GBMT-7 (Elder care) — Verification Log

Protocol: method/verification-protocol.md, Phase 1 (fact-check). Filing under check: Whitepaper No. 7, filed 2026-08-03 — site/elder-care/index.html, site/elder-care/sources/index.html, and the research record in elder-care/docs/, elder-care/research/ and elder-care/baseline/. Check date: 2026-08-10. Independence (V1): this session did not author GBMT-7, did not run its red team, and did not run the 2026-08-10 blind re-score. The filing's own red team and re-scores carry no evidentiary weight here; where they are cited it is as claims under check. Source atlas: read from method/sources/ before any fetch. Two files there — plaw-119-21-obbba-lihtc-and-staffing-moratorium.md and cms-nursing-home-staffing-rule-and-repeal.md — were fetched for this filing by the 2026-08-10 atlas pass and are used below.

Coverage (V3)

Coverage proof: 689 extracted · 689 verdicts · 0 unaddressed.

Four rows additionally carry a dual as-of verdict under V6 (P057, P072, S098, R105 — the 36.5% foreign-born figure): correct as of the filing date, stale as of the check date.

Two honest caveats on that arithmetic, stated here rather than at the foot. First, 312 of the 689 verdicts are CONFIRMED-INTERNAL — the scorecard arithmetic, the process self-descriptions, the withdrawals. Those are real verdicts and several of them are load-bearing (the 24 ranking sums were recomputed from the matrix and reproduce exactly), but they test the record against itself, not against the world. Second, 90 rows are UNVERIFIABLE, concentrated in the research record. An unverifiable verdict is a verdict, not a verification.

Extraction guardrail (V3)

Nothing was delegated: extraction, retrieval and adjudication all ran in this session on the adjudicating model, so the guardrail's named failure mode — a cheap extractor capturing every number and dropping every inference — could not occur by construction, and there is no delegated ledger to diff against.

The guardrail was still exercised, in the form the childcare pass showed it matters. Part 5 ("Institutional quality") was re-extracted a second time from the raw HTML, before consulting the first extraction, with an explicit instruction to capture causal claims, statutory characterisations and inferences rather than only figures and dates. Second extraction: 20 substantive assertions. First extraction (P106–P125): 20. The diff is zero. Both passes independently caught the four assertions in that section that are pure inference or characterisation rather than fact — "not merely under litigation risk," "statutory survival isn't enforcement," "the most rigorous study run on the question," and "No published rebuttal exists" — and three of those four ended up carrying findings. The extraction layer is sound; the ledger's count can be trusted.

The three findings that matter most

1. The filing's best evidence carries four wrong numbers

The whitepaper calls the private-equity mortality result "our best evidence" and gives it its own line in the honesty box. Anchor row 10 on the public sources page states six figures from it. The paper was fetched this pass — NBER WP 28474, revised August 2023, which is the version published as Review of Financial Studies 37(4) 2024 — and four of the six do not survive.

The headline holds. The supporting apparatus does not. The 20,150 figure is the superseded February 2021 version of the same paper, so the record cited one vintage's excess-death count alongside another vintage's venue. The antipsychotic figure is not in the paper at any vintage; it appears to have migrated in from the adjacent PE/nursing-home literature. And the record's own framing of the OLS/IV gap ("notably larger than the naive OLS estimate (+10%)") describes a gap of one percentage point where the paper describes a gap of nine.

Two scope corrections travel with these. The 11% is a local average treatment effect for compliers — patients who go to a PE facility because it is closest — which the authors then generalise via a marginal-treatment-effects analysis that "reveals substantial heterogeneity in treatment effects, including small beneficial effects for some patients," concluding that "PE has nuanced effects, with adverse outcomes for a subset of patients." The honesty box's "an 11% mortality increase … across seven million patients" is therefore wrong twice over: wrong sample, and an average asserted where the paper reports a subset effect. Separately, the whitepaper's "No published rebuttal exists" is a V5 negative that upgrades the record's own honest "No direct econometric rebuttal was located," and does so while the record itself holds a JAMA COVID-era comparison that the red team used to downgrade confidence.

