GBMT-7 (Elder care) — Phase 2 Steelman Log

Protocol: method/verification-protocol.md, Phase 2 (S1–S5). Filing under test: Whitepaper No. 7, filed 2026-08-03, as corrected by Phase 1 (2026-08-10) and re-scored by pass 3 (2026-08-10). Date: 2026-08-10.

Independence (S1). This is a different session and a different worktree from the Phase 1 fact-check, which is what S1 requires. Phase 1's corrections had fully landed before this pass began, so no steelman premise here rests on an uncorrected fact.

Deviation from the childcare precedent, logged not hidden. GBMT-1's Phase 2 built each steelman in a separate agent context and had the adjudicator check them. This pass built and adjudicated all three cases in one session. Mitigations: the S2 tilt statement below was written and committed to disk before any steelman evidence was gathered, and each case records the evidence that defeats it, including the two places where the steelman loses outright. A reader who thinks a single-context build is weaker than three independent ones should discount accordingly — the strongest items here (§71121, the CFC null) are single primary documents that can be checked in a minute, and do not depend on trusting the builder.

Every source cited below was fetched in this session. No figure appears here that was not read inside the fetched document itself.

S2 — The tilt, derived from the filing's own text

Written before any steelman evidence was gathered.

Which direction it leans. Toward a public-programmatic diagnosis: LTSS is a public benefit whose problems are payer architecture (mandatory entitlement vs. capped waiver), staffing regulation, and public insurance design. All nine scored architectures are federal or state public-program instruments. The market side appears only as failure narrative — LTC insurance "collapsed," reinsurance "didn't pencil out," assisted living is a "blind spot," private equity raises mortality.

Which conclusion the evidence was marshalled toward. Toward the binding constraint is direct-care workforce capacity, not Medicaid funding — pre-committed before fieldwork. research-inquiry.md §3 reads "Workforce (treat as the primary binding constraint, per the childcare and housing rhyme)"; §12 pre-orders "workforce wage floor and pipeline first." The verdict came back INDETERMINATE, which is honest. The framing survived the verdict.

Which architecture won. The board is deliberately flat after pass 3, but HCBS de-capping leads workforce weighting under both scale readings and holds the only positive causal cell on the board (WS=4). Cash & Counseling leads caregiver relief decisively.

Whose framing it adopted. The Medicaid-rebalancing frame: home care is under-provided because the entitlement is institutionally biased, waiting lists measure suppressed demand, and the fix is de-capping. Part 4, verbatim: "backwards from the 'home care is cheaper, so of course it's prioritized' intuition."

The stamped verdict. "Medicare mostly doesn't pay for this." — printed in the filing rail at the top and bottom of every screen, restated in Part 1 and in the plain-talk box. It is the most repeated sentence in the document and the one the homepage sells the filing on.

What it argued against, scored down, or never entertained. Four things, established by search across elder-care/ and site/elder-care/ before any case was built:

1. The verdict itself. Nothing in the filing asks whether Medicare's LTSS footprint has moved. "Medicare Advantage" — 0 hits in the entire filing and record. "Dementia" — 2 hits.
2. That capping HCBS might be fiscal design rather than pathology. "Woodwork" — 0 hits. The scorecard's own CC=2 cell for HCBS de-capping is that objection, unnamed, and it appears on neither public page.
3. That self-direction is already federal law and mainstream practice. Part 9's heading says nobody is proposing it. "Community First Choice" — 0 hits.
4. That the financing base under every scored architecture just moved. The filing cites Pub. L. 119-21 §71111 on the public page and engages no other section of that act.

Therefore the three steelman targets — the filing's stamped verdict, its top-ranked architecture, and its Part 9 headline:

S3 / S4 — Steelman A: Medicare is becoming a long-term-care payer

A-1. Under the accounts the public actually means, Medicare is the largest payer of home health care — STEELMAN WINS on framing

Fetched this pass: CMS National Health Expenditure Accounts, Tables 14 and 15 (nhe-tables.zip, cms.gov, 2024 estimates, retrieved 2026-08-10).

Medicare is the single largest payer of home health care in the United States, and the third largest payer of nursing-facility care. Across the two settings an ordinary reader means by "elder care," Medicare pays 26.5 cents of every dollar against Medicaid's 30.1.

The filing's $415B / 61%-Medicaid figures come from KFF's LTSS construct, which excludes Medicare post-acute care by definition. Both constructs are legitimate. The filing publishes one of them as an unqualified fact and names the definitional problem in one direction only — Part 1, verbatim: "federal accounting categories mix short-stay Medicare rehab with long-stay Medicaid custodial care in ways that understate Medicaid's true share." The reciprocal sentence — that the LTSS construct understates Medicare's role in the settings the reader is picturing — is never written.

