Audit optional employer pass-through in the payroll tax references; propose excluding four outputs

[MaxGhenis]

Evidence only: this PR changes no reference, exclusion or published number. It adds reference_audit/2026-10-05-payroll/, which audits the state part of the US payroll_tax references in release dashboard-data-20260930 and drafts two alternative records for Max's ruling (cos d972).

Finding

The output asks for "mandatory employee state payroll taxes" and says to exclude employer payroll taxes. policyengine-us 2.15.17 counts the largest employee share an employer may deduct. Under primary law (16 of 16 adversarial reviews agree):

• optional employer pass-through: MN Paid Leave, CO FAMLI, MA PFML, NY PFL and DBL. The premium is the employer's; the employer may deduct up to a cap and may pay it instead.
• mandatory employee contribution: CA SDI, PA UC, CT Paid Leave, NJ TDI/FLI.

Four scored references count an optional share:

Output	Frozen	Without the optional share	Exact now	Exact if regenerated
scenario_032 MN	2,346.10	2,218.50	6	33
scenario_043 CO	330.73	312.74	16	26
scenario_081 MA	14,192.66	13,387.65	9	9
scenario_082 NY	8,108.03	7,664.92	4	14

Recommendation: exclude the four outputs

• The frozen references rest on the employer's choice to deduct the maximum, which the prompt never states.
• "Mandatory employee state payroll taxes" has two defensible readings here: the $0 the law requires of the employee whatever the employer does, or the statutory employee share withheld unless the employer elects to pay it. The statutes frame the employer's choice both ways, so the prompt's default for unlisted facts does not settle it. A reference that turns on a definition the prompt never states is excluded (2026-09-22 and 2026-09-28 rule 4), as in r24, r25 and r14.
• If "mandatory" is read strictly, regenerate with proposed_regenerations.json.
• Exclusion: GPT-6 Sol 95.00% to 95.09%, still first; five adjacent swaps; 1,928 to 1,924 scored outputs per model. Regeneration: GPT-6 Sol 94.63%, Claude Opus 5.5 94.39% and first on within-1%; ten adjacent swaps.

What is here

• program_classification.json, verification/law_research.json: each program's law, sources, 2026 rates and review verdicts.
• fixes/payroll_mandatory_scope.py: an output-scope adapter in the form of the Maryland one. A sweep of all 1,984 outputs with it moves these four payroll outputs and nothing else; the baseline reproduces every scored reference.
• proposed_exclusions.json, proposed_regenerations.json: the two alternatives, in the format a release installs.
• verification/leaderboard_impact*.{json,csv,log}: both options scored through policybench.cli analyze; the unchanged copy reproduces the published payload. Also scored on top of Audit the state income tax withheld in federal SALT; propose three exclusions #191's SALT proposal.
• verification/model_answers.csv, explanations.json: every model's answer and reasoning category on the twelve outputs with a state part.

Independent review

Eight agents reviewed commit 8714770 in a Claude Code workflow (verification/independent_reviews.json):

• Differential impact. Both options re-scored outside the CLI: 2,438 values match, plus 966 for the Audit the state income tax withheld in federal SALT; propose three exclusions #191 combinations.
• Differential sweep. All 1,984 outputs recomputed with the contributions zeroed as inputs: bit-identical, same four moves.
• Mechanism. Engine path traced; 36 payroll parts hand-computed. It corrected three wording errors.
• Recommendation. Advocates for keep and for regenerate, then a judge. The judge upheld exclusion on rule 4 and the r24, r25 and r14 precedents, and struck the draft's board-effect arguments, which no rule uses.
• Release checklist. Each option installed on scratch clones and the tests run; the full list is in the README.
• Claims. 131 claims checked; five errors and thirteen imprecisions found, all fixed in 088929c.

Side findings: #191's README rounds GPT-6 Sol's 95.6146% to 95.62%, and its scenario_081 record treats the optional MA share as paid (both chipped).

Invariants

• Baseline reproduction: every scored reference recomputes to the published value (sweep_payroll_scope.log); the unchanged scoring copy equals the published payload's modelStats, programStats, heatmap, globalWeights and failureModes.
• Scope: the adapter changes only the list behind employee_state_payroll_tax; across all 1,984 outputs it moves exactly the outputs carrying a program it drops, by exactly that program's amount (propose_changes.py refuses any other move).
• Consistency: the adapter's program list equals the reviewed optional set in program_classification.json (checked by propose_changes.py).
• Idempotence: CountryTaxBenefitSystem applies a reform twice; the adapter's second pass is a no-op.

axiom: n/a: PolicyBench reference audit; no PolicyEngine policy change

🤖 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 5, 2026 4:55pm UTC

[MaxGhenis]

Shipped in release dashboard-data-20261006 (PR #202, squash-merged as 9ce4ade8, now the latest data release and live on policybench.org). This PR's commits were folded into that release as a merge commit, so its content is on main; closing it rather than merging it a second time. Ruling d972: payroll_tax for scenario_032, 043, 081 and 082 is excluded (an optional employer pass-through of a state paid-leave premium). Following the release's independent review, Massachusetts's limits are stated as the Department's published 2026 rates, with the competing reading of Acts 2026 c. 101 noted.