GBMT-1 (Childcare) — Verification Log

Protocol: method/verification-protocol.md, Phase 1 (fact-check). Filing under check: Whitepaper No. 1, filed 2026-08-03 — site/childcare/index.html, site/childcare/sources/index.html, and the research record in childcare/docs/ and childcare/research/. Check date: 2026-08-04. Independence (V1): this session did not author GBMT-1 and gave the filing's own red team no evidentiary weight. Source atlas: primary texts fetched for this pass are committed under method/sources/ with provenance headers.

Coverage (V3)

Coverage proof: 1,029 extracted · 1,029 verdicts · 0 unaddressed. The clusters below account for every row. Two honest caveats on that arithmetic, stated here rather than at the foot where they would be easier to miss: a substantial minority of verdicts are CONFIRMED-INTERNAL (reproducible from the repo, no external referent) or confirmed at secondary tier because 403s blocked the primary source, and ~68 rows are UNVERIFIABLE. An unverifiable verdict is a verdict, not a verification. Every claim was reached and given a disposition; not every claim was reached at the tier V2 asks for.

Extraction guardrail (V3) — the delegated ledger failed its own test

The protocol requires the adjudicating model to independently re-extract one section and diff it against the delegated ledger before trusting it. That guardrail fired.

Sample section: Part 4, "Senate procedure: finance survives, mandates die" (site/childcare/index.html:192-198).

• Independent re-extraction by the adjudicating model, recorded before reading the delegated ledger: 12 substantive assertions (plus 3 source attributions).
• Delegated ledger for the same lines: 4 rows (P025–P028).
• Miss rate: 8 of 12 (67%).

The misses were not random. Every dropped row was a causal claim, statutory characterization, or policy inference:

The delegated pass captured every number and date in the section and dropped the reasoning. That is the exact material V4 calls the INTERPRETATION failure mode — the mode the protocol was written to catch. An extraction layer with this bias does not merely under-count; it removes the protocol's subject matter from the ledger before adjudication starts.

Action taken: the extraction layer was repaired by re-running it at a higher tier against an explicit inferential-claim taxonomy, and re-tested against the same Part 4 guardrail. Consequence for the orchestration design: logged as a correction to method/verification-protocol.md — see the deviations entry.

Cluster A — Reconciliation and the Byrd rule (the protocol's named highest-risk claim)

The protocol singled this out: "its load-bearing procedural analysis (the Byrd rule, reconciliation) is the kind of technical legal claim that fails quietly." It did fail — in a more specific way than expected, and in one respect the filing was closer to right than the secondary record suggests.

Filing's sourcing tier. childcare/research/ws11-legislative.md:7 rests this analysis on a Manhattan Institute article and an EPIC for America explainer — both advocacy-affiliated secondary analyses. The same file records "CRS deep-dive pending" and lists "CRS Byrd compilations" under Pending. The claim was published on the public whitepaper as "the controlling precedent" while the record itself marked its verification as outstanding. That gap is the finding behind all of the verdicts below.

Primary sources reached this pass: 2 U.S.C. § 644 (statute); CRS RL30862, The Budget Reconciliation Process: The Senate's "Byrd Rule" (Sept. 28, 2022), including its table of Byrd rule points of order; CRS R48640, The Senate's Byrd Rule: FAQ (Aug. 21, 2025). All committed to method/sources/.

What the primary record actually shows

CRS RL30862's table of Byrd rule points of order, entry 22 — American Rescue Plan Act of 2021 (P.L. 117-2), records:

• "To Strike Provision(s) from Bill or Conference Report: [none]"
• "To Bar Consideration of Amendment(s): Sanders Amendment No. 972 — basis: "Budgetary changes merely incidental to non-budgetary components" — subject: "To provide for increases in the Federal minimum wage" — waiver motion Rejected, 42-58 (March 5, 2021) — disposition: "Sustained; amendment fell."

Two distinct events, which the filing merges into one:

1. Feb. 25, 2021 — the Parliamentarian advised that the House-passed minimum-wage provision did not comply; Senators removed it from the substitute before floor consideration. Per CRS R48640: "When Senators respond to guidance from the Senate Parliamentarian by choosing to remove or redraft questionable matter in a reconciliation measure prior to its consideration on the Senate floor, then no formal precedent is created, and information that may guide future interpretations of the rule is not published."
2. March 5, 2021 — Sanders Amendment No. 972 sought to add it back; a Byrd point of order was raised on merely-incidental grounds, the waiver failed 42-58 (60 required), and the point of order was sustained.

So nothing was struck from the Act; an amendment to add it was barred. But a Byrd determination on merely-incidental grounds was sustained by the Chair and is recorded in CRS's compilation — which is a stronger footing than the pre-floor-removal story the secondary sources tell.

