Correct the amterr lab's layer-3 claims and make it rerunnable

[MaxGhenis]

Corrects the error-case lab at paper/snapshot/labs/amterr and makes it rerunnable. Context: TheAxiomFoundation/axiom.org#297 rewrote the Colorado SNAP QC report after recomputing this lab; this PR fixes the lab itself. Each finding below was recomputed from the committed replay and the FY2024 QC file before the text changed.

What was wrong, and what the text says now

July text	Recomputed
"10 of those 37 … $3.28M/yr = 3.3% of replayed error dollars"	10 cases, $4.47M (4.5%). $3.28M does not reproduce (HWGT × AMTERR gives $4.47M; HWGT × |engine − RAWBEN| gives $2.91M).
"on the facts the agency recorded, the verified engine returns the reviewer-certified correct benefit"	The engine equals FSBEN in 7 of the 10. All 7 are solver no_change rows, where the replayed input is the file's own input. FSBEN is Mathematica's calculated benefit (tech doc PDF pp. 75, 91).
Example case 202312-40441	Carries cause 15 only, so it is outside the 10.
All 10 framed as computation	7 of the 10 broad-coded findings concern the computation (4 × 520/80, 520/75, 520/98, 362/52). The others are 331/44/17, 350/44/17, 161/6/10. The 7 carry codes 10 (5 cases) and 20 (2); none carries 17 or 19.
"no single-variable original value + correct math reproduces the issuance"; the 37 as an "upper bound on computation-side"	The solver moves only the ELEMENT1 input. It moved nothing in 17 of the 37 (14 no_change, 3 eligible but unmoved) and in 16 of the 246 reproduced cases. The upper-bound reading is withdrawn: of the 26 cases with a layer-2 computational finding, 13 reproduce ($3.6M), 10 do not ($3.5M) and 3 were not replayed.
Broad class ≈ "~1.0 point of the rate"	10.5% of Colorado error dollars (1.05 points of 9.97), of which 5.7% (0.56) reproduces from a single input, 4.0% (0.40) does not, 0.9% (0.09) was not replayed. The 7 candidates are 1.8% (0.18 points; 0.16 on dollars above the $56 threshold).
14 software cases = "automation fed itself the wrong input"; 2 = "computation logic itself wrong"	Of the 14, 10 moved the software-coded element ($1.8M), 2 moved the same unearned-income total through another finding ($1.7M), 2 moved rent ($0.8M). In the 2 that do not reproduce, the software finding is on an element the solver never moved. No software-coded finding is on rent; one mass-change finding is on the medical deduction. The combined "$7.6M ≈ 6.7%" figure is withdrawn.
"$88.8B issuance" nationally	$91.4B (sum HWGT × RAWBEN); $88.8B was sum HWGT × FSBEN.
National codes 20/21 $107.1M; 10/22 $1,210M (14.9%)	$106.9M; $1,188M (14.6%).
FY2025 "being measured now"	FY2025 was published June 24, 2026: Colorado 10.09%.
Weight vintage unstated	May 2026 posting. The August 2026 posting changed only HWGT/FYWGT (and the _OLD pair); case-level results are identical and dollars move slightly (data entry $18.47M → $18.43M). A table gives both.

Reproducibility

• audit_claims.py → claims_audit.json recomputes every figure under both postings. It also regenerates native_decomposition.json and ../phase_a_classification.json from the May posting and checks them against the committed copies (both match). phase_a_classification.json was already committed one directory up; ANALYSIS.md now cites that path.
• Paths: amterr_replay.py and reconstruct_co_fy2024.R take AMTERR_LAB_DIR (default: the script's directory) and SNAP_QC_REPO. No user-specific paths remain. No computation changed.
• README.md gives pins and commands: giannella/snap_qc 741e10bf, axiom-oracles d34aa6fa, rulespec-us b53ce208, axiom-rules-engine de0efdc7, both QC postings by sha256.
• Reruns at those pins (2026-10-03) reproduced co_fy2024_reconstruction.csv, fy2024_reconstruction_national.csv and amterr_replay_results.json byte for byte. The July engine binary was not recorded; the rerun used the 2026-08-06 build of the same commit attested in paper/snapshot/cert/CERT_REPORT.md.

