.
What I checked holds: every stored place and word was made again by the second program (45,817 figures agree), both programs' makers give the approved scripts' own places on the four real cases, and all three planted errors were caught.
An agent
shared:pictures-marks
Built
Build the shared pictures of squares, dates on a line and a large figure that more than one piece draws.
It built the three pictures with pages to check them, and flagged where two groups sharing a row would print their words on top of each other.
Where two groups share a row, their brackets overlap (the second stepped 8px out), which reads fussy at El Paso's last two rows and in Tyler's hundred
An agent
shared:pictures-bars
Built
Build the shared split bar and the chart of where a contract finishes by the day the fix is in place.
It built the two pictures to the approved drawings and listed the keys it had to add for them.
The split bars read cleanly at every real case, in both places.
An agent
shared:pictures-work
Built
Build the drawing of how the work was run, its small mark, and the steps of a notice, with every figure read from its stored path.
It built the drawing of the agent work, its small mark and the steps of a notice, and noticed the approved section 3 page showed figures older than the data.
In section three's six boxes at about 197 pixels the counts are tight
An agent
Shapes agree: makers and pictures
Built
Make the six makers of chart places, in both programs, and the eight shared pictures take one shape, so the piece builders could hand one straight to the other.
It made both programs' makers and the pictures agree, ran both programs on 39 real cases with all 2,103 compared values matching, and passed the full gate.
the Python twins give the same keys. I ran both programs on 39 real cases and all 2103 compared values matched.
An agent
piece:first-screen
PooledThe data
Build the dashboard's first screen for real, from its work order and sheet, in a copy of the project of its own.
It built the dashboard's first screen (the map with a label for each branch, the money across the agency, the headline) from stored figures, storing 34 sentences with the conditions they stand on and a twin in the second program that makes them again.
Nothing printed without a path: every figure carries data-path and every sentence data-said (the guard checks both)
An agent
piece:view-sat
PooledThe data
Build San Antonio's branch view for real, in Tyler's pattern, from its work order and sheet, in a copy of its own.
It built San Antonio's view in Tyler's pattern, storing 16 sentences with their conditions and a twin that makes them again, and reported where the pattern's stored places crowd at San Antonio.
San Antonio's view now reads like Tyler's: the same frame, type, colours and folds, with every sentence stored and every figure traced.
An agent
piece:view-hou
PooledThe data
Build Houston's branch view for real, in Tyler's pattern, from its work order and sheet, in a copy of its own.
It built Houston's view (the steps of the payer's notice, the attendants behind half the visits) with 10 stored sentences and their twin, and flagged a stored note that names the state's appeal clocks beside a health plan's notice.
Houston's own parts read as the approved drawing did, now in Tyler's look. The attendants flow down into half of the visits, the steps carry the mark on the first notice, and the fold prints what the lead ruled.
An agent
piece:view-aus
PooledThe data
Build Austin's branch view for real, in Tyler's pattern, from its work order and sheet, in a copy of its own.
It built Austin's view with 10 stored sentences and their twin, drawing the card of the one thing to do in its own file because the pattern would not draw it without a headline.
Blocks 2 and 3 themselves (the card, the large figure, the bar with its numbered sentences) read as finished work next to Tyler's view, not as a wireframe.
An agent
piece:section-1
PooledThe industry
Build section 1 of the breakdown page (why Ben built this) for real, from its work order and sheet, in a copy of its own.
It built section 1 of the breakdown page (the state's report row, El Paso's blind spot, the day a quarter's visits lock) from stored figures and rules, storing 28 sentences with their twin.
The source of the monthly instruction (s1_1c_plain_2): views:shared.rules[id=monthly_instruction].source_said is null, so the sentence prints with no source (decision 11)
An agent
piece:view-mca
PooledThe data
Build McAllen's branch view, with the change of health plan, for real from its work order and sheet, in a copy of its own.
It built McAllen's view with its picture of one client's change of health plan and the rules behind it, storing 30 sentences with their twin, and checked its groups of clients against the database with its own queries.
Checked by query: the same four clients, the same three in time and the same one late (her 276 payer rejections, the others 0)
An agent
piece:section-3
PooledThe data
Build section 3 of the breakdown page (the data, and how it was made and checked) for real, from its work order and sheet, in a copy of its own.
It built section 3 of the breakdown page with 31 stored sentences and their twin; as the lead had ruled, the sentence saying nothing was added to the data until the lead agent had run the checks again is stored but does not print, because the project's record shows 16 commits made with no lead agent in the loop.
Stored with the condition that that record row says nothing (views:pieces.section-3.records.lead_gap.said is null); the row says 16 today, so the sentence does not print and the lead agent's row of the drawing is not drawn.
An agent
piece:section-2
PooledThe industry
Build section 2 of the breakdown page (the rules, and how they were learned) for real, from its work order and sheet, in a copy of its own.
It built section 2 of the breakdown page (four of the state's rules shown working at Tyler, and how the rules were learned and checked), storing 62 sentences with their twin, and changed 9 of them and six labels as the lead had ruled.
Nothing the sheet asks for is left unsourced.
An agent
piece:view-ftw
PooledThe data
Build Fort Worth's branch view for real, in Tyler's pattern, from its work order and sheet, in a copy of its own.
It built Fort Worth's view with its own picture of how many of its newest attendants' visits were typed in by hand, against the whole agency, storing 10 sentences with their conditions and their twin.
Four sentences carry conditions beyond the sheet's column, each true today and stated in method.conditions_added: buys_headline also stands on the first action's kind being daily_clock_out_check, buys_meaning on its worth_path naming the large figure's path
An agent
piece:view-elp
PooledThe data
Build El Paso's branch view for real, in Tyler's pattern, from its work order and sheet, in a copy of its own.
It built El Paso's view with its two grids of planned hours (El Paso's and the other six branches'), storing 16 sentences with their conditions and their twin.
The grids' words sit beside their rows (the shared picture's and maker's rule, L267: a label on every mark, no legend), where the approved drawing printed a small key under each grid.
An agent
piece:package
PooledThe data
Store what the seven emails read that no other file stores, and build the letter component every email is mounted on, in a copy of its own.
It stored what the seven emails say that no other file stores (49 sentences, each with the conditions it stands on, and a twin that makes them again) and built the letter component every email is mounted on.
Twenty conditions beyond E1's 81 are stored (all stated in method.edits): this quarter is the one in progress where a path types SFY26-Q4 (Tyler 6, San Antonio 2, Fort Worth 3 and 6, El Paso 1, McAllen 2)
An agent
letter:letters-rest
PooledThe data
Build the six other emails on the package's letter component, each with its head, six sentences, its chart of contracts and its attachments, in a copy of its own.
It built the six other emails on the letter component, each with its head, its six sentences and its chart of contracts under the right sentence, plus the attachments that were stored. Fort Worth's and El Paso's were not stored yet, so those two letters printed none.
The letter itself reads well, in the look of Tyler's letter: the head, the margins, the one fold with its ring, the headline over the chart, the stored attachments and the note on dollars.
An agent
letter:letter-tyl
PooledThe data
Build Tyler's email whole, with its six pictures, on the package's letter component, in a copy of its own.
It built Tyler's email whole on the package's letter component, with six pictures drawn from stored places, and moved one chart figure that the shared maker had put on another contract's point.
After the shared fix chart's maker, one rule of my own (in letter-tyl.ts and its twin) moves a figure the maker flipped above its point onto the right of its own point where it would sit on the mark above it, because the maker put Plan B's 82 on Plan A's 83 at Jul 22
An agent
assemble:whole
Built, also pooled
Put both pages together whole from the landed pieces, run the whole chain once, shoot every page, and say where pieces do not read as one page.
It put every piece together into the dashboard and the breakdown page, stored the six words no piece owns with their twin, ran both programs over everything (52,946 figures agreed) and shot every page state.
The frame joins the pieces, and both pages read top to bottom as one piece of work.
An agent
check:letters-a
VerifiedThe data
Read four of the seven emails against the stored files, the database and the rules, trusting nobody, and say what you would not sign.
It checked every figure, sentence and place in four of the emails with its own resolver and database queries, signed every figure, and found that the dollar note's claim rests on no verified source and that several sentences could be taken wrongly.