2. "Ten-year moratorium" is nine years and three months, and the record's open dispute is closed

Pub. L. 119-21 §71111 (fetched to method/sources/):

The Secretary of Health and Human Services shall not, during the period beginning on the date of the enactment of this section and ending September 30, 2034, implement, administer, or enforce the amendments made by the provisions of the final rule …

Enactment was 2025-07-04. The moratorium therefore runs 9 years 3 months, not ten. CMS's own repeal rule (90 FR 55687) says "until September 30, 2034" and "before October 1, 2034" throughout.

Deviations log #2 records this as unreconciled — "CMS's own repeal cites the moratorium as running to 2035, one law-firm summary states 2034" — and anchor row 7 publishes the dispute. The primary text closes it: 2034, in both the statute and the agency rule. The figure the record attributes to CMS does not appear in CMS's rule.

Everything else in this cluster verified at primary tier and holds: 3.48 total / 0.55 RN / 2.45 NA HPRD, the 24/7 RN requirement, the phased 2026–2029 implementation, rescission effective 2026-02-02, vacatur in American Health Care Association v. Kennedy (N.D. Tex.) and Kansas v. Kennedy (N.D. Iowa), and the survival of the facility-assessment and Medicaid staffing-spend provisions (the repeal touches only §§483.5 and 483.35).

Two record-level claims in the same cluster did not survive. Phase 0 states the rule was vacated "on statutory-authority grounds" in both courts; CMS records a major-questions holding for N.D. Tex. only and states no ground for N.D. Iowa. Phase 0 also states "HHS withdrew its own appeals of both losses in fall 2025"; the repeal rule records the Texas case as "appealed June 2, 2025 to the Fifth Circuit" and says nothing about withdrawal. Neither opinion has been read by this project or by the atlas (UNOBTAINED.md §3), so the reasoning attributed to them rests on the characterisation of the agency that repealed the rule.

And the pre-2024 baseline is not what the record says it is. ws06-09-findings.md describes "the old, pre-2024 federal floor (0.3 HPRD)." 42 CFR 483.35, fetched from eCFR in its post-repeal form, contains no numeric HPRD floor — it requires "sufficient nursing staff," licensed nurses on a 24-hour basis, and an RN for at least 8 consecutive hours a day, 7 days a week. "38 states plus DC already exceeded the old federal floor" is therefore true in the trivial sense that any numeric standard exceeds a non-numeric one, and the whitepaper prints it in a numbers block immediately beside "3.48 HPRD in the dead federal rule," where a reader will take it to mean something much stronger. The record's own counterweight — "29 states require under 3.5 HPRD, 15 under 2.5 … state law is a genuinely uneven patchwork, not a uniform substitute" — appears on neither public page.

3. The workforce evidence is mostly about people under 65

This is the finding that reaches furthest, and it is not a wrong number — it is a scope the filing never states on the page a reader actually reads.

The whitepaper's central hypothesis is tested on Medicaid HCBS: ARPA §9817 spending, HCBS waiting lists, HCBS wage movement. Its own primary source says who those people are. KFF's 10 Things About LTSS — cited by name in Part 1 — states: "Among Medicaid enrollees who use HCBS, 63% are under age 65," and "Over half (57%) of Medicaid enrollees who use LTSS are under 65." KFF's waiting-list brief — cited by name in Part 4 — states: "Most people on waiting lists or interest lists have intellectual or developmental disabilities," comprising 89% of lists in non-screening states and 49% elsewhere.

So the 710,000 figure that Part 2 uses as the workforce hypothesis's main outcome variable, and that Part 4 uses as the institutional-bias story's headline, is majority-I/DD and majority-under-65. The sources page discloses the wider scope in one clause of Record 0 ("older Americans and adults with disabilities"). The whitepaper's subtitle says "a feasibility assessment of long-term services and supports for older Americans," and nothing in nine parts or the honesty box tells the reader that the evidence base is mostly not that.

