GUBMENTPlain talk · policy frontier
Filings / Mental health / Sources / Deviations Log (GBMT-11)
GBMT-11 · Research record · No. 11

Deviations Log (GBMT-11)

mental-health/research/deviations-log.md
This is a working research document from the mental health filing, published as written — including the parts later corrected. It is the underlying record for Whitepaper No. 11, not a summary of it.

Freeze point: mental-health/docs/research-inquiry.md as of protocol draft 2026-08-09 (commit carrying the GBMT-11 protocol on main). Per M6, each divergence from the frozen protocol is recorded with its effect. Entries are one line each.

# Date Deviation Reason Effect
1 2026-08-11 Protocol freeze recorded at Phase 0 close rather than at execution open (empty log at start). Prior Claude session claimed the lane and hit a session limit before writing freeze + findings; this pass re-opened primary sources and freezes the 2026-08-09 protocol text as the commitment. Freeze is retrospective to protocol draft; no protocol text was altered mid-pass.
2 2026-08-11 Anchor 8 prior (“~1/3 of working-age awards”) replaced by SSA awards share 12.7% (2023 disabled-worker awards). SSA DI ASR Chart 10 / awards-by-diagnosis table contradicts the circulating prior. H8 award-share leg must be rewritten before support; stock % and access–award links handed to §8 / GBMT-12.
3 2026-08-11 Contemporary bed figure carried as state-hospital staffed beds only (TAC 2016: 11.7/100k), not whole-system psych beds. Protocol prior band was definition-sensitive; Phase 0 found no clean all-payer staffed inventory in the gate pass. §3 must not treat 10–15/100k as whole-system supply; NRI/SAMHSA inventory still owed.
4 2026-08-11 H2 acceptance leg rests on one strong NAMCS root (Bishop 2014) plus secondary HRSA citations, not yet two independent methods. Phase 0 gate verified the prior is live; full H2 support needs secret-shopper/claims in §2. H2 remains open (acceptance leg live); KC1 still fires on the missing national by-payer series.
5 2026-08-11 988 “doubled” prior accepted as directionally true via monthly KPI pairs and GAO volume growth, not a single official annual multiplier. No single SAMHSA “2× in two years” headline series; Jan-to-Jan answered contacts ~2.5× (2022→2024). H4 volume leg supported; diversion leg still thin pending §4.
6 2026-08-11 H1 supported without a second survey series for AMI/SMI at >1.5× vs NSDUH. NHIS lacks AMI/SMI; support uses locked AMI vs SMI (~4.2×) within NSDUH plus flat 2021–2024 prevalence. Definitional-sensitivity leg holds; literal cross-survey criterion marked partial in §1.
7 2026-08-11 H2 Method 2 (Brahmbhatt secret-shopper) has no commercial comparator arm. Study limitation; commercial wedge rests on Bishop NAMCS same-frame contrast. H2 still Supported (two methods show Medicaid usable access <50%); disclose commercial gap on Method 2.
8 2026-08-11 Contemporary bed supply locked as two constructs (state-hospital ~10.8/100k; HCRIS whole-hospital ~28.4/100k), not one band. Phase 0 owed whole-system inventory; McBain 2025 HCRIS delivers it with different magnitude than TAC. Scorecard must name construct; Phase 0 deviation 3 narrowed further.
9 2026-08-11 H3 supported on statutory/CMS “cannot purchase” path despite McBain null on waiver-state bed counts. Formal criterion’s OR branch met; comparative waiver evaluations thin. Do not claim §1115 MH waivers have already raised Construct B supply.
10 2026-08-11 Post-2014 national MH spend uses Commonwealth/IHME 2019 Medicaid dollars as supporting, not BHSUA replacement. No newer SAMHSA BHSUA payer table or clean NHEA MH line located. Spending cells cite vintage explicitly.