Everything else holds by my own checks. Every figure is the stored one, every stored sentence prints word for word, every place is where its figure belongs, and every fact I could query in the database agrees.
An agent
check:breakdown-a
VerifiedThe data
Read sections 1 to 3 of the breakdown page against the stored files, the database, the rules and the project's records, trusting nobody, and say what you would not sign.
It checked sections 1 to 3 of the breakdown page at the source and found two of the page's sentences about how the work was checked untrue as built: the lead agent's reading of every report, and the build gate.
Everything else in my part I checked at the source and it holds: every figure and place, every stored sentence word for word, every rule against the inventory, and every world claim against the database.
An agent
check:letters-b
VerifiedThe data
Read three of the emails against the stored files, the database and the rules, trusting nobody, and say what you would not sign.
It checked three of the emails against the database and the rules and found McAllen's "Plan A visits" misleading, Fort Worth's "attendants typed in" wrong about who keys, and a count of rejections stated as the state's own.
These hold by my own queries. AUS: 116 notices on Dec to Feb care, next highest ELP 26, 111 records missing, 42 of those under appeal, 43 under appeal in all
An agent
check:the-whole
VerifiedThe data
Check the whole: the title and path, run the second program again, read the diff of the night, and sweep every stored file for real names, trusting nobody.
It ran the second program again (52,946 figures agreed, 9 of 9 planted errors caught), linted the code and swept every stored word for real names, and found a health plan's name in three stored rows and that no planted error touched the words of the first screen, the sections or the letters.
52946 figures compared, 0 disagreements, 427 checks passed, 0 failed.
An agent
check:dashboard-top
VerifiedThe data
Read the dashboard's first screen against the stored files, the database and the rules, trusting nobody, and say what you would not sign.
It re-made every figure and place on the first screen with its own page parser and database queries and signed all of them, then held each sentence to the verified rules and found five things it would not sign.
Every figure, place and pick on the screen I would sign. They agree with my own queries and my own arithmetic to the cent and to the third decimal.
An agent
check:views-a
VerifiedThe data
Read three branch views against the stored files, the database and the rules, trusting nobody, and say what you would not sign.
It re-made every contract grade and ranking behind three branch views from the database and found them right, then found two false sentences: San Antonio's "By when" and "the middle branch" beside Tyler's and Houston's ranks.
My own database recomputation agrees: every contract grade for the three service areas (Q2, Q3, today) and all three levers for all seven branches match the stored values to the second decimal.
An agent
check:views-b
VerifiedThe data
Read four branch views against the stored files, the database and the rules, trusting nobody, and say what you would not sign.
It re-made the figures of four branch views and checked 426 stored places, all right, then found McAllen's "the middle branch" false, Austin's count of notices missing its scope, and McAllen's rules cited in plain words.
Every figure I made again matches: 13 scores so far, 21 lever values, turnover 69 in 100 and highest
A check of the lead's
The seven checkers returned. Reading all seven reports whole before settling anything.
A check of the lead's
The chain is green again: the same bytes twice, 52,967 figures agreed with none disagreeing, nine of nine caught.
A message of Ben's
A message of Ben's
A check of the lead's
You're right, and I agree it should be a pass of its own. I shot the breakdown page at 2560 after reading your message.
A message of Ben's
A message of Ben's
A message of Ben's
A message of Ben's
A message of Ben's
An agent
ground:pictures
Built
Read only: check the morning list's picture items at their source, and name each fix's files and smallest change.
It traced each picture item to the code behind it, looked at the committed shots, and corrected two items: by its estimate from the stored places, the tucked labels sat on about 10 to 12 of 21 rows, not 4, and Houston's wrapping table was on the dashboard, not in section 4.
The floor comes from axisFor (analyze/groups/a-score.ts L896-904): the ten strictly under the lowest mark, with the pass mark 80 always one of the marks. So the floor can never be above 70
An agent
ground:rules-sources-and-small-words
VerifiedThe industry
Read only: check the morning list's items on the rules' source lines and small words at their source, and name each fix's files and smallest change.
It traced each rule's source line, McAllen's rule sources and the county sentence to the code and records that make them, and found project docs calling the unverified county sentence verified, and three more rules whose source lines hold raw dates, none printed yet, that a narrow fix would miss.
So the date can be said in words while the record still proves it: set figure to "effective Mar 12, 2025" and written to the form the lines hold.
An agent
ground:chain-and-landing
Built, also verified
Read only: work out how a change lands (the chain of commands, the copies, the room) and check item 42's odd records at their source.
It traced the chain of commands a change must run through, the copies and the room, and checked item 42 in the database: it confirmed the real health plan's name sat only in the stored views file with no guard reading it, and found that the 'waiting to be keyed' count overstates unkeyed visits at every branch, not only McAllen.
Not printed: grep of the whole pages, analysis.json, reconcile.json, receipts.json and E1-emails.json finds no plan name, and the only hits in views.json are these three.
An agent
ground:san-antonio-and-the-ladder
VerifiedThe data
Read only: check the morning list's items on San Antonio, the ladder's stored words and the dollar note at their source, and name each fix's files and smallest change.
It checked the stored words of the one thing to do at all seven branches against the stored data, confirmed San Antonio's "By when" false, and found the dollar-note item only half true and a one-day slip in how a lock date is said.
Every fixture in tmp/home-care-ops-fixtures holds the same false San Antonio by-when (Today, with 28 to 245 visits already lost), so the fixtures never exercised the "nothing lost yet" case the old phrase is true for.
An agent
ground:rules-section
VerifiedThe data
Read only: check the morning list's items on the rules section and the receipts' words at their source, and name each fix's files and smallest change.
It checked each item on the rules section at its source, confirmed against the project's own record that the lead-agent sentence was false, and found that one listed item was no defect and one proposed fix would guard nothing.
The sentence is stored in views.json and prints on the page, and the record contradicts it.
An agent
ground:data-section-and-gates
VerifiedThe data
Read only: check the morning list's items on section 3, the build's gates and the markup at their source, and name each fix's files and smallest change.
It read the page scripts and the second program's record and confirmed that the scripts that write the pages never checked the second program's result and that nothing checked the views file when a page was built, and showed where a fix must go so Ben's sentence keeps printing.
So reconcile.json can go stale, fail or vanish after the receipts ran, and the page still builds.
An agent
cross-check
VerifiedThe data
Re-check at the source the readers' claims the fixes rest on, try to show each wrong, and find where fixes collide, what order they must land in, and what no reader covered.
It re-checked the readers' key claims at the source and found every one holding, five of them with corrections, mapped which fixes touch the same files and the order they must land in, and named the parts of two items no fix covered.
So views.json is proven at receipts time, just never at page-build time.
A check of the lead's
I've changed my position. I also checked whether it can be built from real records, and it can.
A message of Ben's
A check of the lead's
You're right to disagree, and the project's own record backs you: that is not how the rules were learned.
A message of Ben's
A message of Ben's
A message of Ben's
A check of the lead's
The handoff is read whole. Now the section 11 checks, run myself rather than trusted
A message of Ben's
An agent
build:fx-a
PooledThe data
Change the analysis program and its twin so San Antonio's "By when" is true, the county sentence claims only what was found, and a still-open count per month is stored.
It changed the first program and its twin so San Antonio's "By when" reads "Today. More of their visits lock every day.", the county sentence says "We found no per provider county list published in Texas.", and a still-open count per month is stored and rechecked.
Both San Antonio actions now store "Today. More of their visits lock every day." That is true: 185 visits locked, and a lock falls on each of the 95 days from Jul 15 to Oct 17.
A check of the lead's
Job A has been going 27 minutes. It has changed the files its order names, plus two it doesn't: section 3's words (section-3.ts and its twin). I'll see what that is before answering.
An agent
build:fx-pics
Built
Draw versions side by side of three pictures for Ben's eye (the contracts' chart at the quiet branches, "the other six" on the tracks, the map's plates); nothing lands.
It drew side-by-side versions of three pictures for Ben to point at, measuring each on the page; none of them landed.