Three related overstatements sit on top of it:

• "71% of it went to wages." MACPAC's Figure 1, fetched this pass, attributes $26.3B to "workforce recruitment and retention" — provider rate increases, wage add-ons, hiring and retention bonuses — with workforce training a separate $3.9B. The record gets this right (`ws03`: "71% ($26.3B) to workforce recruitment/retention"); the whitepaper narrowed it.
• "$37.1B ARPA HCBS funding." MACPAC: states "planned to spend an estimated $37.1 billion in state and federal funds." Both qualifiers are dropped on the public page, and the abstract further calls it "earmarked for direct-care wages."
• "the national HCBS waiting list grew from ~692,679 (2021) to over 710,000 (2024): the opposite of what capacity-targeted funding should produce." KFF's own brief says "In most years since 2016, there have been roughly 0.7 million people on waiting lists or interest lists," warns that the counts "are not necessarily comparable across states or over time," and records a 2023 survey-definition change that widened what states report. A 2.5% move across a definitional break in a flat series is not evidence about a funding intervention — and ws03 records, but neither public page mentions, that only 5% ($1.7B) of the ARPA money targeted waiting-list reduction at all. That single unpublished fact is the strongest available objection to the inference, and it is in the filing's own research record.

The verdict on the hypothesis itself — INDETERMINATE — is not disturbed by any of this. If anything the primary sources support it more firmly than the filing does: MACPAC states outright that "Section 9817 did not include a requirement to evaluate," that states "encountered difficulty isolating the impact of each initiative," and that the national evaluation's report "is anticipated in 2026." What is disturbed is the framing of the evidence that produced it.

Cluster A — Baseline and spending (P038–P052, S085–S091)

Reached at primary tier. ASPE Research Brief, Long-Term Services and Supports for Older Americans: Risks and Financing, 2022 (August 2022, REVISED): "we estimate that over half (56%) of Americans turning 65 today will develop a disability serious enough to require LTSS." CONFIRMED, including the definition the record states (2+ ADLs expected to last 90+ days, or severe cognitive impairment — the HIPAA trigger). KFF's 10 Things About LTSS confirms $415B (2022), Medicaid 61%, out-of-pocket 17%, HCBS $284B vs institutional $131B — all four verbatim.

CORRECTED — the 2013 crossover is a Medicaid fact, published as an all-payer one. Part 1 reads: "Medicaid … covers 61% of the $415B this country spends annually on long-term care, with home- and community-based spending having overtaken institutional spending nationally in 2013." KFF's brief supports the 415B/61284B/$131B figures and separately reports that Medicaid LTSS spending on HCBS rose "from only 10% in 1988 to 62% in 2020." Nothing in it dates an all-payer crossover to 2013. The 2013 crossover is the well-documented Medicaid-LTSS event; attaching it to the all-payer sentence changes what it says. ws02 and anchor row 9 carry the same conflation.

OVERSTATED — "not the 70% figure still circulating from a superseded estimate." Both figures are ASPE's. The 2019 brief (fetched) is HRS-based and reports "70 percent of adults who survive to age 65 develop severe LTSS needs before they die"; the 2022 brief is DYNASIM4-based. The 2022 brief does not present itself as superseding the 2019 one — it states that its results "differ from those reported in Favreault and Dey (2016)" and cites Johnson (2019) as a live reference. Different model and different construct is not the same as superseded, and "citation trap" (anchor row 1) overstates it further.

OVERSTATED — the §2 confidence tier was not updated after its own red team. ws02's confidence line still justifies medium-high because "multiple KFF compilations converge." Red-team item #6 established that these share upstream NHEA accounting assumptions and are restatements of one convention. Deviations #23 records the line-edit pass as complete; this line, and ws04's "Reinsurance was actually tried by a state, and it failed" headline (softened by red-team #8 to "suggestive negative evidence from one unread secondary-sourced state study"), were both missed.