A-2. The custodial-care exclusion is statutory, and the filing is right about it — STEELMAN LOSES OUTRIGHT

42 U.S.C. §1395y(a)(9), fetched from uscode.house.gov: no payment may be made "where such expenses are for custodial care (except, in the case of hospice care, as is otherwise permitted under paragraph (1)(C))."

The filing's core insight — that Medicare does not fund custodial long-term care, and that most people believe it does — is correct at statutory tier and is the most useful thing on the page. Recorded here because S4 requires stating what a surviving conclusion beat.

What the exclusion does not reach is the rest of this case: it binds Parts A and B. It does not bind Part C supplemental benefits, and it does not bind CMMI models.

A-3. The Part C channel exists, is large, and is unmeasurable — STEELMAN WINS on existence, LOSES on size

• 42 U.S.C. §1395w-22(a)(3)(D), added by Pub. L. 115-123 §50322(a)(2) (fetched): "For plan year 2020 and subsequent plan years … an MA plan … may provide supplemental benefits … that, with respect to a chronically ill enrollee, have a reasonable expectation of improving or maintaining the health or overall function of the chronically ill enrollee and may not be limited to being primarily health related benefits." Congress removed the health-relatedness limit for chronically ill Medicare enrollees six years before this filing.
• MedPAC, Report to the Congress, March 2025, ch. 11 (fetched): in 2024 MA "enrolled about 33.6 million beneficiaries (54 percent of Medicare beneficiaries with both Part A and Part B coverage)." In 2025, 28% of rebates — $53 per enrollee per month — are projected for non-Medicare supplemental benefits.

And the same chapter defeats any attempt to size it. MedPAC, verbatim: "the data that Medicare collects are insufficient for examining the use of supplemental benefits, making it impossible to know how much plans spend on each type of benefit, which enrollees use each benefit … or whether service use differs." So the steelman cannot claim a quantity of LTSS bought through Part C. It can claim the authority, the enrollment base, and the money — and must stop there.

A-4. Medicare already pays for respite care, nationwide, and the filing does not know it — STEELMAN WINS

CMS, GUIDE (Guiding an Improved Dementia Experience) Model (cms.gov, fetched):

The model began on July 1, 2024, and will run for 8 years. … Respite services payment – CMS reimburses participants up to $2,500 annually per eligible patient for services that temporarily relieve qualifying caregivers of their caregiving responsibilities, including in-home care, adult day center programs, and facility-based respite.

Plus care navigation, a 24/7 support line, and required caregiver training and support.

Part 7 of the whitepaper is about family caregivers and the single architecture with an RCT behind it. A nationwide Medicare model paying for caregiver respite and dementia care management, running for twenty-five months before the filing date, appears nowhere in the filing or its record — "GUIDE model" returns 0 hits.

Counter, stated because it is real: GUIDE is a CMMI model with a fixed term, not a benefit; $2,500 a year is small against the cost of care; and a model can be cancelled. It is not proof that Medicare has become an LTSS payer. It is proof that the sentence "Medicare mostly doesn't pay for this" is no longer safe as a flat statement about caregiver support.

Verdict on Steelman A: partially survives (S4 outcome 2)

The verdict is right about the custodial benefit — statutorily right, and worth keeping. It is stated too broadly. Medicare pays a third of the nation's home-health bill, a fifth of its nursing-facility bill, may buy non-health-related supports for chronically ill enrollees in plans covering 54% of beneficiaries, and pays for dementia respite nationwide. Applied to the record per S4, not footnoted: the verdict line is narrowed to what the statute actually says, and Part 1 now carries the NHE numbers and names the construct it is using.

S3 / S4 — Steelman B: the cap is fiscal design, and Congress has already partly removed it

B-1. The waiver cap is the price of the waiver, not only a preference for institutions — STEELMAN PARTIALLY WINS

42 U.S.C. §1396n(c)(2)(D), fetched: a §1915(c) waiver may be granted only if

under such waiver the average per capita expenditure estimated by the State in any fiscal year for medical assistance provided with respect to such individuals does not exceed 100 percent of the average per capita expenditure that the State reasonably estimates would have been made in that fiscal year … if the waiver had not been granted.

Phase 1 verified the filing's legal description of the entitlement/waiver asymmetry and found it sound; nothing here disturbs that. What the filing omits is the condition that produces the cap. Congress made HCBS available by waiver on the condition that it cost no more per person than the institutional care it replaces. An enrollment cap is one of the two levers a state has to satisfy that test. So a waiting list is what a budget-neutral optional benefit looks like — not only what institutional preference looks like. Part 4 tells the reader the structure is "backwards" and never tells them what the structure is for.