Measured on the page: as it is now, the words meet the branch's own label on 12 of 21 tracks
An agent
build:fx-d
Pooled, also verifiedThe industry
Rebuild the rules section around Ben's approved opening in its new order, remove the two lists of decisions, and correct box 4's sentence about the lead agent.
It rebuilt the rules section around Ben's own opening in its new order and, after checking the project's record at its source, stored box 4's corrected sentence about the lead agent with its twin.
I verified the record at its source: carryover lines 109 to 111 say the lead worked from a digest and read reports only where a story turned.
An agent
build:fx-i
Built
Make the title's lines carry their stored path, keep section 4's "Read the email" on its page, and give the fold's table room in every window.
It made the title's lines carry their stored path and kept section 4's "Read the email" on its page; Houston's fold table was still 12.6 px too wide, so it stopped as ordered and the lead closed the gap at landing.
Version A is built in full, and Houston's table still does not fit, by 12.6 px, so I stopped there as the order says and did not stack the fold.
An agent
build:fx-c
PooledThe industry
Say every rule's date in words, complete the stored rules for how pictures place their marks, and put Appendix G back on the first rule's sources.
It made every rule's source line say its date in words ("effective Mar 12, 2025") while the record still proves the day, put Appendix G back on the first rule, and completed the stored rules for how pictures place their marks.
Stored result: 23 dated rows, every one said and proved. On the 31 whole pages, 254 printed source lines hold no raw date.
An agent
build:fx-b
PooledThe industry
Say McAllen's rule sources in the documents' own names with the version key never printed, and take a real health plan's name out of three stored record rows.
It rewrote McAllen's rule sources in the documents' own names with the version key never printed, and pointed three stored rows at another line of the research report, taking a real health plan's name out of the stored file; it also flagged that its own new title row had brought in the state contractor's name, which the lead then moved.
fold_rules_2_source now says 'EVV Visit Transaction Rejection Guide, dated Jun 5, 2026'.
An agent
build:fx-e
PooledThe data
Make the dollar note print only at the foot of an email that prints a dollar, and give San Antonio's attachment a still-open column.
It made the dollar note print only at the foot of an email that prints a dollar (four of seven), and turned San Antonio's attachment column into "Still open", adding to the 1,161 in its sentence.
The column now reads 0, 0, 0, 32, 330, 494, 305 and adds to 1,161, the figure in sentence 4.
An agent
build:fx-h
VerifiedThe data
Make the page refuse to build unless the second program finished on the exact files it reads, plant errors in the pages' own words, and guard the stored files against real company names.
It made the second program name the exact views file it checked and the three page scripts check that name before they build, added three planted errors in the pages' own words (12 planted, 12 caught), and built a guard against real company names in the stored files.
In scratch only, with no stored file touched: whole.ts's own build, handed the views as they were before my change, threw 'views.json refused: reconcile.json does not name it' and wrote no page.
A message of Ben's
A check of the lead's
Both views runs took under a second, which is suspicious. Checking:
An agent
skeptic:sk-gate
VerifiedThe data
In a throwaway copy, show that the page builds on files the second program did not finish on.
It made 20 attacks on the build gate in a throwaway copy; 7 got through and it reported 5 holes, among them that switching off the one line that runs the check left every test green and the page printing a changed sentence.
All three exit 1: 'views.json refused: reconcile.json does not name it (the views changed after the twin ran)'. 0 of 39 page files changed.
An agent
skeptic:sk-plants
VerifiedThe data
In a throwaway copy, show that the planted errors in the pages' words, or section 3's flags test, pass when they should fail.
It made 23 attacks on the planted errors and section 3's test; 11 got through and it reported 4 holes, among them that a second program which merely copied the stored words would still catch all twelve plants.
Real run: exit 0, finished true, 53,127 figures, 0 disagreements.
An agent
skeptic:sk-never
VerifiedThe data
In a throwaway copy, show that the guard against real company names in the stored files passes when it should fail.
An agent
read:the-pages
VerifiedThe data
Read every fix on the rebuilt pages, comparing the text of every page state before and after, trusting no builder, order or log.
It compared the text of all 37 page states before and after the fix pass and found 28 of 29 fixes holding; the one that did not was a proof tile still naming its check in the program's words.
Joined with one space, the two parts match his V59 words character for character.
An agent
build:fx-k
VerifiedThe data
Proved able to fail (harness/gate-proofs.log), each file put back by sha256 after: whole.ts line 76 commented out: 6 of 21 fail; wrapped in try/catch: 6 fail; inside a function never called: 6 fail.
A message of Ben's
A workflow run
verify-6a-analysis
To attack the first batch of new dashboard figures (step 6a) from three sides (the arithmetic, a recount from the raw tables, and whether a page could print every value without working anything out), then have separate agents try to disprove each serious finding of the first two; the third side's list went to the lead.
The lead
6a is committed (e86d1eb). Launching an adversarial review of it in the background while I build 6b.
A workflow run
build-branch-page-sections
To build the four sections of the first dashboard's branch page in parallel from stored figures, then have a fresh critic check and fix each section.
Ben
What he wrote is his to release.
A workflow run
r2-per-hundred-one-in-n
To store two new kinds of figure for the redesign (what happened to every 100 visits or approved hours, and phrases like '1 in 5'), each with a Python twin, then have three checkers try to break them and one agent fix what was real.
Ben
What he wrote is his to release.
A workflow run
r1-mock-critics
To test the first redesign mock before Ben saw it: five readers with no context who each saw one screenshot, a mechanical tester, a design critic, and a fact-checker of every sentence and number.
The lead
SAT's cliff reads cleanly now, and the Public Sans variant gives plain numerals. One clip left: Tyler's second audit note runs off the right edge. Fixing, then producing the reference screenshots.
A workflow run
v2-build-brief-appendices
To write the seven appendices of the v2 build brief, have a checker try to break each against the real code, data and packages, revise each, and have two critics read them as a set.
Ben
What he wrote is his to release.
A workflow run
v2-harmonize-and-foundation
To bring the seven appendices into line with the lead's rulings, land the first new stored figures for the v2 first screen with Python twins, check the result, then build; only the data builder ran, because the check blocked the foundation.
The lead
Understood, I'm letting it work. I had already written the viewer when your message came in, so I tested it once and then shut it down. It works. [...] From that one test: all seven appendix revisions and the first data commit are done, and the conformance check is still running. The builders and the reviewer come next.
A workflow run
v2-s1-foundation
Fix the last conflicts in the build plan's appendices and build the dashboard kit's foundation (colours, fonts, data doors, honesty guards, workbench), then have an outside reviewer check it and a repairer fix what it found.
The lead
The workflow is resumed and running again. [...] Still queued: 1. The checker confirms all 30 rulings landed in the appendices. 2. Three builders run in parallel: the rest of Tuesday's data, the kit's tokens, and the foundation plus the workbench page. 3. A reviewer who did not write the code checks the result.
A workflow run
v2-s2-hero-three-skins
Build the first page pieces and the San Antonio sample panel in three candidate looks, run two critic-and-repair rounds on it, and build the gate 1 page where Ben picked a look.
The lead
The build passed. One thing in the route table needs a look: the workbench page /lab is listed in the production build. I'm checking whether it prerenders as a 404, as the builder claimed.
A workflow run
v2-s2-revision-1
Apply the lead's own revision of the sample panel, run one critic-and-repair round, and refresh the gate 1 page.
The lead
S2 finished with every check passing, after two critic rounds that logged 30 and then 26 defects. I'm going to look at the final shots myself before any of it reaches you.
A workflow run
v2-s3-s4-first-screen
Make Ben's pick, Field guide, the only look, build the first screen's pieces and compose it at /v2, run two critic-and-repair rounds, and build the gate 2 page.
Ben
What he wrote is his to release.
A workflow run
s5-branch-room
Store one new figure with its twin in the second program, build five modules for each branch's room, and compose the seven rooms under the branch strip.
Ben
What he wrote is his to release.
A workflow run
s5-domain-grounding
Research how Texas agencies really handle payer-mismatch rejections, the fix, the two deadlines, the state's vocabulary and undelivered hours, then audit the page's sentences against what was found.