**UNVERIFIABLE — 209M. * *ACL^′sownNFCSPpagepublishesafundingtablecoveringFY2013–FY2016(145,586,000 rising to $150,586,000) and nothing later, despite a 2026 modification date. No appropriations document was reached this pass. The whitepaper additionally labels the figure "OAA caregiver support, total," which is a scope claim about Title III-E alone — Title VI-C's Native American caregiver support and Lifespan Respite are separate OAA lines.

Cluster B — Workforce (P053–P075, S092–S099)

Covered above. Reached at primary tier: MACPAC's January 2026 issue brief confirms $37.1B, the $26.3B workforce line, the $1.7B waiting-list line, the absence of an evaluation requirement, the isolation difficulty, and the pending GSA/AIR evaluation. KFF independently corroborates "more than $26 billion of the planned $37 billion."

Dual as-of verdict (V6) — the foreign-born figure. 36.5% of home health aides foreign-born is 2019 ACS data (MPI/American Immigration Council). "Over twice the rate for the workforce overall" holds against the 2019 base the record uses (17%). Against the current base it does not: the foreign-born share of the US labor force was 19.2% in 2024, which makes the ratio 1.9×. Correct as of the filing date on the record's own base; stale as of the check date as a present-tense claim. This is a wording fix, not a retraction.

Cluster C — Financing and insurance (P076–P093, S100–S104)

Reached at statutory tier. Pub. L. 112-240 §642 (fetched from govinfo): "Title XXXII of the Public Health Service Act (42 U.S.C. 300ll et seq.; relating to the CLASS program) is repealed," with ACA Title VIII repealed as a conforming change. Signed 2013-01-02; CLASS enacted 2010-03-23. §643 created the Commission on Long-Term Care, confirming R072. The whitepaper's CLASS sentence is CONFIRMED.

Reached at statutory tier — WA Cares. RCW 50B.04.060(1): "Beginning July 1, 2026, approved services must be available and benefits payable." RCW 50B.04.010(3): a benefit unit is "up to $100 … adjusted annually for inflation." RCW 50B.04.060 caps lifetime receipt at "365 benefit units." The WA Cares Fund's own site states the benefit as "$36,500 (grows over time with inflation)" and the contribution as 0.58%. Rate and cap CONFIRMED.

OVERSTATED — "unchanged since enactment" (anchor row 6). RCW 50B.04.080(1): "The initial premium rate is .58 percent … Beginning January 1, 2026, and biennially thereafter, the premium rate shall be set by the pension funding council at a rate no greater than .58 percent," at "the lowest amount necessary to maintain the actuarial solvency" of the fund. The rate may in fact still be 0.58%, but it is no longer a fixed statutory rate, and the anchor's "unchanged since enactment" describes a statute that no longer exists in that form.

OVERSTATED — "WA Cares' first-ever payment went out July 1, 2026." The statute makes benefits payable from that date. The record's own anchor says "benefits first became payable July 1, 2026." No payment event was verified by the record or by this pass. The whitepaper also runs the two framings against each other on the same page: the Part 3 heading says "the new one hasn't paid a claim yet" while the paragraph beneath it says a payment went out.

UNVERIFIABLE — the Milliman study. ws04 discloses that "the original study has not yet been read in this record." It was not reached this pass either. The "little potential to generate savings" quotation, which carries the private-LTC row's only evidenced scorecard cell, remains secondary-sourced.