B-2. The woodwork objection fails — on the filing's own citation — STEELMAN LOSES OUTRIGHT

The standard forty-year objection to HCBS expansion is the woodwork effect: expand home care and people who would never have entered a nursing home come out of the woodwork, so spending rises without substitution. The word appears nowhere in this filing or its record. I built the case and went to the evidence, and the evidence is already in the filing's own footnote.

McGarry & Grabowski, J Am Geriatr Soc 71(10):3143–3151 (2023) — PMID 37326313, the exact paper ws13 cites for "$0.74" — is titled Medicaid home and community-based services spending for older adults: Is there a "woodwork" effect? Its abstract, fetched:

We did not find evidence of a woodwork effect in those states that expanded Medicaid HCBS more aggressively … A $1 increase in HCBS spending was associated with $0.74 increase … in total LTSS spending, suggesting each dollar directed to HCBS was offset by $0.26 savings from decreased nursing home use. Increased HCBS waiver spending was associated with more older adults receiving LTSS at a lower cost per beneficiary served relative to the nursing home setting.

The steelman loses this point, and the filing gains one it never claimed. Its record holds the best available refutation of the standard objection to its own leading architecture and reports it only in the negative — "partial offset only, self-financing unsupported." That is accurate and one-directional. Both halves belong on the page.

B-3. Congress enacted a partial de-capping in the same statute the filing cites — STEELMAN WINS OUTRIGHT

Pub. L. 119-21 (2025-07-04) — the act the whitepaper cites on the public page, by section number, for the staffing moratorium. Ten sections later, in a subchapter titled "Expanding Access to Care," §71121 amends §1915(c) by adding a new paragraph (11). Fetched from govinfo:

Beginning July 1, 2028, notwithstanding paragraph (1), the Secretary may approve a waiver that is standalone from any other waiver approved under this subsection to include as medical assistance … payment for … home or community-based services … provided pursuant to a written plan of care to individuals described in subparagraph (B)(iii).

Paragraph (1) is the institutional level-of-care test. Subparagraph (B)(iii) requires the state to establish "needs-based criteria … regarding who will be eligible for home or community-based services under a waiver approved under this paragraph without requiring such individuals to have a determination described in paragraph (1)." Section 71121(b) appropriates $50,000,000 for FY2026 and $100,000,000 for FY2027 to build state HCBS delivery systems.

That is the filing's leading architecture, partially enacted, in the statute the filing cites for something else. GBMT-7 describes HCBS de-capping as a candidate architecture and scores it as though it were untried policy.

Three details make this sharper rather than softer, and two of them cut against the filing's framing:

• §71121(B)(v) requires the state to attest that per-capita expenditure for the new population "will not exceed the State's average per capita expenditure for medical assistance for individuals receiving institutional care." Congress expanded access and re-imposed the cost-neutrality condition of B-1 in the same breath. The constraint is not an artifact of institutional bias; it is the durable design.
• §71121(B)(ii) requires the state to demonstrate that the new waiver "will not result in a material increase of the average amount of time that individuals [meeting the institutional level-of-care test] will need to wait to receive home or community-based services under any other waiver." Congress wrote queue competition between HCBS populations into the statute — the mechanism the filing's Part 4 waiting-list argument never considers.
• §71121(C) bars any payment under the new paragraph from being used "to make payments to a third party on behalf of an individual practitioner for benefits such as health insurance, skills training, and other benefits customary for employees" where Medicaid is the practitioner class's primary revenue source. The money appropriated to expand HCBS is expressly barred from funding the benefits a workforce-first architecture depends on — a live statutory constraint on §12's sequencing prior, in the same section.

Counter, and it is a real one: §71121 vindicates Part 4's diagnosis. Congress looked at the same institutional-bias problem and acted on it. The filing is not wrong about the architecture; it is wrong that the architecture is hypothetical, and it missed that the enacted version carries the budget constraint and the anti-workforce-benefits clause with it.

Verdict on Steelman B: partially survives (S4 outcome 2); B-3 wins outright

The rebalancing case survives — B-2 defeated the strongest available objection to it using the filing's own source. What does not survive is the presentation: the cap's statutory purpose is missing, the woodwork refutation the filing owns is unpublished, and the enacted partial de-capping is absent from a page that cites the enacting statute.

S3 / S4 — Steelman C: the best-evidenced fix is already law, and already mainstream

The target is the Part 9 heading, verbatim: "The scorecard: the best-evidenced fix isn't the one anyone's proposing."