Ben
What he wrote is his to release.
A workflow run
why-device-comps
Build three different design answers to 'Why is San Antonio failing?' from stored figures, and a page for Ben to point at one.
Ben
What he wrote is his to release.
A workflow run
s5-copy-corrections
Apply the audit's lists of wrong and badly worded sentences to the page, then have a fresh reader check the rendered page as a Texas operations director would (a resend: the first launch failed to parse).
The lead
The room is real and the structure works. It also carries exactly the copy defects the research flagged, including my 781 error sitting next to the correct 1,308. Fixing those now.
A workflow run
s5-scope-rule
Make every panel about the 24 clients take its figures only from their own stored record and every branch panel only from the branch's, then have a fresh reader re-run the arithmetic.
The lead
Both confirmed exactly. May 1 + 95 days is Aug 4, and that date is stored on the cohort object the page isn't using. Fixing the systemic rule, not the symptoms.
A workflow run
s5-class-sweep
Fix each remaining kind of error on all seven branches at once, then have a final fresh reader check every sum on the page.
The lead
A workflow run
project3-cold-read
Twelve readers who knew nothing about the project read the real dashboard and a page of three draft answers to 'why is San Antonio failing?' cold; 120 checkers then tested the first ten problems each reader raised (230 in all) against the stored figures, and one agent merged it all into a report the lead folded into its list of repairs for the next build.
Ben
What he wrote is his to release.
A workflow run
project3-s5b-build
Overnight, to repair the dashboard from the cold read's list and Ben's rulings: new stored figures with second-program checks, six builders, a composer, two fresh critics reading only screenshots, and a fix round.
Ben
What he wrote is his to release.
A workflow run
catchup-verify-s5b
The next morning, after Ben asked to be caught up, to re-run every check the overnight report claimed had passed and try to prove its other claims wrong against the code, the data and git, before Ben's next review of the page.
Ben
What he wrote is his to release.
A workflow run
s5c-gate3-comps
To build six drafts Ben could point at (three first screens, three action sections), with the new figures they needed stored and checked, then a fresh critic and a fix round for each page.
The lead
A workflow run
s5d-voice-sheet-factcheck
Before any plain-language rewrite went on the page, to try to prove false each of the lead's 66 rewritten sentences against the stored data, the test states and the rules; 38 came back not plainly true.
The lead
A workflow run
s5c-comps-trapcheck
Read-only, to check every sentence the six drafts could print against the known wording traps and the stored data, so the drafts Ben picked had a fix list ready.
The lead
A workflow run
s5e-map-master
Ben
What he wrote is his to release.
A workflow run
world-research-pass
A read-only research pass, run before the modeled agency was rebuilt, to learn from the state's documents and other public sources how Texas home care really works and test the modeled world against it: ten core questions, eight more earned by three question-finders, a checker on every report, and a completeness critic.
Ben
What he wrote is his to release.
A workflow run
world-research-round-2
Round 2 of the research on the modeled world: fourteen questions that round 1's critic and the lead judged still open, each researched from the state's documents and the web, each attacked by its own checker, then a critic.
The lead
The shell truncated the long command. I'll write the synthesis with the file tool and launch round 2 in the same step.
A workflow run
world-research-round-3
Round 3: the last three questions round 2's critic earned (span billing, the other plans' suspension duties, whether the state's usage report can be run for the quarter in progress), each researched and attacked, then a critic who ruled no fourth round was needed.
The lead
Round 2 closed all fourteen questions and the critic earned a small third round of three. I'll fold round 2 into the synthesis, commit, and launch round 3 in the same step.
A workflow run
world-brief-appendices
To turn the three research rounds into six reference documents the builders would work from (the agency, the rules, the database design, the calendar and measures, the map, what the page may say), each drafted, attacked by a reviewer, then fixed.
The lead
Round 3 is done. Pulling its three verdicts and the critic's call on whether anything remains.
A workflow run
w1-world-without-stories
To build the modeled agency without its problem stories: its settings, a 31-table database, 1,500 clients, about 250,000 visits, claims and events, and a validator, then a read-only critic and one fix round.
Ben
What he wrote is his to release.
A workflow run
w2a-integrator-prep
Prepare the rebuilt modeled agency so five branch stories could be built at once: tune each branch's healthy level, add the lead's rulings, and set up the story files.
The lead
W1 finished: six agents, about three hours, no errors. Before I trust any of it, I'll pull the critic's verdict and the fixer's gate from the result, then run the gate myself in the build worktree and read the branch table.
A workflow run
w2b-five-stories
Build five branch stories in parallel, merge them, check them against the brief and the tables, and fix what the check found.
The lead
W2a finished in about half an hour with a green gate and every tuned cell within two points. Two items were left to me. I'll read the full return, re-run the gate myself, and land it.
A workflow run
w2c-integrator-second-pass
Apply the lead's rulings after the stories: make the state's 24 graded contracts computable, make Houston's failed quarter real, start hiring warm, amend the brief to the landed figures, then check and fix.
The lead
W2b finished: eight agents, about two hours, no errors. The San Antonio builder surfaced something real and refused to paper over it: with the ruled trickle of 24 members, the branch composite bottoms out near 85 by arithmetic, while the Plan B contract cell, the grain the state actually scores, is at 70.
A workflow run
w3a-restructure-and-port
Split the two programs that work out the figures into seven groups and move every old figure onto the rebuilt database, then check and fix.
The lead
W2c is landed: both trees at 22c4315. Launching W3a from that commit.
A workflow run
w3b-seven-groups
Build the seven groups of figures at once, each in its own copy; for each, a second builder who did not read the first program's code worked the figures out again in Python, and where the two disagreed, a builder in the first program's role settled what a re-read of the brief could settle and sent the rest to the lead.
The lead
W3a is landed at 2bacd74 in both trees, and the seven W3b worktrees are cut at that commit with their datasets generated. Launching W3b.
A workflow run
(no run record) w3c-integrate
Merge the seven groups and apply the lead's shared changes; a Windows Update restart killed it about twenty minutes in.
The lead
Launching W3c with the 24 rulings and the shared changes.
A workflow run
w3c-resume
Finish the killed merge from where it stopped, apply the lead's 24 rulings, then check every stored figure and fix what the check found.
Ben
What he wrote is his to release.
A workflow run
w4-fixtures-and-kit
Rebuild the six test states on the new figures and make the page kit read them, rewording sentences whose meaning changed, then check and fix.
The lead
W3c is landed on both trees at d63607b. Removing the seven W3b worktrees, pointing the W4 script at the landed commit, and recording the landing in the carryover.
A workflow run
w5-map-data
Redraw the Texas map from the Census Bureau's 2025 county file and store and double-check every map figure.
The lead
W4 is landed at c84255e on both trees. Now the main-tree steps for Ben's look (fixtures, the judging room, the /v2 shot) and, in parallel, W5 gets pointed at the landed commit.
A workflow run
w3d-tidy
A tidy pass: store and double-check the last figures the contract-first screen needed, fix the two problems the test states had shown (a deadline's wording and a crash on a quarter's last day), and make every stored figure name where it comes from, then check and fix.
Ben
What he wrote is his to release.
A workflow run
s5f-territory-and-contracts
Store the contract-by-branch figures for a new first screen and build comp A6 (the county-true territory map with the graded contracts up front) for Ben to rule on, then check and fix.
The lead
W3d finished: 3 agents in 84 minutes (a fix round ran). Reading the critic and fixer, then the lead's gate.
A workflow run
s5f-fix2-strip-and-tyler
Carry out two of the lead's rulings on comp A6 (the branch strip follows each branch's contract status; Tyler's name plate moves above its pin), then check and fix.
The lead
Reading the work order's strip clause, the strip module's props, and where the label sides are authored.
A workflow run
s6a-content-spine
Write what the page should say for the company and each branch using only stored figures, with a skeptic checking each writer, then assemble it into one content spine.
Ben
What he wrote is his to release.
A workflow run
s6b-a7-seven-one-anatomy
Build comp A7, a trial first screen with the seven branches as rows sharing one anatomy and the map as the selector, as a picture for Ben to point at.
Ben
What he wrote is his to release.