Cluster D — Waiting lists and institutional bias (P094–P105, S105–S110)

Reached at statutory tier, and it holds — this is the filing's soundest load-bearing claim. 42 U.S.C. §1396a(a)(10)(A) requires state plans to make medical assistance available including "at least the care and services listed in paragraphs (1) through (5)" of §1396d(a); §1396d(a)(4)(A) is "nursing facility services … for individuals 21 years of age or older." §1396n(c)(1): "The Secretary may by waiver provide that a State plan … may include as 'medical assistance' … payment for … home or community-based services," and §1396n(c)(9) contemplates "any waiver under this subsection which contains a limit on the number of individuals who shall receive home or community-based services." Mandatory entitlement, optional capped waiver, in the statute. KFF puts 40 of 51 jurisdictions with lists in 2024, which carries "most states."

CORRECTED — "eighteen months." ws05 dates Indiana's waitlist growth July 2024 → February 2026. That is nineteen months.

CORRECTED — Illinois "has no waitlist at all." Phase 0 §5 disqualifies Illinois as a case study on that ground. KFF's current brief reports Illinois waiting lists growing from 14,444 to 15,905 between 2023 and 2024, and lists Illinois among the states that do not establish eligibility until selection. The case-selection reasoning rests on a false premise; the Indiana finding itself is unaffected.

OVERSTATED — Texas is one of "six" non-screening states. KFF's brief: "Among the eight states that do not screen people for eligibility on any lists, six have only waiting lists, one (Texas) has only interest lists, and one (Washington) uses both."

OVERSTATED — the §5 digest headline. "Indiana confirms the 'architecture, not money' hypothesis" outruns the findings file's own corrected header, "CONSISTENT WITH H5.1 (n=1, correlational, confounded)."

Cluster E — Institutional quality (P106–P125, S111–S116)

Covered in finding 2 above. One addition: the Health Affairs 22-state panel is cited on both public pages and in the scorecard as "2025." It was published in the March 3, 2026 issue (State Nursing Home Minimum Staffing Mandates: Increased Staff Levels, Minimal Impact On Finances And Closures, 2010–23); the 2025 in its DOI is a submission identifier. Its findings verify at author-institution tier: ~5% direct-care staffing increase (0.18 HPRD, driven by LPNs +0.06 and CNAs +0.13), +273Klaborcostagainst+546K net patient revenue, margins unchanged, no closure effect — and, worth recording because it bounds the result, eleven of the twenty-two states had no staffing mandate at all across the window.

Also unverifiable and worth naming: "modeled to save 13,000 lives/year" (Phase 0 §8) appears nowhere in the fetched text of 89 FR 40876 and carries no citation in the record.

Cluster F — Assisted living (P126–P134, S117–S123)

CORRECTED — the second GAO report does not say what the filing says it says. GAO-18-179 (published 2018-01-05) verifies verbatim: "Twenty-six state Medicaid agencies could not report to GAO the number of critical incidents that occurred in assisted living facilities," and "Thirty-four states made critical incident information available to the public." GAO-26-107884 exists and verifies verbatim on spending: "federal Medicaid and Medicare spending for services provided in assisted living facilities totaled at least $12 billion in 2024. This amount is likely an undercount because of data limitations" ($3.5B Medicaid + $8.5B traditional Medicare). But its title is Assisted Living Facilities: Information on Federal Spending and Medicaid Coverage (2026-06-02) and it contains no critical-incident-reporting finding.

The whitepaper's "two federal watchdog reports eight years apart both found most states can't count assisted-living safety incidents at all" attributes a 2018 finding to a 2026 report about a different subject. The sources page's "confirm the same undercounting gap" does the same more softly; ws07's "confirms the gap persists" and the scorecard's AL-standards basis column both repeat it. The two reports do document undercounting — of different things. That is a weaker and more accurate claim, and it is the one the record should make.

UNVERIFIABLE — the census comparison. NPALS (2022) and the CASPER-derived SNF count were both blocked this pass (cdc.gov returned 403 to every method tried; the KFF indicator page 404'd). The record's own hedging on this pair is good and is left standing.