C-1. Congress enacted it twice, and pays states a bonus to adopt it — STEELMAN WINS

Both fetched from uscode.house.gov this pass:

• 42 U.S.C. §1396n(j), "Optional choice of self-directed personal assistance services," added by Pub. L. 109-171 §6087(a) (Deficit Reduction Act of 2005). It made the Cash & Counseling design a permanent state plan option: "A State may provide, as 'medical assistance', payment for part or all of the cost of self-directed personal assistance services."
• 42 U.S.C. §1396n(k), "State plan option to provide home and community-based attendant services and supports" — Community First Choice, added by Pub. L. 111-148 §2401 (ACA). Its paragraph (2): "the Federal medical assistance percentage applicable to the State … shall be increased by 6 percentage points."

Congress did not merely propose this instrument. It wrote it into the Social Security Act twice and attached a six-point federal match bonus to make states take it.

C-2. It is in every state, and the filing's own record says so — STEELMAN WINS; this is a public-page failure, not a research gap

ws08-findings.md, line 14, committed in the filing's own record:

Today: self-direction is available in all 50 states + DC via 1915(c)/(i)/(j)/(k) and 1115 authorities; ~1.5M people self-directed Medicaid HCBS in 2023.

Corroborated independently this pass at named-professional tier — AARP LTSS State Scorecard / Applied Self-Direction, National Inventory of Self-Directed LTSS Programs, 2023: "there are more than 1.5 million people who self-direct their LTSS. This represents a 23 percent increase since 2019. … Most states (44) reported an overall increase in self-direction enrollment and six of those states have more than doubled self-direction enrollment since 2019."

The research record found this and the public page contradicts it. This is the same shape as the Phase 1 pattern — the record gets it right; the whitepaper narrows it — and here the whitepaper does not narrow the claim but inverts it.

The charitable reading, tested: "nobody's proposing it" could mean "no federal bill leads with it." Even that is shaky — §1915(k)'s 6-point FMAP bonus is standing federal law, and Pub. L. 119-21 §71121 is a 2025 federal HCBS expansion. On the plain reading a reader takes from the page — a policy nobody is pushing — it is false.

C-3. The filing's own cited evaluation reports a null on the statutory version, and the filing prints only the other half — STEELMAN WINS; most damaging item in this pass

The HCBS row's WS=4 — the only positive causal cell on the board — cites Ne'eman, Health Services Research 61(2):e70018 (2026), PMID 40845205. Its abstract, fetched:

This study finds that BIP resulted in a 13.24% (95% CI: 1.14%, 25.34%) increase in the size of the HCBS workforce in participating states, while finding no statistically significant effect for CFC (1.51%, 95% CI: −12.77%, 15.79%). … Federal policymakers seeking to increase investment in HCBS should ensure that additional funds are tied to measurable performance targets.

The paper's title is Evaluating the Affordable Care Act's LTSS Rebalancing Programs — plural. It evaluates two. ws13, ws11-scorecard.md, and both published source pages report the BIP result and omit the CFC null. "Community First Choice" appears zero times in the filing and record.

This matters twice over. CFC is the statutory instrument closest to both headline rows — it is attendant services delivered on the self-direction model, i.e. the scaled statutory version of Cash & Counseling, and it is an HCBS expansion, i.e. the scaled version of de-capping. The best available evaluation of it finds no workforce effect, in the same paper and the same design that supplies the board's strongest positive cell. The filing cites the paper for the finding that helps and not for the finding that does not.

C-4. What defeats the steelman

Scale is not effect. 1.5M participants is not evidence the instrument works; C-3 is if anything evidence that scaling it through the state plan option did not produce the workforce gain. And the filing's substantive claim — that Cash & Counseling is the only architecture on this board anchored by a real randomized trial — is untouched by any of this. Only the framing falls.

Verdict on Steelman C: the steelman wins on the published claim; the underlying finding survives

"The best-evidenced fix isn't the one anyone's proposing" is false and is withdrawn. What replaces it is more interesting and is the filing's own material: the best-evidenced fix is already law in every state and has been scaled without anyone measuring whether the scaling worked — and the one study that looked found nothing.

S5 — Scorecard consequences

The board was independently recomputed before any sensitivity was run

All 24 weighted sums under the published pass-3 table were recomputed from the reconciled matrix in this session. All 24 reproduce exactly, including the three-way tie at 21 and both joint firsts. The published arithmetic is sound; nothing below is an arithmetic correction.

Sensitivity 1 — federal LTC social insurance does not robustly rank last

Pass 3's replacement headline: "Federal LTC social insurance now ranks last under all three weightings … Unlike the row it displaced, this one ranks last on evidence."