A workflow run
(no run record)
Store the new figures the first screen needed, each with a twin in the second program; it ended with the session near midnight after five of ten items.
The lead
A7 landed at 665acb5 on both trees and origin. Launching the two parallel slices now: S6c stores the first screen's keys with twins and commits by name; A7b draws the cards without committing.
A workflow run
s6d-a7b-cards
Build comp A7b, the same seven-branch content as A7 drawn as cards, so Ben could compare rows and cards.
The lead
A7 landed at 665acb5 on both trees and origin. Launching the two parallel slices now: S6c stores the first screen's keys with twins and commits by name; A7b draws the cards without committing.
A workflow run
ground-for-wrap
Ben
What he wrote is his to release.
A workflow run
s6c-rest-then-s6e
The first launch of the chain to finish the first screen's stored figures and store its sentences; the lead stopped it five minutes in to add the grounding run's corrections.
Ben
What he wrote is his to release.
A workflow run
s6c-rest-then-s6e
Finish the first screen's stored figures with their twins, then store each branch's cause, insight and actions as sentences built from stored values, each half followed by a critic and a fix round; the second fix round died on the billing stop.
The lead
Both trees are at 62c3117 with item 6 intact. Now the script edits: the ladder path, the base commit, and the three rulings in the S6c, S6e and critic prompts.
A workflow run
showcase-facts
Gather the facts for a homepage showcase of project 3: the homepage's look, which process numbers code can prove, how a live piece could be embedded, and the copy rules.
Ben
What he wrote is his to release.
A workflow run
showcase-lesson-facts
Inventory the two lessons the showcase teaches: the state rules that hide problems, each sourced, and the agent orchestration as it actually ran, counted from the journals.
Ben
What he wrote is his to release.
A workflow run
showcase-sh0-sh1
Make every process number the showcase prints come from a file written by code, with a planted-error check, then build three showcase versions; the versions part was stopped and rerun in parallel.
The lead
Brief committed at 101f220, and the showcase branch is fast-forwarded to it. Before launching the showcase builders, a quick look at how the dashboard chain is doing:
A workflow run
showcase-sh1-parallel
Build three different homepage showcase sections at once, one per builder, for Ben to judge by eye.
Ben
What he wrote is his to release.
A workflow run
showcase-sh2
Build the showcase as Ben had just ruled it (the homepage slot and a case-study page of its own); stopped after three minutes when Ben proposed a different homepage treatment.
Ben
What he wrote is his to release.
A workflow run
showcase-sh2-case-study
Build the case-study page alone; stopped about a minute in when Ben asked to pause building the site until the vision was decided.
Ben
What he wrote is his to release.
A workflow run
case-study-reference-hunt
Hunt great explanatory pages in five genres as design references for the case-study page and build a board for Ben to keep or kill; the billing stop cut off four hunters and the curator.
Ben
What he wrote is his to release.
A workflow run
ground-the-next-brief
Ground the next brief read-only: mine each branch for what is worth discussing and what its email would say, check the live public page's claims against the rules and the data, and audit which sentence defects were really fixed.
Ben
What he wrote is his to release.
A workflow run
p3-fix-round-board-truth
Run the fix round the stored sentences never got and build the case-study reference board; the machine's memory crash killed the fix-round builder, and a resume briefly restarted both builders.
Ben
What he wrote is his to release.
A workflow run
p3-seven-emails
Write the seven hand-off emails from stored values only, have a skeptic try to refute every sentence, and print them by script for Ben to read.
Ben
What he wrote is his to release.
A workflow run
(no run record)
Continue the sentence fix round alone after the crash: check the dead builder's four commits, then carry on from step 5; it finished steps 5 and 6 and was partway through step 7 when the machine froze again.
Ben
What he wrote is his to release.
A workflow run
p3-fix-round-board-truth
The fix round the stored sentences never got, relaunched with dev servers forbidden after the machine froze: a builder carried on from two that had died, a critic tested the second program's checks and a fact-checker every sentence, and a fixer fixed the four blocking defects and most of the rest (the reference board and the public-page truth pass were switched off).
Ben
What he wrote is his to release.
A workflow run
p3-structure-spec-and-handoff
Spec the structure of the breakdown page and the dashboard and write the next session's handoff; stopped within a minute when Ben said the job did not need a workflow, and the lead wrote the handoff itself.
Ben
What he wrote is his to release.
A workflow run
slice3-exact-spec
To write exact specs for the breakdown page and the dashboard from the stored files: nine readers, two writers, three critics and two revisers.
Ben
What he wrote is his to release.
A workflow run
slice3-consolidate
Gather everything the two pages still owe into one register, move both specs onto the fix round's stored words and shorten them, check all three files, and fix what the checks found.
Ben
What he wrote is his to release.
A workflow run
slice4-wireframes
To build two different wireframes of each page as static pages filled with the stored words, have each checked and fixed, and set them side by side for Ben to pick.
The lead
Both summaries read right. Now my own spot-check: I'm pulling eight of the headline figures the specs cite for Tyler and reading each by path.
A workflow run
w1-surgical-world-change
To change the modeled world's settings for Fort Worth and Austin only so both read near perfect: design, measure in scratch worlds, two adversarial challenges, revise, and apply on a side branch.
The lead
The memory sampler finished on its own. Starting W1 now, as planned. First I'm reading its brief, and E1's diagnosis of what W1 should change, so the workflow is written against them and not from memory.
A workflow run
p3-clarity-handoff-cold-read
Four fresh readers, read-only, tried to start work from the clarity round's handoff and seven build orders and reported guesses, gaps and anything that contradicted Ben or the record.
Ben
What he wrote is his to release.
A workflow run
p3-apply-cold-read
One agent per build order applied the cold read's findings under the lead's rulings, checking each at its source and in both versions of the data.
Ben
What he wrote is his to release.
A workflow run
p3-round-two-pieces
The clarity round: seven builders rebuilt the inside of the wireframe blocks Ben found unclear, and one assembler joined them into the two round-two pages, with no critics; it ran 11:40 to 12:25 while Ben was away (V36).
The lead
Committed (a9fceed). Now section 5, the launch: confirming main, saving the git baseline, checking the room, and starting the memory alarm.
A workflow run
ground-final-build
Nine read-only readers mapped what the final build stands on: the two approved pages, what is still owed, the component kit, the look, the public site, the stored data and the w1 merge, the reach of Ben's reframing (V37), and the record of the agent work.
Ben
What he wrote is his to release.
A workflow run
p3-round-two-pieces
One builder drew section 1 of the breakdown page on the purpose Ben had just settled (V39).
Ben
What he wrote is his to release.
A workflow run
mcallen-light-issues
Three read-only researchers looked for light, real issues for McAllen: what its records hold but do not show, what the project's research already verified, and what published sources say about a South Texas branch.
Ben
What he wrote is his to release.
A workflow run
p3-round-two-pieces
One builder rebuilt McAllen's view around the issue Ben picked (V43): one client's change of health plan on Jan 1 behind the money that locked unpaid.
Ben
What he wrote is his to release.
A workflow run
p3-w2-mcallen-out-of-the-storm
On Ben's go (V45), one builder took McAllen out of the modeled January storm and rebuilt the data, and a checker who trusted neither the builder nor the order tested eight claims with its own runs.
Ben
What he wrote is his to release.
A workflow run
p3-final-piece-sheets
Step 0 of the final build: thirteen writers each wrote the sheet of one piece of the two approved pages (what it prints, where each figure is stored, what is owed, its faults), building nothing; launched on Ben's 'Go' to the lead's question at 18:46.
Ben
What he wrote is his to release.
A workflow run
p3-final-the-slice
Two fresh readers read the three build orders for Tyler's view, the pattern the other views copy, as their builders would, before anything was built.
Ben
What he wrote is his to release.
A workflow run
p3-final-the-slice
Stage B: one builder stored the branch views' words and chart places in a file of their own, made again by a new group of the second program.
Ben
What he wrote is his to release.
A workflow run
p3-final-the-first-night
Six fresh readers read the first night's build orders as each kind of builder would and listed what they would have to guess (111 findings, 14 blocking); the lead rewrote the orders from them.