Cluster G — Family caregiving (P135–P155, S124–S131)

CORRECTED — "for a comparable period." Part 7: "Independent, credible estimates for a comparable period range from $234 billion to $1.01 trillion, depending on population scope, wage-rate assumption, and hours source — three named, defensible methodological choices, not chaos." The endpoints are not from a comparable period. $234B is CBO 2013 on 2011 data; $1.01T is AARP's 2026 update on 2024 data. The record knows this: ws08 and the anchor table restrict "comparable era" to CBO $234B and RAND $522B, and anchor row 4 records that AARP's own series moved $350B (2006) → $600B (2021) → $1.01T (2024) "partly from methodology revision, not just growth." Elapsed time is a fourth driver of the published spread, and naming only three understates how much of the range is calendar rather than method. Phase 0 §1 has the same three-knob framing.

OVERSTATED — the BRFSS trend. "CDC's own national health survey found caregivers' mental-health gap versus non-caregivers widened, not narrowed, over the last decade." Red-team item #19 already downgraded this: the 2021–22 endpoint sits inside the pandemic window, persistence is unchecked, and the item's own verdict was "downgraded from 'strongest non-advocacy evidence' to suggestive, confounded by pandemic timing." The whitepaper prints the un-downgraded version.

UNVERIFIABLE — the underlying numbers. BRFSS 19.6%/12.9%, the Cash & Counseling satisfaction and cost figures, CBO 2013 and RAND 2014: none reached at primary tier. cdc.gov and cbo.gov both blocked (cbo.gov's DataDome block is already recorded in the atlas's UNOBTAINED.md §§1–2 from the childcare pass); the RAND PDF 404'd; the ASPE Cash & Counseling landing page resolves but does not carry the figures. The substance — that Cash & Counseling is the domain's one real RCT, that it improved caregiver-reported strain and health, and that it cost more rather than less — is consistent across ASPE's own framing and is left CONFIRMED at secondary tier.

Cluster H — International (P156–P166, S132–S137)

CORRECTED (citation) — NBER w31870 does not contain the Germany figures. The paper is Long-Term Care in Germany (Geyer, Börsch-Supan, Haan & Perdrix, NBER WP 31870, November 2023). It was fetched in full (3,448 lines of extracted text). It contains no vacancy series, no demand-growth series, and no workforce-growth series. Its only "13%" is the 2021 share of foreign nationals in the German LTC workforce. It cannot support "workforce vacancies up 110% against 45% demand growth and only 13% workforce growth," which ws10 and the scorecard both attribute to "NBER/Geyer" and the whitepaper cites as its Part 8 source. Nor can a November 2023 chapter support a claim about Germany's 2026 premium.

The underlying claims may well be true — German LTC vacancy data comes from the Bundesagentur für Arbeit, and the 2026 contribution-rate loan is a live German policy story — but the record's actual source for the 2026 rate is a German consumer-finance website, and no primary German source was reached by the filing or by this pass. Under V2 this is UNVERIFIABLE with a wrong citation attached, which is the more serious half.

CORRECTED — "Japan … is already raising copays." ws10: copay increases toward 30% "are already being proposed — active benefit-tightening." Proposed is not raising. The record has it right; the whitepaper does not.

Cluster I — The scorecard (P167–P196, S138–S147, R187–R226)

This cluster is the filing's strongest, and it is worth saying so plainly in a log otherwise full of corrections.

Every figure was recomputed independently from the reconciled matrix. All 24 weighted sums under the corrected scale reproduce exactly. All 24 under the published-scale sensitivity reproduce exactly. The row sums reproduce. The counts reproduce: 30 cells at 3 (recounted from the published table), exactly 20 rows in the re-score log's section B, exactly 14 in section E, 26 + 14 = 40. The published table on site/elder-care/index.html matches the reconciled matrix cell for cell. The withdrawal of the private-LTC-reform headline is correctly reasoned and correctly scoped ("the claim is not robust," not "the claim is false"), and the sensitivity analysis publishing both scale readings is exactly what S5 asks for. Verdicts: CONFIRMED-INTERNAL throughout.

