[MaxGhenis]

What this does

It moves PolicyBench's frozen US references from policyengine-us 2.41.0 to 2.60.0, the newest PyPI release when the sweep began (01:45:50 UTC on 2026-10-23). It follows the method and records of reference_audit/2026-10-12-engine-upgrade/ (Max, 2026-09-28: "we should be using the latest pe for this always!").

The build is staged for the next board release in reference_audit/2026-10-22-engine-upgrade/reference-build/. That directory holds the regenerated reference_outputs.csv, the exclusion record, and the sidecar with its fourth engine_upgrade revision. The published board (release 20261014) and its snapshot stay on 2.41.0 until the next board release installs the build, re-scores and re-freezes. The README gives the install steps.

What changes (1,984 outputs)

Output	2.41.0	2.60.0	Why
scenario_031 (CA) SNAP	$0	$288	policyengine-us#10036 removes the net income test from California's MCE (CDSS ACL 12-62, 13-32, 14-56, 15-42). This one-person elderly household is categorically eligible, and its $0 calculated benefit becomes the $24/month minimum (7 U.S.C. 2017(a)). A blind Opus reference-adversary judge also ruled the $0 reference_wrong.
scenario_028 (PA) free school meals	0	1	policyengine-us#10238 models Pennsylvania's universal free breakfast (Act 33 of 2023, 24 P.S. § 13-1337.1(c.2)). Lunch stays paid.
scenario_117 (AR) free school meals	0	1	#10238, Arkansas Act 123 of 2025 (A.C.A. § 6-18-722(f)), from SY 2025-26.
scenario_123 (PA) free school meals	0	1	#10238, Pennsylvania Act 33 of 2023.

The output definition is "whether PolicyEngine returns positive annual free school meal support". Both breakfast statutes are standing law enacted before the 2026-07-03 freeze. Neither is unconditional: Pennsylvania reimburses participating schools to the extent funds are available, and Arkansas gives the breakfast on request. #10238 assumes a student at a covered school and models neither condition, and the bases and narratives say so.

The record still has 49 exclusions, so each model is still scored on 1,935 outputs.

Exclusions: none regenerated, 16 rechecked

Departure from the brief: the four r24 outputs stay excluded.

scenario_002, 107 and 121 federal income tax and scenario_121 state income tax land exactly on their records' alternative values: $1,890.30, $211.23, $875.34 and $212.23 (#10185, released in 2.50.0).

Those records are not engine defects. They are reference_depends_on_unlisted_input: the prompt lists only "disability benefits" and does not say who paid for the coverage, and that decides taxability (26 U.S.C. 105(a); 104(a)(3)). reference_audit/2026-09-22/root_causes.json classes r24 as unlisted_input. #10185 reaches the employer-paid reading through a new input, tax_exempt_disability_benefits, that defaults to zero.

The builder regenerates only engine-defect records. The 2026-10-12 upgrade kept scenario_080, 052 and 081 excluded on the same ground: an engine release does not settle facts the prompt leaves unstated. So the four keep their decided $0 and are rechecked.

Scoring them on the engine's default is a methodology call, queued for Max.

The other twelve rechecks:

• scenario_118 SNAP moves from $2,868 to $2,400. #10158 models New York's SSI state supplement ($87 a month), which counts as SNAP income. The record's unlisted input (whether the mortgage is on the home) still decides it.
• Eleven others keep 2.41.0's values. They are rechecked because their decided values differ from the engine's, as in every upgrade.

Engine releases

Each named PR (#10036, #10158, #10185, #10238) is an ancestor of 2.60.0's release commit, 2bc4ccc970.

A first sweep on 2.59.0 (19:49 UTC on 2026-10-22) was cut off by a host restart. When the session resumed, 2.60.0 had been published and moved the three school-meal flags (#10238 is not in 2.59.0). So the sweep began again on 2.60.0; verification/superseded_2_59_0.json keeps the record.

Publication check: 2.60.0 was still PyPI's newest release at 02:11 UTC (verification/sweep_timing.json). The next release reruns the check before it publishes.

Invariants (stated, and tested in tests/test_build_references_upgrade_1022.py)

• No silent move. Every output that moves beyond the exact-match tolerance from its base reference is named by exactly one action: approved, regenerated, new exclusion or rechecked. Otherwise the builder refuses. A Hypothesis test checks that the draft places every moved output exactly once.
• Rule 5. An excluded output keeps its decided value. Only an engine-defect record that lands within $1 of its audited target can be regenerated, and an unlisted-input record never is.
• Byte preservation. Every unchanged CSV row and every earlier sidecar revision keeps its bytes.
• Order independence. The records do not depend on the order of the actions' lists.
• Differential reproduction. Re-planning the committed computed.csv with final_actions.json and writing it with the committed provenance reproduces the committed CSV, sidecar, exclusion record and traces byte for byte. No engine is needed.
• Pinned provenance. The sidecar pins the builder, harness and actions by sha256. The build's computed.csv is byte-identical to the sweep's first output and gives every scored reference.
• Reviewed draft. final_actions.json is the drafter's output with figures added. It has the same placements, values and causes.
• Narratives. They are hand-corrected from the traces (no LLM writer call), and each figure is checked against the trace.

Tests

• new suites: 137 passed (tests/test_build_references_upgrade_1022.py, tests/test_sweep_timing_1022.py)
• full suite (pytest -m "not slow") on the locked 2.41.0 environment: 5,452 passed, 58 skipped, 0 failed
• CI
• independent review r1 (GPT-6.1 Sol, at 830c7ae): APPROVE. It agrees the four r24 cells stay excluded. Its one minor finding (overstated breakfast statutes) and one nit (031's January–September figures) are fixed in 87f40d5.
• r2 delta review of the fixes (GPT-6.1 Sol, at 87f40d5): APPROVE. Both findings are resolved, and nothing new is wrong.

The release commit 2bc4ccc970, built from source, gives the wheel's 1,984 outputs byte for byte (verification/release_commit_check.json).

🤖 Generated with Claude Code

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated
policybench-site	Ready	Preview	Oct 23, 2026 2:34am UTC

[MaxGhenis]

Merge audit (head 87f40d5):

• CI: green on this head (lint, test 11m40s, app, Vercel). gh pr checks exits 0; mergeable, state CLEAN.
• Independent review: GPT-6.1 Sol (Subfleet, read-only).
  • r1 at 830c7ae: APPROVE, with one minor and one nit, both fixed in 87f40d5.
  • r2 at 87f40d5: APPROVE.
  • Verdicts: ~/reviews/policybench-engine-upgrade-2026-10-22/review/r1_out.md and r2_out.md.
• Local:
  • full suite on the locked policyengine-us 2.41.0 environment: 5,452 passed, 58 skipped, 0 failed;
  • new suites at head: 137 passed.
• Publication check: PyPI's newest is still 2.60.0 at 02:49 UTC. Nothing published changes: the build is staged for the next board release.
• Open, and not a merge blocker: d1878 asks whether to score the four r24 cells at the engine's employer-paid default. This build keeps them excluded, as the method does.