Ben
What he wrote is his to release.
A workflow run
p3-final-the-first-night
Stage SL1b: one builder, in its own copy, grew the machinery that stores the pages' words and places so the night's builders could add their pieces, each made again by a twin in the second program.
The lead
A workflow run
p3-final-the-first-night
Stage C: one builder built Tyler's branch view for real as the pattern, then a checker who trusts nobody read every sentence, figure and mark against the stored files, the database and the rules.
The lead
A workflow run
p3-final-the-first-night
The recheck: two fresh readers read the rewritten night orders again and said what a builder would still have to guess (39 findings, 5 blocking); launched by the lead under Ben's go for the night (V49).
The lead
A workflow run
p3-final-the-first-night
Stage B2: two builders in copies of their own stored the rows every piece shares (the state's rules with their sources among them) and the sentences that read the receipts.
Ben
What he wrote is his to release.
A workflow run
p3-final-the-first-night
Stage C2: three builders in copies of their own built the pictures more than one piece draws (marks, bars, and the drawing of the work).
Ben
What he wrote is his to release.
A workflow run
p3-final-the-first-night (stage pieces)
Eleven builders, each in a copy of the project of its own, built one piece each of the two pages for real: the dashboard's first screen, the six other branch views, sections 1 to 3 of the breakdown page, and the emails' package.
The lead
Eleven copies whole at 2a9b992 in under a minute. Launching stage E.
A workflow run
p3-final-the-first-night (stage letters)
Two builders built the emails on screen: Tyler's whole with its six pictures, and the six others with their head, sentences and chart of contracts, plus the attachments that were stored (Fort Worth's and El Paso's had none yet).
The lead
Both letters' copies are whole at 2d55de5, with the package's Letter in place. Launching stage F.
A workflow run
p3-final-the-first-night (stage assemble)
One assembler put every piece together into the two whole pages, ran both programs over everything once, and shot every page state.
The lead
Everything of the night is landed on night1 (1,706 tests, 52,892 figures agreed, no copies left). Launching the assembler in the build tree, and filling the stage F rows into the log.
A workflow run
p3-final-the-first-night (stage check)
Seven checkers who trusted nobody read every figure, sentence and place of both whole pages against the stored files, the modeled agency's database and the verified rules, and said what they would not sign.
The lead
Stage G is landed on night1 at 4b9846c. Merging the docs into the build tree and writing the log rows for the assembler's return and landing, then the checks launch.
A workflow run
morning-list-grounding
Before the fix pass, six read-only readers checked the open items of the morning list, the night's findings Ben had just ruled on, at their source in the build tree, and a seventh re-checked the claims the fixes rest on and found where the fixes collide.
Ben
What he wrote is his to release.
A workflow run
p3-final-the-fix-pass (stage a, job A)
Job A of the fix pass: one builder changed the analysis program and its twin for San Antonio's "By when", the county sentence and a stored count of still-open visits.
Ben
What he wrote is his to release.
A workflow run
p3-final-the-fix-pass (stage build)
Six fix-pass builders worked side by side, each in its own copy: McAllen's rule sources, every rule's dates, the rules section, the emails, the frame and fold, and three picture pages for Ben to point at.
The lead
Six copies made, all whole, at 9f34d13. The handoff shows one uncommitted change; I expect it's just my clock fix, so I'll confirm and commit it, then launch the six builders.
A workflow run
p3-final-the-fix-pass (stage gates, job H)
Job H: one builder made the page scripts check that the second program had finished on the very views file they read, changed 'Ben's own series' to 'my own series', planted errors in the pages' own words, and built a guard against real company names in the stored files.
The lead
A workflow run
p3-final-the-fix-pass (stage skeptics)
Three skeptics, each in a throwaway copy, tried to make job H's new checks pass on bad input: the build gate, the planted errors and the real-names guard.
Ben
What he wrote is his to release.
A workflow run
p3-final-the-fix-pass (stage read)
One reader who built nothing checked every fix on the rebuilt pages, comparing the text of all 37 page states before and after.
Ben
What he wrote is his to release.
A workflow run
p3-final-the-fix-pass (stage guards, job K)
Job K: one builder wrote the tests the skeptics showed were missing and proved each one can fail.
Ben
What he wrote is his to release.
What this does not claim
Scroll both ways, or pinch to zoom.
The template
{0} client landlines were removed after the landline review and no alternative device has been ordered, so the {1} attendants still working those homes have no way to clock in.
The blanks
Stands only when
The sentence
20 client landlines were removed after the landline review and no alternative device has been ordered, so the 19 attendants still working those homes have no way to clock in.
querylandline-cohorte-storiesbranches[id=TYL].story.landlines.homesbranches[id=TYL].story.landlines.attendants_still_on_the_round_with_no_methodbranches[id=TYL].cause.key is clocking_inbranches[id=TYL].story.landlines.homes above 1branches[id=TYL].story.landlines.devices.ordered is 0branches[id=TYL].story.landlines.attendants_still_on_the_round_with_no_method above 1The first program
82 cross-checks inside the analysis pass
The second program
reads the raw tables: 250,088 visit rows, 254,439 transaction rows
52,904 figures compared
0 disagree
428 checks passed, 0 failed
node scripts/home-care-ops/analyze.ts --out <rebuild>py -3 -B scripts/home-care-ops/reconcile.pydata/home-care-ops/reconcile.json2026-10-01T07:06:35ZThe second program reads the same raw tables and shares no query or code with the first, and the first stores where each figure comes from.
The page does not build unless the check finished on the exact file it reads with nothing disagreeing.
By group, each figure compared by the second program
Scroll both ways, or pinch to zoom.
a. the score and the pace
one accepted visit
changed from 22,265 to 22,266
Caught by usage score counts, by branch and quarter
caught
b. delivered against authorized hours
one delivered hour
changed from 68,757.5 to 68,758.5
Caught by authorized and delivered hours, by branch and quarter
caught
c. the money
one cent
changed from 17,820,505 to 17,820,506
Caught by the agency's dollars on the clock
caught
d. the workforce and the three levers
one visit in a lever's count
changed from 8,829 to 8,830
Caught by the three levers for the quarter, from counts
caught
e. the stories, the grids, the frequencies
one visit in a grid
changed from 8,690 to 8,691
Caught by what happened to every visit delivered this quarter, by branch and pooled
caught
f. the calendar, the events, the zone
one client on a marked day
changed from 10 to 11
Caught by days a landline review removed phones, how many, and the Monday each is marked on
caught
g. the map of Texas
one point across the map
changed from 85.83 to 86.83
Caught by where each branch's marker sits on the map of Texas
caught
h. the pages' words and places
a place moved by one point
changed from 20 to 21
Caught by where every mark sits in each branch's pictures, made again from the stored figures: the contracts' chart, what the fix is worth, the week by week line and the three tracks
caught
h. the pages' words and places
one word changed, "care" to "work"
changed from Tyler is doing the care, and two of its contracts are heading to fail. to Tyler is doing the work, and two of its contracts are heading to fail.
Caught by the words of each branch's view, made again from their templates, their conditions and the stored figures
caught
h. the pages' words and places
one word changed, "pass" to "fail"
changed from A contract needs 80 points out of 100 to pass. to A contract needs 80 points out of 100 to fail.
Caught by the words of the dashboard's first screen, made again from their templates, their conditions and the stored figures
caught
h. the pages' words and places
one word changed, "under" to "over"
changed from STAR+PLUS Plan A is heading to 74 and STAR+PLUS Plan B to 74, under the line of 80, and HHSC fee for service is heading to 82. to STAR+PLUS Plan A is heading to 74 and STAR+PLUS Plan B to 74, over the line of 80, and HHSC fee for service is heading to 82.
Caught by the words of the seven letters, made again from their templates, their conditions and the stored figures
caught
h. the pages' words and places
one word changed, "verified" to "guessed"
changed from This is how I sourced, pooled and verified my information about this world, the industry and its data. to This is how I sourced, pooled and guessed my information about this world, the industry and its data.
Caught by the words of link 3 (the data), made again from their templates, their conditions and the stored figures
caught
Modeled data. No real people, clients, or companies.