Three defects, none arithmetic:

• S198 — "moved 26 of the 40 ranked cells." The re-score log, deviations #30, the scorecard, the whitepaper and three other places on the same sources page all say the board moved on 25 and differed on 26. The Record 4 correction paragraph is the only place that says 26 moved.
• The two cells that carry the federal-LTC-insurance row's "evidenced" status rest on Cluster H. The scorecard's own basis column for that row cites the German vacancy figures (R174) and the 2026 rate loans (R173) as what makes it evidenced. The re-score log's proudest structural finding — "federal LTC social insurance ranks last … on evidence, which is what the published bottom rank claimed to be and was not" — therefore rests partly on a miscited source. The ranking does not move (both cells sit at 2 either way), but the claim that this row's last place is evidence-backed needs the citation fixed before it is repeated.
• ws12's sequencing conclusion was not updated. It still says the protocol's workforce-first order "remains directionally sound," while the pass-3 board withdraws the scorecard's support for it under both readings — a fact the whitepaper and the re-score log both state.

Cluster J — Process, counts and self-description

• CORRECTED — "12-workstream protocol" (whitepaper receipts). The sources page says eleven (§2–§12, nine executed to findings) in two places, and the record agrees.
• CORRECTED — "21 entries" (sources page, Record 3). The committed log has 31, and the same page's own receipts line says 31.
• OVERSTATED — "all public" (whitepaper receipts). The published deviations digest presents 19 rows covering entries 1–21 and 30–31. Entries 22–29 — the whole PACE late-pass sequence, including the M3-criteria-written-after-the-fact disclosures and the CMS-directory extraction correction — appear on neither public page. M6 makes the deviations log part of the record, not an appendix to it; publishing two thirds of it under a wrong total is the one process claim on these pages that does not hold.
• OVERSTATED — deviation #23 ("completed the line-edit pass queued in #19"). Two of the six files still read pre-red-team: ws02's confidence tier and ws04's §4 headline.

Corrections applied on this branch

Dependency order per the corrections protocol: research record → anchor table → scorecard basis → both public pages.

Public-page replacement wording is recorded inline in the diff. The load-bearing ones:

• "mooted by a ten-year congressional moratorium" → "mooted by a congressional moratorium running to September 30, 2034"
• "71% of it went to wages" → "71% went to recruitment and retention"
• "an 11% mortality increase in private-equity-owned nursing homes, across seven million patients" → "an 11% mortality increase for patients steered into private-equity-owned nursing homes, in a 4.2-million-patient study"
• "No published rebuttal exists." → "No direct econometric rebuttal has been found, and the authors themselves report the effect falls on a subset of patients."
• "two federal watchdog reports eight years apart both found most states can't count assisted-living safety incidents at all" → "a federal watchdog found in 2018 that most states studied couldn't count assisted-living safety incidents, and in 2026 that the government still can't count what it spends on the setting"
• "Independent, credible estimates for a comparable period range from $234 billion to $1.01 trillion" → "Independent estimates range from $234 billion to $1.01 trillion — and they are not measuring the same year"
• "Japan … is already raising copays" → "Japan … is already proposing copay increases"
• "Germany needed federal loans … amid a workforce vacancy rate that rose 110% in a decade" → vacancy figure withdrawn pending a citable source
• Part 2 gains the scope disclosure and the 5% waiting-list-targeting fact; the honesty box gains a line naming the under-65 majority.