Verdicts

Replacement wording (public whitepaper, Part 4)

Reconciliation's Byrd rule declares six categories of provision extraneous; the one that governs here strips provisions whose budgetary effects are "merely incidental" to their non-budgetary components. The closest on-point determination is the $15 minimum wage: the Parliamentarian advised in February 2021 that it did not qualify, it was dropped from the American Rescue Plan before floor consideration, and when it was offered again as an amendment that March a point of order was sustained against it on merely-incidental grounds, the waiver failing 42-58. A childcare wage mandate faces the same objection; wage funding is budgetary and survives that test — though CRS cautions every such determination is case-specific, and money-moving provisions can still fall to the rule's outyear-deficit test, which is why reconciliation-built programs get written with sunsets.

What this does and does not change

It does not overturn the filing's operative conclusion. "Fund compensation rather than mandate it, and house standards inside federal spending programs" survives — the primary record supports the direction. What falls is the confidence and the mechanism: a case-specific judgment was published as a bright-line rule ("dies identically," "controlling precedent"), and a compressed generalization ("anything that moves money survives") contradicts both CRS and the filing's own sunset finding. Per the protocol, this is applied to the record, not footnoted.

Cluster B — bounded numeric and citation claims

Delegated to the low-risk adjudication layer per the efficiency architecture, with primary-source verification required and search-summary figures barred. 22 claims checked: 12 CONFIRMED · 3 CONFIRMED-with-caveat · 4 CORRECTED · 2 STALE · 1 negative positively verified.

Cluster C — the committed FTE model (adjudicated directly)

The whitepaper's headline finding — the workforce is the binding constraint — rests on a model in this repo rather than on an external source, so it was re-run rather than looked up.

childcare/baseline/scripts/fte_model.py reproduces every published figure exactly:

This is the committed-pipeline discipline working as advertised: the arithmetic is reproducible from the repo by a checker who did not write it. Two qualifications:

Model caveats already disclosed by the filing and confirmed as disclosed: representative rather than state-by-state licensing ratios (deviation #4), and a CURRENT_WORKFORCE constant of 1.05M carried as a parameter rather than a live pull.

Cluster D — the scorecard (the pass's most serious finding)

The scorecard produces the filing's central architecture recommendation, and the series method (M6) requires "anchored ordinal scales authored first; every cell cites a workstream finding; rankings under multiple objective weightings with rank-stability reported; independent re-score with a reconciliation log." Audited cell by cell, then re-run.

What reproduces. rank.py regenerates rankings.txt exactly (diff empty). The matrix really is 14 dimensions × 4 weightings. The reconciliation log's shuffled sample re-score is real and its two applied adjustments are recorded. That part of the method was delivered.

S5 — rank sensitivity under Phase 1's corrections

The corrected ws11 finding leaves one comparative standing (A8, CONFIRMED): standards housed inside federal spending programs are the most Byrd-defensible form. That bears on procedural for the two architectures that do exactly that — a3 (scored 3) and a8 (scored 2). Re-scoring both to 4:

The top-4 set does not move under any weighting — the filing's recommendation survives. What moves is a3's primacy: a8 ties a3 under child_dev_first (3.95 each) and leads under labor_supply_first (3.95 vs 3.90). scores.csv is left at committed values — the re-score is a reported sensitivity, not a silent restatement.

Superseded 2026-08-11. This sensitivity was computed against the board as it stood in Phase 1, before deviation #10's structurally-blinded re-score of 2026-08-07 moved 47 cells. It is left here as the dated Phase 1 record, but it no longer describes the committed board: a8 now leads all four weightings outright, so there is no a3 primacy for the procedural re-score to erode. The full table this paragraph cited to scorecard/rationale.md was replaced when that file was rewritten for the re-score; the current sensitivity of record is scorecard/rank-sensitivity.py. See deviation #15.

Why D1 is the pass's most serious finding. The Byrd corrections changed how confidently one sentence could be stated. D1 changes what the scorecard is: not an evidence-linked instrument but a single analyst's judgement applied against scales they wrote, with ~5% of cells tagged. That is a legitimate method — but it is the method the filing's own red team, Attack 5, warned about in its own words ("if the scores encode the priors, stability across weightings measures internal consistency, not truth"), and the record claimed a stronger one. The honesty box already says "one analyst, one desk" and flags the missing second scorer; it does not say the cells are largely uncited, and scales.md says the opposite.

Cluster E — the meta-record (274 rows, all adjudicated)

protocol-review.md (97) · research-inquiry.md (21) · adjudication-criteria.md (22) · deviations-log.md (27) · red-team.md (43) · steelman-feasibility.md (33) · incompatibility-log.md (32). 274 of 274 adjudicated, zero unaddressed in this batch.

Tally: ~178 CONFIRMED-INTERNAL · ~23 CONFIRMED external · ~68 UNVERIFIABLE · 2 OVERSTATED/CORRECTED · 3 transcription slips.