7 branches, 1,500 clients, 250,088 visits over 226 days.
The dataset rebuilds from its seed byte for byte.
The fixed starting phrase: pecan-hollow-sfy26
version 2.0.0; 1,353 attendants; 52 counties; 7 service areas; 24 contracts
Both programs read only what a real agency's systems would hold, never the hidden settings the agency was modeled from.
Neither program reads a sim_ column.
103 checks on the dataset pass
887 research claims tested against their sources
Modeling choices, as stored
Where the map comes from
These counties are derived by a published rule from published service areas and are attributed to no real agency. We found no per provider county list published in Texas.
| What it says | Command | Exit | Ran at | Head at run |
|---|---|---|---|---|
| rebuilds from its seed byte for byte | node scripts/home-care-ops/generate.ts --out <rebuild> && node scripts/home-care-ops/analyze.ts --out <rebuild> && node scripts/home-care-ops/views.ts --out <rebuild> | 0 | 2026-10-01T07:05:57Z | 7f569c9d864f18d5bc21249df9a12fdcac9a1f18 |
| 103 checks on the dataset pass | node scripts/home-care-ops/validate.ts --out <rebuild> | 0 | 2026-10-01T07:06:29Z | 7f569c9d864f18d5bc21249df9a12fdcac9a1f18 |
| 82 cross-checks inside the analysis pass | node scripts/home-care-ops/analyze.ts --out <rebuild> | 0 | 2026-10-01T07:06:04Z | 7f569c9d864f18d5bc21249df9a12fdcac9a1f18 |
| read figures and disagreements from data/home-care-ops/reconcile.json | py -3 -B scripts/home-care-ops/reconcile.py | 0 | 2026-10-01T07:06:35Z | 7f569c9d864f18d5bc21249df9a12fdcac9a1f18 |
| 12 planted, 12 caught | node scripts/home-care-ops/planted-errors.ts | 0 | 2026-10-01T07:07:03Z | 7f569c9d864f18d5bc21249df9a12fdcac9a1f18 |
| 101 workflow runs | node scripts/home-care-ops/receipts.ts --count | 0 | 2026-10-01T07:07:31Z | 7f569c9d864f18d5bc21249df9a12fdcac9a1f18 |
| 692 agent launches | node scripts/home-care-ops/receipts.ts --count | 0 | 2026-10-01T07:07:31Z | 7f569c9d864f18d5bc21249df9a12fdcac9a1f18 |
| 887 research claims tested against their sources | node scripts/home-care-ops/receipts.ts --count | 0 | 2026-10-01T07:07:31Z | 7f569c9d864f18d5bc21249df9a12fdcac9a1f18 |
| 746 held | node scripts/home-care-ops/receipts.ts --count | 0 | 2026-10-01T07:07:31Z | 7f569c9d864f18d5bc21249df9a12fdcac9a1f18 |
| 88 left as "cannot check" rather than guessed | node scripts/home-care-ops/receipts.ts --count | 0 | 2026-10-01T07:07:31Z | 7f569c9d864f18d5bc21249df9a12fdcac9a1f18 |
| 53 refuted | node scripts/home-care-ops/receipts.ts --count | 0 | 2026-10-01T07:07:31Z | 7f569c9d864f18d5bc21249df9a12fdcac9a1f18 |
| 96 numbered rulings in my own series (V, D, R and W), counted through V67 | node scripts/home-care-ops/receipts.ts --count | 0 | 2026-10-01T07:07:31Z | 7f569c9d864f18d5bc21249df9a12fdcac9a1f18 |
| 149 numbered rulings in the lead agent's series (K and L), counted through L111 | node scripts/home-care-ops/receipts.ts --count | 0 | 2026-10-01T07:07:31Z | 7f569c9d864f18d5bc21249df9a12fdcac9a1f18 |
How the errors were planted: planted-errors.ts: 8 stored figures and 4 stored sentences, each changed by one unit (the unit is stored with it), at least one in each of the twin's 8 groups (5 in "h. the pages' words and places"), in a scratch copy
Every sentence you meet next, on the dashboard and in the emails, is built this way.
Link 4 · The diagnosis
This is the regional director's dashboard, live, opened on Tyler. The state grades each contract, never a branch.
Scroll both ways, or pinch to zoom.
Tyler
The visits happen. Too many are typed in by hand instead of clocked in.
Where each contract is heading
| Dec to Feb | Mar to May | so far | where it finishes | |
|---|---|---|---|---|
| STAR+PLUS Plan A | 94 | 86 | 78 | 74Will not pass |
| STAR+PLUS Plan B | 92 | 87 | 77 | 74Will not pass |
| HHSC fee for service | 92 | 91 | 85 | 82Passes |
under the line
The one thing to do
What it buys
STAR+PLUS Plan A
STAR+PLUS Plan B
Why it is happening
Against the other six branches
Clocking in
40
100
7th of 7
Accepted the first time they are sent
80
100
5th of 7
Planned hours delivered
50
100
1st of 7
The stored sentences, word for word
Every figure, contract by contract
| Contract | Dec to Feb | Mar to May | so far | Heading to range | Status | Last day the fix can be in place, in our model | |
|---|---|---|---|---|---|---|---|
| STAR+PLUS Plan A | 94 | 86 | 78 | ↓ | 74 73 to 74 | Will not pass | Aug 4 |
| STAR+PLUS Plan B | 92 | 87 | 77 | ↓ | 74 73 to 76 | Will not pass | Aug 3 |
| HHSC fee for service | 92 | 91 | 85 | ↓ | 82 81 to 83 | Passes | not needed |
Where the figures come from
Every branch has a letter like Tyler's. All seven open at the end.
Link 5 · The effect
The hand-off
One email to each branch manager, from the regional director. Pick a branch to read its email, then open a sentence to see the figures behind it and how they were checked.
Scroll both ways, or pinch to zoom.
Every contract has dropped every quarter. Two finish under the line.
Counted from the raw records and made again by a second program: 52,904 figures compared, 0 disagree.
Scroll both ways, or pinch to zoom.
20 of 110 clients lost the home phone their attendants clocked in on.
20 homes lost the home phone
the branch's other clients
Most of the visits typed in by hand came from those homes.
84 in every hundred came from those homes
from other homes
788 of the 938 visits typed in by hand this quarter
Counted from the raw records and made again by a second program: 52,904 figures compared, 0 disagree.
Scroll both ways, or pinch to zoom.
The phones came out in April and May, and nobody working those homes has a way to clock in.
phones taken out
86 days since the first phone came out, 0 devices ordered
Apr 20first phone taken out
May 18last phone taken out
Jul 15today
One figure per attendant on those homes
19 have no way to clock in
left the agency
Counted from the raw records and made again by a second program: 52,904 figures compared, 0 disagree.
Scroll both ways, or pinch to zoom.
Our model's last day to order is Jul 27, 12 days from today, and nothing has been ordered yet.
time left to order, 12 days
5 business days to ship
Jul 15today
Jul 22in use if ordered today
Jul 27last day to order
Aug 3 last safe day, STAR+PLUS Plan BAug 4 last safe day, STAR+PLUS Plan A
Aug 31quarter ends
One device per home that qualifies: 20 homes qualify, 0 ordered
not ordered
Counted from the raw records and made again by a second program: 52,904 figures compared, 0 disagree.
Scroll both ways, or pinch to zoom.
The sooner the fix is in place, the higher each contract finishes. After the last safe day, it fails.
Counted from the raw records and made again by a second program: 52,904 figures compared, 0 disagree.
Scroll both ways, or pinch to zoom.
The sooner the fix is in place, the higher each contract finishes. After the last safe day, it fails.
Neither contract has finished a quarter under the line. Left this late, this quarter would be the first.