Steelman A contests the WS=2 cell on a ground the pass-3 scale amendment was written to catch. That cell's basis is "Japan projects a 570,000-worker shortfall by 2040 and the Netherlands 266,000 by 2035." Those are projections of national workforce shortfalls in countries that have LTC social insurance. They are not estimates of the instrument's effect. The comparator that has no such program — the United States — has the same shortfall: the filing's own anchor 2 records PHI's projection of 9.7 million total direct-care openings, 2024–2034. A cell scored below neutral on a phenomenon the counterfactual also exhibits is the defect the amended scale exists to prevent; under the amended anchors, "no evidence in the record either way" is a 3.

Moving that one cell 2→3 (row becomes 2·3·3·3·3, sum 14):

"Ranks last under all three weightings" becomes "ties for last under all three weightings" on one cell. The pass-3 replacement headline is not more robust than the pass-2 headline it withdrew — it is the same shape of claim, resting on the same kind of below-neutral cell, one step further along. Published as a sensitivity per S5; not entered as the cell's value, because deciding a cell is the filing's call, not the checker's.

The CC=2 cell on the same row is left alone, deliberately. It rests partly on HHS's CLASS review, which is instrument-specific domestic evidence of a design-stage failure. It is a fair below-neutral score even though the row is "WA Cares nationalized" — i.e. mandatory, which is the channel CLASS failed on. That tension is worth a future pass; it is not resolved here.

Sensitivity 2 — Phase 1's German withdrawal does not move the row

Checked and reported because it would have been the obvious move. Phase 1 withdrew the German vacancy figures (miscited to NBER w31870) and flagged the 2026 rate-loan claim as consumer-finance-sourced. Removing both leaves CC=2 supported by CLASS, the Dutch Wlz and Japan's copay proposals, and WS=2 supported by Japan and the Netherlands. Neither cell moves; the row's rank is unchanged. The filing survives this one.

Sensitivity 3 — two contested items the board cannot express

Both are structural, so no re-score is offered; the absence is the finding.

• There is no dimension that can register that an architecture has already been enacted. Pub. L. 119-21 §71121 partially enacts HCBS de-capping from 2028-07-01, with $150M of implementation money, a per-capita cost cap and a bar on funding worker benefits. None of that can move a cell on a board whose five objectives are cost, quality, workforce, caregiver relief and autonomy. GBMT-1 scored procedural feasibility; GBMT-7 does not.
• There is no row for a Medicare-financed instrument. GUIDE-shaped dementia care management with paid respite, and MA supplemental benefits for chronically ill enrollees, are live federal instruments serving objectives the board scores. Adding architectures to a published board is the filing's decision, not the checker's — recorded as a recommendation.

Does the recommendation move?

No. The filing's published position is that the board is mostly flat, that the differences between these instruments are mostly differences in how hard it looked, and that §12's workforce-first sequencing draws no support from the board. All three survive Phase 2 intact, and B-3's §71121(C) independently strengthens the third: the statute Congress actually passed forbids spending HCBS expansion money on direct-care worker benefits.

What moves is what the filing says about three specific things: its verdict, its lead architecture's novelty, and its Part 9 headline.

Convergence, and the one pattern worth stating

Three cases, three targets, built against the filing's verdict, its top-ranked architecture, and its loudest headline. Two of the six load-bearing findings are omissions from documents the filing already cites:

• ws13 and the scorecard cite Ne'eman 2026 for BIP and drop the CFC null in the same abstract.
• ws13 and the scorecard cite McGarry & Grabowski 2023 for "$0.74" and drop the conclusion in its own title.

And one is an omission from a statute the filing cites by section number on the public page — Pub. L. 119-21, cited for §71111, silent on §71121.

The pattern: GBMT-7's sourcing is good and its reading of its own sources is one-directional. In each case the half of the source that supports the filing's frame is published and the half that complicates it is not — not suppressed, just not carried across. Phase 1 found the same shape from the other end ("the record gets this right; the whitepaper narrowed it"). Two independent passes converging on one failure mode is the most useful thing this verification produced about GBMT-7, and it is a reading discipline, not a sourcing budget.

What this pass did not settle

• No steelman was built against Part 5. The private-equity mortality result is the filing's best evidence and has no row on the board; the contrary literature was not searched this pass.
• The size of the Part C channel is unmeasurable, by MedPAC's own statement. Steelman A's strongest quantitative claim stops at authority and enrollment.
• §71121 has not been implemented. It takes effect 2028-07-01; no state has applied under it, and whether the Secretary approves anything is unknown.
• This pass was built and adjudicated in one session, not in independent builder contexts. See the deviation at the head of this file.