Coverage gap noted against this pass itself: deviations-log entries #10–#16 — this verification pass's own self-documentation — post-date the extraction ledgers and appear in none of them. Recorded here so the coverage arithmetic below is not quietly wrong.

Cluster F — workforce, cost and baseline (171 rows)

~110 CONFIRMED (including CONFIRMED-INTERNAL, reproduced from committed data and scripts) · 5 CORRECTED findings across ~25 rows · 3 OVERSTATED across ~5 · 2 UNVERIFIABLE across ~7.

Cluster G — precedents, quality and politics (74 rows)

51 CONFIRMED · 9 CORRECTED · 9 OVERSTATED · 2 UNVERIFIABLE · 1 REFER-UP.

Cluster H — market response, facilities, demand, private capital (82 rows)

56 CONFIRMED · 5 OVERSTATED · 2 UNVERIFIABLE · 3 citation errors · plus carried-over corrections not re-litigated. Several claims were pushed a tier deeper than the filing's own citations — SEC 10-Q/10-K rather than trade press for the chains, GAO rather than TPC/CRS for §45F.

Cluster I — delivery, federalism, civil society, architecture, sequencing (122 rows)

~96 CONFIRMED · 3 CORRECTED · 1 OVERSTATED · 1 UNVERIFIABLE (carried from A12) · 1 REFER-UP.

Confirmed at primary tier and worth naming, because most of this cluster held: CCDBG's FY2015–FY2020 authorization and continued lapse (P.L. 113-186 text); the Unifying Framework's exact 15-organization signatory list including all four named unions/associations (original PDF); NHSA's endorsement of BBB is genuine (live NHSA page — upgrading G9's UNVERIFIABLE); and BBB's House passage 220–213 with Senate death on whole-bill grounds unrelated to childcare (Manchin statement, Dec. 19, 2021).

Pattern across this pass

Three of the four corrections above share one shape: the research record was more careful than the public page. ws11 marked its CRS check pending while the whitepaper called the claim "the controlling precedent"; ws04 adjudicated H4.1 indeterminate while the sources page stated it as a finding; the sources page gave the Byrd ground correctly as "merely incidental" while the whitepaper rendered it "a private-sector mandate."

The filings are not being researched carelessly. They are losing their hedges in translation to the public page — and the public page is what gets read and cited. That is a publication-stage defect, and it will recur in the other five filings unless the translation step is checked as its own layer rather than assumed faithful. Recommended for the remaining filings: diff each public claim against its record sentence and treat any lost qualifier as a finding in its own right.

Status of this pass

Coverage arithmetic (V3), stated honestly:

| Unaddressed | 0 |

The coverage proof now closes: 1,029 extracted · 1,029 verdicts · zero unaddressed. Six adjudication batches ran after the first commit of this log, covering the long tail that the earlier draft correctly reported as outstanding. Two caveats on that arithmetic, stated rather than buried: (1) a substantial minority of verdicts are CONFIRMED-INTERNAL (reproducible from the repo, no external referent) or CONFIRMED-at-secondary-tier where 403s blocked the primary source — those are counted as adjudicated, not as primary-verified; (2) roughly 68 rows in the meta-record are UNVERIFIABLE, and an unverifiable verdict is a verdict, not a verification. The honest one-line summary is: every claim was reached and given a disposition; not every claim was reached at the tier V2 asks for.

What is complete:

• V1 independence — checker did not author the filing; prior red team given no weight.
• V3 extraction — three ledgers committed, guardrail run, failed, layer repaired, re-tested, passed.
• V2 source hierarchy on Cluster A — pushed from advocacy secondary sources to statute and CRS.
• V5 — two negatives verified positively against actual text rather than summaries: CRS RL30862 contains no minimum-wage entry (consistent with R48640's statement that pre-floor removals are not published), and the enacted P.L. 117-2 contains no minimum-wage provision at all (full-text search of the official govinfo PDF — see A11). The remaining named negative (no peer-reviewed ARPA evaluation) sits with the unreturned Cluster B.
• V6 dual as-of verdicts — Cluster A verdicts hold identically as of the 2026-08-03 filing date and as of the 2026-08-04 check date; none of the corrections is a STALE-type supersession. The filing was wrong on the day it shipped, not overtaken by events.
• Corrections landed in dependency order: record → both public pages, with visible dated correction notes.

What remains for GBMT-1 Phase 1: 2. The ~988 unadjudicated ledger rows — the bulk are low-risk numerics suited to the cheap adjudication layer, but they are unadjudicated, not presumed fine. 3. The scorecard cell bases (research/scorecard/) — extracted into the record ledger, not yet checked against their cited workstream findings. 4. Anchor rows 4, 7, 9, 11, 16, 17, 18, 20, still blank in the filing by rule, unexamined here.

Phase 2 (steelman) has not run and per S1 must run in a different session from this one.