Date: 2026-08-11 Model: claude-opus-5 via the Anthropic Message Batches API (scripts/batch-rescore.py, batch msgbatch_013vUDQ6mt6JV8mz2H95zfsx). Blinding: structural. The scorer received one tool-less request containing the scoring brief (ws09-blind-brief-2026-08-11.md), the anchored scale (00-scales.md), and the evidence packet below inlined as text. It had no filesystem access and no search, so it could not reach ws09-scorecard.md, ws09-red-team-log.md, or deviations-log.md by any route.

Reconciled in: ws09-rescore-log.md. What follows after the packet manifest is the scorer's output, verbatim and unedited.

Packet manifest

Evidence was inlined from /private/tmp/gubment-batch-rescore-gbmt11-12/mental-health/ (batch working directory). Evidence files are not re-copied into git — they match the committed workstreams/phase0 on origin/main (SHA-256 checked at reconcile). Brief + raw scores are the research artifacts.

What the scorer did not receive: ws09-scorecard.md, ws09-red-team-log.md, deviations-log.md, any site/mental-health/*, and this reconciliation log.

Evidence SHA note: At reconcile (2026-08-11), every evidence/ws0*.md and phase0-findings.md byte-matched the corresponding files on this branch (origin/main tip). 00-scales.md is the authored abstract scale shipped in the brief packet (aligned with the scorecard's O1–O5 anchors, including the symmetric evidence floor already baked into pass 1).

GBMT-11 Mental Health — blind re-score

1. The matrix

Cell IDs below: rows a1–a10 = 1–10, columns O1–O5 = A–E.

2. Cell-by-cell basis

a1 — Medicaid rate floor + administrative simplification

a2 — CCBHC expansion

a3 — IMD repeal / broad MH IMD waiver

a4 — Bed rebuild / state-hospital reinvestment

a5 — 988 + mobile crisis + stabilization continuum

a6 — Parity enforcement with ERISA teeth

a7 — Assisted outpatient treatment

a8 — Collaborative Care / primary-care integration

a9 — Workforce liberalization

a10 — Do-nothing comparator

3. What my matrix implies

On my numbers the board is mostly flat, and the flatness is the finding: 32 of 50 cells sit at 3, and 15 of those are unevidenced-neutral. a2_ccbhc_expansion and a5_crisis_continuum tie as strongest — they are the only rows with cited positive effect evidence on two or more substantive axes (CCBHC: measured wait-to-evaluation and client-volume gains under Medicaid; crisis continuum: measured crisis connection plus two independent continuum studies on ED use and arrests). a9_workforce_liberalization and a10_do_nothing tie as weakest, for opposite reasons: a9 because the record documents failure of the licensure→usable-access channel it depends on, a10 because the record documents unrelieved acute stress, closed panels and unclosed parity gaps on the status-quo trajectory. a3_imd_repeal_waiver and a4_bed_rebuild come out numerically identical (3,3,3,3,4) — I am saying that plainly: the evidence base gives one a documentary financing constraint offset by a bed-count null, and the other essentially nothing but need documentation, and both land on the neutral floor. a1 and a7 sit mid-board with two evidenced cells each; a6 is single-cell evidenced (O4) and constrained by KC3; a8 is the only row with zero evidenced non-neutral cells.

4. Evidence-coverage count

Total: 18 cells scored on cited evidence, 8 evidenced-wash, 24 unevidenced-neutral (one row, a2, has its 3s split 1/1).

5. What I could not score, and what would change my mind

Structurally unscoreable in this base:

• O4 for every row except a6 and a10. The record has one parity/financial-protection evidence stack (MHPAEA enforcement + state exams + KC3), and it touches only the parity architectures. Nothing in ws01–ws08 measures OOP exposure for Medicaid, CCBHC, crisis, AOT, bed, CoCM or workforce instruments. Eight O4 cells are pure floor.
• O3 for everything except a1, a2, a9, a10. The workforce-participation evidence (Bishop / Brahmbhatt / Zhu) is diagnostic of the problem, not of any instrument's effect on it, except where a rate/PPS design was actually evaluated.
• a8 entirely. There is no CoCM evaluation, no CoCM billing-code uptake data, and no integration outcome anywhere in this base — only NACHC financing-barrier statements. If someone scores a8 above 3 on O1, they are scoring the general CoCM literature, which is not in this evidence base.
• a4's core claim (O2). The whole rationale for bed reinvestment — that added acute/forensic capacity relieves boarding and admit constraints — has no cited instrument evaluation here. KC2 removed the historical baseline; nothing replaced it with a contemporary reinvestment evaluation.
• Realized access under commercial coverage anywhere. KC1 plus KC3: Brahmbhatt has no commercial arm, no by-payer wait series exists, and the self-funded majority is band-only. Every commercial O1 judgment in this board is proxy-free.

What would move specific cells:

• 1A/1C → 4/5: a rate-change design with an appointment-offer or wait outcome, or evidence of new panel openings at post-change rates near 90–100% of Medicare (ws05 hints the elasticity is concentrated there but does not measure panel entry).
• 2B → 4: a multi-state CCBHC ED/hospitalization DID that is not heterogeneous, or a crisis-throughput measure tied to the required crisis criterion.
• 3B → 4 or 2: a waiver-vs-non-waiver boarding or admit-financing evaluation. ws03 explicitly logs this comparison pair as not located; it is the single highest-value missing study on this board.
• 4B → 4: any evaluation of a state acute/forensic reinvestment program on contemporary constructs (Construct A or B) with a boarding or admit outcome.
• 5B → 5: a multi-state causal 988→ED/arrest diversion estimate surviving an independence audit. ws04 flags this as absent but plausible in future; its arrival would also raise 5A.
• 6D → 5, or 6E → 4: evidence of census-style federal reach into self-funded plans (KC3 explicitly bars 5 without it), or restored NQTL investigator funding plus resolution of the May 2025 non-enforcement posture.
• 7A/7B → down: a comparison-group AOT design separating court leverage from the ACT/ICM package. The current 4 on 7B rests on pre/post in a selected cohort and would not survive a well-identified null.
• 9A/9C → 3: any evaluation of an interstate compact, a supervision-ratio change, or peer-specialist Medicaid billing showing a payer-participation or appointment effect. These are the two cells on this board most likely to be scored differently by a co-scorer, because they rest on the record's instruction about headcount architectures plus channel-level evidence, rather than on an evaluation of the instrument itself. I flag that as my largest disclosed judgment call.

One anchor-fit complaint. O5's anchors reward instruments that are already running, which mechanically advantages the do-nothing comparator (10E = 4) and disadvantages nothing except genuinely novel machinery. I applied the anchors as written rather than adjusting, but the resulting O5 column is nearly constant (nine 4s and one 3 out of ten... in fact seven 4s and three 3s) and carries little discriminating information on this board.