Invariants and tests (tests/test_amterr_lab.py)

Invariants the audit must satisfy for any input:

• reproduced + not reproduced = replayed, in cases and dollars; each cause class splits exactly into reproduced / not reproduced / not replayed;
• computation candidates ⊆ broad-coded misses ⊆ non-reproduced cases, with dollars ordered the same way;
• an unmoved non-reproduced row replays the file's inputs, so the engine equals FSBEN;
• layer-2 classes partition the error cases and their dollars;
• shares lie in [0, 1], are unchanged by rescaling weights, and points = share × 9.97;
• any-presence membership is monotone in the code set;
• case-level results and case counts are identical under both postings.

Tests: committed-file checks of those identities; Hypothesis property tests on the classifier and partitions (synthetic cases); a differential check that ANALYSIS.md quotes the audit's figures; a regeneration test against both raw postings, which skips where the QC files are absent (CI). The tests also compare the sha256 values the audit records with the live files, so editing an audited artifact fails in CI. Locally: 19 pass with the data, 18 pass and 1 skips without it. hypothesis joins the dev extra.

Review

Round 1 of the independent review (Opus 5.5 via Subfleet) requested changes: the "upper bound" sentence still overclaimed, the no-data tests did not lock the audited files' hashes, and four smaller points (FSBEN page citation, the audit's stated moved-input convention, the solver description, the replay script's docstring). Commit 8786e7d addresses each; round 2 is running.

Notes

• analysis/cause_shares.json still cites "ANALYSIS.md:L62-L67" for the $3.28M claim. That line range is the pre-revision file (commit 755a7f3). I left the artifact unchanged because six downstream artifacts pin its hash; FACTS K2 now carries the commit-qualified pointer.
• Two data-gated tests (test_fixed_donor_decomposition, test_uhip_decomposition raw regeneration) fail on this Mac on clean main because the committed artifacts record an older macOS version in environment.platform. Unrelated to this PR; they skip in CI.
• The manuscript's layer-3 paragraph (paper/index.qmd) and FACTS D4/C4 are out of scope here and flagged separately.

axiom: n/a: analysis documentation, audit script and tests; no policy encoding.

🤖 Generated with Claude Code

[MaxGhenis]

Independent review (Opus 5.5 via Subfleet, read-only):

• Round 1 on 93bc46a: REQUEST CHANGES. Two majors (the "upper bound on computation-side" sentence still overclaimed; the no-data tests did not lock the audited files' hashes) and four minors (FSBEN page citation, the audit's stated moved-input convention, solver description, replay docstring).
• Round 2 on 8786e7d: APPROVE. All six findings and the nits resolved; no blocking findings. Two optional precision suggestions were applied verbatim in a112f25.

The reviewer had no shell in either round and checked by reading the code, artifacts, R script and tech doc, with hand arithmetic from the replay weights. Tests were run by the author: 19 pass with the QC data; 18 pass and 1 skips without it; audit_claims.py --check is current. Out-of-scope follow-up the reviewer flagged (paper/FACTS.md D4, paper/index.qmd, and the classification string in analysis/cause_shares.*) is tracked separately.

[MaxGhenis]

Merged (squash, c4333b3) at reviewed head a112f25. Gates at that head: test check passed (gh pr checks exit 0), MERGEABLE, not a draft, no CHANGES_REQUESTED. Independent review (Opus 5.5 via Subfleet; reports archived at ~/reviews/snap-qc-sim-pr92-2026-10-03/ on Max's Mac): round 1 REQUEST CHANGES on 93bc46a; round 2 APPROVE on 8786e7d; delta confirmation APPROVE on a112f25 (that commit applies round 2's two optional suggestions verbatim; the author confirmed the 8786e7d..a112f25 diff contains only those two edits). Follow-ups: manuscript, README and simulator copies (task_99975d8c), the cause_shares upper-bound labels (task_4d6d9ce4), local macOS-version test failures (task_095ddfeb).