What this pass did not settle

• Neither vacatur opinion has been read. The atlas records four failed routes (UNOBTAINED.md §3); none were re-attempted here. Everything about the courts' reasoning still comes from the agency that repealed the rule.
• cdc.gov and cbo.gov remain unreachable, which leaves the BRFSS caregiver figures, the NPALS census and the CBO 2013 valuation UNVERIFIABLE. The cbo.gov block is the same DataDome mechanism the childcare pass documented.
• No primary German or Japanese source was reached. Cluster H is the weakest-sourced section of the filing and this pass improved its citations without closing its evidence gap.
• 90 UNVERIFIABLE rows. Most are secondary-tier record claims that were never load-bearing, but the concentration in ws10 and in the state-level HPRD list is real, and deviations #11 already flagged the latter.
• This is a fact-check, not a steelman. Whether the filing's conclusions survive the strongest opposing case is Phase 2, which must run in a different session per S1.

Phase 2 follow-up: four presentation fixes applied after the steelman (2026-08-10)

A review of the Phase 2 steelman branch before merge found four defects of presentation and completeness. None of them changes a finding — every substantive Phase 2 finding was independently re-verified as CONFIRMED against primary sources — but three of the four are the kind of gap the protocol exists to catch, so they are recorded here rather than fixed silently.

1. The steelman fetched five primary sources and committed none of them. M-tier convention is that every source a verification pass relies on is persisted in method/sources/ behind a provenance header, so a later pass reads it from disk instead of re-fetching a URL that may have moved. Phase 2 cited fetched text for Pub. L. 119-21 §71121, CMS's National Health Expenditure Tables 14 and 15, the GUIDE model, MedPAC's March 2025 ch. 11, and 42 U.S.C. §1395w-22(a)(3)(D) while adding zero files. All five were re-fetched and committed: method/sources/plaw-119-21-obbba-lihtc-and-staffing-moratorium.md (extended with a §71121 section — the atlas already held that public law for §§70422 and 71111, so a second file for the same act would have been the duplicate the atlas is designed to avoid), cms-nhe-tables-14-15-home-health-and-nursing-facility.md, cms-guide-model-dementia-care-and-respite.md, medpac-march-2025-ch11-medicare-advantage.md, and usc42-medicare-custodial-care-exclusion-and-ma-supplemental-benefits.md (which also archives §1395y(a)(9), the custodial-care exclusion the surviving half of the verdict rests on). Every figure and quote Phase 2 published from these sources string-matched the fetched text, with one exception recorded in the NHE file: the steelman prints Medicaid's 2024 home-health share as 22.6% where CMS's own percent-distribution row prints 22.5 (a rounding convention on a non-load-bearing cell). The headline figure — Medicare at 32.9% of home health care — matches CMS exactly.

2. A statutory qualifier was dropped in the public paraphrase. Part 4 of the whitepaper said §71121 "bars the money from paying for direct-care workers' health insurance or training." §71121(C) bars that only "in the case of a class of practitioners for which [Medicaid] is the primary source of revenue" — a qualifier steelman-log.md states correctly and the public page had lost. Restored on site/elder-care/index.html. The un-qualified version overstates the constraint on a workforce-first architecture.

3. A stale count on the whitepaper's receipts line. It read "deviations log (34 entries)" against a log of 37. Corrected. The sources page already read 37.

4. The corrected rail verdict had not propagated off the filing. Phase 2 narrowed the verdict from "Medicare mostly doesn't pay for this" to "Medicare doesn't pay for custodial care. It pays for plenty of the rest." — but the withdrawn claim was still being published in three other places outside the steelman's original scope: site/index.html (the GBMT-7 investigation card and the folder card), site/404.html (the on-file index), and site/index-v2.html (the alternate homepage's elder-care feature, headline and standfirst). All four instances now carry the narrowed claim, phrased for their length. A withdrawn claim that survives on the homepage is still a published claim.

Not fixed, and named rather than left silent: elder-care/report/gubment-elder-care-report.html is a generated derivative of site/elder-care/index.html (see scripts/build-report-pdf.py) and still carries the pre-Phase-2 rail verdict and the 34-entry count. Patching the two strings by hand would produce a half-regenerated artifact; it needs build-report-pdf.py re-run, which also re-renders the PDF, and that is a separate step from this correction pass.