11 2026-08-11 H4 diversion leg treats mobile-crisis / BHCC ED studies as continuum evidence, not 988 refute. Formal criterion requires multi-state 988 causal diversion; those studies are different instruments. Architecture #5 may still cite continuum steelman in §7.
12 2026-08-11 H5 “material” supply response includes modest Allegheny elasticity (0.16) plus CCBHC PPS access package. Protocol did not define elasticity magnitude; two independent change designs show detectable response. Do not claim fee bumps alone close H2 acceptance gap.
13 2026-08-11 Anchor 10 verified with mandatory May 2025 nonenforcement disclosure. ERIC litigation / Departments statement paused new 2024-rule enforcement. Scorecard must date-stamp “2024 rule teeth.”
14 2026-08-11 KC3 fires even though EBSA CAA reviews reach some ERISA plans. Protocol KC3 targets exams generating most compliance evidence; state exams miss self-funded majority. O4 commercial = band-only / unknown with federal spot-check caveat.
15 2026-08-11 H7 international leg uses qualitative OECD bed dispersion + peer cases, not a fitted outpatient-controlled boarding regression. No clean multi-country joint bed/outpatient/boarding series located this pass. Criterion (a) met directionally; disclose non-modelled.
16 2026-08-11 H8 adjudicated Not supported as written; awards plurality path dead at 12.7%; overflow-valve rewritten as access→applications pathway (Swenson). Phase 0 deviation 2 + SSA Chart 10; scarcity→awards fails direction test. Seam language: 12.7% awards / ~28.6% stock (GBMT-12); no one-third-of-awards.
17 2026-08-11 Protocol §3 qualitative weight emphases committed as numeric vectors summing to 1.0 (W1 0.30/0.30/0.20/0.15/0.05; W2 0.35/0.05/0.15/0.35/0.10; W3 0.10/0.20/0.30/0.10/0.30). M6 requires explicit weightings for rank-stability; protocol named emphasis only. Rankings in ws09-scorecard.md are reproducible; vector choice is a logged judgment.
18 2026-08-11 Scorecard uses symmetric evidence floor (below-neutral requires cited failure; absence → 3), matching elder-care pass-3 / housing-media-drugs correction — not an asymmetric “no support → 2” reading. M6 childcare failure mode + HANDOFF scale-asymmetry warning; protocol did not freeze band text. Unevidenced rows (#8/#9) float at 3 rather than inventing 1s/2s; do-nothing 2s rest on documented status-quo shortfalls.
19 2026-08-11 Measurement infrastructure (national by-payer wait/acceptance series) not scored as an architecture row among protocol §8 list 1–10. Red-team Attack 7: KC1 makes measurement the headline shape; VA standards are the precedent (ws07); inventing a full #11 mid-pass without an evaluation package would invent cells. Honesty box names missing row; whitepaper / next pass must score it or record explicit exclusion.
20 2026-08-11 Red-team cell moves: #2 O1 5→4, O2 4→3; #6 O1 4→3. Attacks 1 and 5: Mathematica wait metrics match band 4 not 5 (DY4 softening; clinic-reported); required crisis scope + heterogeneous ED DID ≠ O2=4; NQTL corrections ≠ §2 realized-access proxy for O1. “CCBHC leads all three” withdrawn — #5 leads W1; #2 leads W2/W3. #6 W2 swing narrowed.
21 2026-08-11 Structurally blinded re-score reconciled: 7 cells moved (#4 O2 4→3, O5 3→4; #6 O5 2→3; #7 O1 4→3, O2 3→4; #8 O5 4→3; #10 O1 2→3); #3 O2 kept at 4 vs blind 3; #9 O1/O3 kept at 3 as judgment splits vs blind 2s. M6/M10 batch msgbatch_013vUDQ6mt6JV8mz2H95zfsx; evidence rooted in ws02–ws07 (need≠instrument; O1≠hospitalization OR; O1 OR includes crisis path; O5 wash on EBSA/CoCM). Full table in ws09-rescore-log.md. #5/#2 leadership and #10 pole hold; AOT W2 near-top withdrawn; #3≡#7 and #8≡#9 numerically; scorecard → pass 2.
← All Mental health research documents Sources digest Read the whitepaper