Counted from the raw records and made again by a second program: 52,904 figures compared, 0 disagree.
| Attendant | Left | Mobile setup | Homes with no usable landline | Visits with no way to clock in | Latest such visit |
|---|---|---|---|---|---|
| Eleanor Rios | No | 1 | 88 | Jul 14 | |
| Dalia Sanders | No | 1 | 88 | Jul 14 | |
| Latoya Walker | No | 1 | 88 | Jul 14 | |
| Eleanor Porter | No | 1 | 88 | Jul 14 | |
| Pamela Adams | No | 1 | 88 | Jul 14 | |
| Rachel Solis | No | 1 | 88 | Jul 14 | |
| Danielle Young | No | 1 | 88 | Jul 14 | |
| Garrett Harris | No | 1 | 38 | Jul 14 | |
| Keisha Fairchild | No | 1 | 38 | Jul 14 | |
| Erica Washington | No | 1 | 38 | Jul 14 | |
| Whitney Porter | No | 1 | 37 | Jul 13 | |
| Carla Henderson | No | 1 | 37 | Jul 13 | |
| Tiffany Reed | No | 1 | 32 | Jul 14 | |
| Deborah Ibarra | No | 1 | 32 | Jul 14 | |
| Kimberly Stovall | No | 1 | 31 | Jul 14 | |
| Patrice Kirkland | No | 1 | 31 | Jul 14 | |
| Julieta Yancey | No | 1 | 31 | Jul 13 | |
| Imani Dunlap | No | 1 | 28 | Jul 14 | |
| Darnell Whitfield | No | 1 | 28 | Jul 9 | |
| Natalie Montoya | Jul 11 | No | 1 | 12 | Jun 14 |
Scroll both ways, or pinch to zoom.
STAR+PLUS Plan B is heading to 66, under the line. The rest are heading to pass.
Counted from the raw records and made again by a second program: 52,904 figures compared, 0 disagree.
In our model, STAR+PLUS Plan B cannot pass this quarter now. Its earlier quarters are over the line so far, and this one is open to a payer's review from Dec 4.
26 clients have STAR+PLUS Plan A on the state's file and STAR+PLUS Plan B on our EVV setup, and every one of their 5,693 sends since their plan changed has been rejected.
Have Noemi Hinojosa change the plan on the 26 clients' EVV setup from STAR+PLUS Plan B to STAR+PLUS Plan A today, and once it is changed, have the EVV coordinators correct their 1,161 open visits, oldest first.
The first of those visits locks today, Jul 15, and 185 of their visits have already locked unpaid.
Changed today, STAR+PLUS Plan B still finishes under the line in our model, at 75 instead of 66, and about $69,692 of care on those 1,161 visits can be billed as each is corrected and accepted before it locks.
| Month | Still open | Already locked |
|---|---|---|
| Apr | 0 | 47 |
| May | 0 | 56 |
| Jun | 0 | 56 |
| Jul | 32 | 26 |
| Aug | 330 | 0 |
| Sep | 494 | 0 |
| Oct | 305 | 0 |
| Plan changed on | Clients |
|---|---|
| Jan 1 | 2 |
| May 1 | 12 |
| Jun 1 | 8 |
| Jul 1 | 4 |
STAR+PLUS Plan A finished Dec to Feb at 77, and the training its notice requires is due Jul 20.
STAR+PLUS Plan B and STAR+PLUS Plan E also finished that quarter under 80, at 76 and 77, and neither has a notice on file.
That quarter 83 managed care authorizations were keyed into the EVV system after their start date, 39 days late on average. HHSC fee for service, whose authorizations the EVV system retrieves from the state, passed the same quarter at 86.
Scroll both ways, or pinch to zoom.
Every contract is heading to pass this quarter.
Counted from the raw records and made again by a second program: 52,904 figures compared, 0 disagree.
Run the training STAR+PLUS Plan A's notice requires by Jul 20, with the 3 EVV coordinators. If it is missed, STAR+PLUS Plan A may temporarily withhold Medicaid claims payments until requirements are met.
Name a second person who can key managed care authorizations before Sep 1, when the quarter that decides the notice starts by our reading. If a second notice follows, it asks for a corrective action plan within 10 business days.
Of every hundred hours El Paso's clients were authorized for, 8 had nobody sent to work them in Dec to Feb, 10 in Mar to May and 13 so far this quarter.
El Paso had nobody to send for 1 in 8 of the hours its office planned this quarter, against fewer than one in 200 at the other six branches together.
Scroll both ways, or pinch to zoom.
Every contract is heading to pass this quarter.
Counted from the raw records and made again by a second program: 52,904 figures compared, 0 disagree.
El Paso loses 69 of every hundred attendants a year, the highest turnover of the seven, and at 12 of 40 handovers after an attendant left, the home had a visit nobody worked before another attendant arrived.
Gloria Salinas, please check each of the 14 clients who had a visit with nobody assigned in the days up to Jul 14, assign an attendant today to every one who still has nobody, and fill the 2 handovers still open.
An hour nobody works can never be made up or billed, and this quarter the hours nobody was sent to work at El Paso would have billed about $30,989 by our estimate.
Scroll both ways, or pinch to zoom.
Every contract is heading to pass this quarter.
Counted from the raw records and made again by a second program: 52,904 figures compared, 0 disagree.
This quarter the state accepted 97 in 100 of Austin's visits the first time they were sent, and on Jul 14 every one of its 187 homes in service had an attendant.
Payers have sent 116 recoupment notices on Austin's paid claims for care given Dec to Feb, the most of the seven, and the next highest branch drew 26.
111 of them are for missing records, out of 132 notices for missing records on last winter's claims across the agency, and a payment whose records cannot be produced can be taken back in full.
Carla Salinas, please pull the 111 missing-records notices and find the records behind each paid claim, starting with the ones under appeal, oldest notice first.
The 116 notices cover about $7,082 of paid claims, and 43 of them are still under appeal.
| Reason | Notices | Paid claims they cover |
|---|---|---|
| a payer's instruction not followed | 3 | $208 |
| records missing | 111 | $6,741 |
| visit changed after the match | 2 | $133 |
Scroll both ways, or pinch to zoom.
Every contract is heading to pass this quarter.
Counted from the raw records and made again by a second program: 52,904 figures compared, 0 disagree.
HHSC fee for service, the lowest, is heading to 91, and the rest of the quarter would have to fall to 69 before it reached the line.
Fort Worth's attendants with under 6 months of service typed in 14, 17 and 22 in every hundred of their accepted visits over the last three quarters, this one so far. Across the agency the same group went 17, 15 and 14, and Fort Worth is now the highest of the seven branches.
Since Dec 1, 71 in every hundred of the hand-typed visits the state accepted from Fort Worth were missing only the clock out, against 33 in every hundred at the middle branch of the seven.
Every morning from today, have the 2 EVV coordinators call each attendant whose visit the day before has a clock in and no clock out, ahead of the next shift, starting with those who have under 6 months of service.
On our own figure for Fort Worth, visits with one punch typed in have cost 5 of the 60 clocking points so far this quarter. Fort Worth hired 18 attendants since Jun 1, the most of any branch.
McAllen has already lost about $4,182 by our estimate on 88 delivered visits that locked unpaid, and 75 of them are STAR+PLUS Plan A visits from Dec to Feb.
That quarter one client's visits were rejected 290 times, and so far this quarter the most for any one client is 10.
Scroll both ways, or pinch to zoom.
Every contract is heading to pass this quarter.
Counted from the raw records and made again by a second program: 52,904 figures compared, 0 disagree.
Hannah Zamora and Ursula Ibarra, please correct the 10 rejected visits still open and fix the cause before saving.
They are worth about $523, they have been sent 22 times with 22 rejections counted against the grade, and the first of them locks Sep 6.
Please also type in the 43 visits nobody has keyed yet, oldest first. The first of McAllen's open visits locks Jul 31.
| Code | Rejections | Visits | Not cleared |
|---|---|---|---|
| Ex0002C | 36 | 35 | 5 |
| Ex0003C1 | 26 | 26 | 1 |
| Ex0007C1 | 25 | 23 | 3 |
| Ex00057C1 | 23 | 23 | 1 |
| Ex00057C2 | 21 | 21 | 4 |
| Ex0002C1 | 21 | 21 | 1 |
| Ex00059C | 16 | 16 | 1 |
| Ex00049C2 | 15 | 15 | 1 |
| Ex00034C1 | 12 | 12 | 0 |
| Ex00065C | 11 | 11 | 1 |
| Ex0003C3 | 10 | 10 | 2 |