Certify policyengine-uk 2.130.1 with policyengine-uk-data 1.58.1

[MaxGhenis]

Fixes #616

Certified versions

	Version	Basis
UK data	policyengine-uk-data 1.58.1, enhanced_frs_2024_25 (sha256 056e80fd…ce8563784)	release manifest at HF commit 89be6b73 (manifest sha256 545797bf…)
UK model	policyengine-uk 2.130.1	legacy_compatible_model_package: the publisher's claim in that manifest (==2.122.2, ==2.130.1)
Core	policyengine-core 3.33.0	unchanged; the claim also adds ==3.33.0
US	policyengine-us 2.2.1 and its datasets	unchanged

Before this PR, main served uk-data 1.58.0 with policyengine-uk 2.126.12 (#585).

• uk-data 1.58.1 is the batch Max approved on 12 October (d1360). It contains uk-data Preserve nullable SPM statuses in wrapper summaries and JSON #518, Update PolicyEngine Core to 3.32.10 #540 and Select UK local-authority lookup by dataset identity #541.
• policyengine-uk 2.130.1 is the first PyPI release after 2.127.0.
• The changed fragment makes this a patch release.

Why now: the UK Autumn Budget is on 28 October, with no model freeze (cos d1152), so the latest qualified pair is what runs on the day.

Certification

• The claim. HF commit 89be6b73, parent 2526edbb, changes only releases/1.58.1/release_manifest.json. It appends policyengine-uk ==2.130.1 and policyengine-core ==3.33.0. The UK hub wrote it after reviewing the evidence, and checked that the uploaded file matches the reviewed claim byte for byte. uk-data's own validate_release_manifest accepts the result.
• The command. scripts/bundle.py certify-data --country uk --data-producer populace --manifest-uri hf://model/policyengine/policyengine-uk-data-private@89be6b73…/releases/1.58.1/release_manifest.json --model-version 2.130.1, run in an environment synced from this branch's lockfile. The gate refuses an unclaimed version: 2.127.0 exits 2 with "matches neither the build-time model (2.122.2) nor any publisher compatibility claim".
• Regenerated together:
  • the bundle manifest;
  • both TRO sidecars (bundle.py generate --include-tros with the HF token; bundle.py check --include-tros passes);
  • the three policyengine-uk pins in pyproject.toml;
  • uv.lock (uv 0.12.13, --upgrade-package policyengine-uk only): 12 lines, and its 2.130.1 hashes match PyPI's;
  • the test constants that pin the certified pair.

Evidence

The full evidence is in ~/reviews/w4-cert-2.130.1/EVIDENCE.md, which the UK hub reviewed (~/reviews/uk-hub/reviews/claim-2.130.1/review-evidence.md). Every figure is from full policyengine_uk.Microsimulation runs. The aggregates, the reform scores and the UC counts use the whole 1.58.1 file. The 19 contract population checks use its seeded 1,000-household subsample, so a failure outside that sample could escape them. The data-change comparisons also use the 1.58.0 file and the UK hub's seeded builds. Only aggregates and counts are reported.

• Contract. W1's UK Budget reform contract 1.0.0 holds on 2.130.1 with 1.58.1: 673 pass, 0 fail, 0 error (6 pending on d1344, 79 unsupported). All 19 population checks pass. This uses the runner with Give contract assertions unique ids and declare full-run population coverage #612 and Check the UC allowance response on the award before deductions #614 (see below).

• Finiteness and reforms. Every aggregate is finite in 2024, 2025, 2026 and 2030. A 21% basic rate from 6 April 2026 raises £6.66bn of income tax in 2026-27 (£6.67bn on the current pair).

• Target fit. 78.3% of 644 national targets are within 10%, against 73.6% for the current pair. This is the policyengine-model UK audit session's measurement, using uk-data 1.58.1's own create_target_matrix.

• Most of the move is the data. In 2025, household net income falls £10.60bn: −£10.36bn from the data and −£0.25bn from the model. The data move is the release's code effect plus unseeded rebuild variation in each published build. The UK hub's seeded comparison puts the code effect at −£5.57bn, and the rest is noise in the two builds (policyengine-uk-data#507).

• The model moves are each attributed to a merge, by running every first-parent merge from 2.127.0 to 2.130.1 and splitting rollup #2380 into its members:

  • #2365: UC −0.27% (2024).
  • #2366: gov_spending +£0.54bn.
  • #2271: income tax +£0.34bn. 1.58.1 holds 2,251 Gift Aid records, so #2271's "no Gift Aid" held only for the 2023-24 file it measured.
  • #2119: ESA −1.26%, HB −0.3%.
  • #2047: senior relative poverty −0.14pp.
  • #2031: JSA £2.0m → £1.4m.

  2.126.12 and 2.127.0 agree to 1.6e-8 relative, and the released 2.130.1 wheel equals its source build.

Universal Credit and the contract's allowance check

• The old check failed on 2.130.1. It required that no universal_credit payment fall when every UC standard allowance rises 10%. On 2.130.1, one unit in the subsample fails: a capped award whose deductions rise.

• 2.126.12's pass was an artefact. 2.126.12 applies no UC deductions below 1e6 benefit-unit weight, so it ran the subsample with deductions off; #1905 removed that gate.

• The law allows a payment to fall this way (this establishes that the model's behaviour is lawful in kind, not that the observed unit's deduction amounts are right; see #2410):

  • the cap counts UC before deductions and holds a capped award flat (SI 2013/376 regs 80(2), 81);
  • deductions are then taken from the award (SI 2013/380 Sch 6);
  • the UK hub's independent legal check returned AGREE_WITH_CHANGES.

• Check the UC allowance response on the award before deductions #614 restates the check:

  • no award before deductions falls by more than a penny;
  • each payment moves by its award's change less its deductions' change, to the penny.

  Both versions hold on the full file (0 award falls, 0 accounting breaks). How much deductions rise belongs to policyengine-uk#2410.

Known gaps, unchanged by this pair

• Northern Ireland: NI households have no constituency or local-authority codes (policyengine-uk-data#580), so those runs raise "no households".
• Engine skew: the data is calibrated with policyengine-uk 2.122.2 and served with a later model (policyengine-uk-data#581).
• git_sha: the manifest's build.built_with_model_package.git_sha records uk-data's commit, not the model's.

Order and gates

This PR's contract evidence uses #612 and #614 (W1's runner: assertion ids, and the UC check on the award). Neither changes what this PR certifies. #614 holds for its own independent review.

Independent review (GPT-6.1 Sol): APPROVE (~/reviews/w4-budget-ci-2026-10-11/pr617-round1-sol.md).

• It reproduced the certification scope, pins, lock, hashes and append-only claim, and that the gate accepts 2.130.1 and refuses 2.127.0.
• Regeneration reproduces the manifest, pins and TROs byte for byte.
• Every numeric claim reproduces from the raw evidence, and an independent full-file UC rerun matches exactly.
• 149 targeted tests passed.
• Its three low documentation findings are fixed above.

Full suite on 480644f, with the HF token: 1 failed, 1684 passed, 9 skipped. The one failure was the projection-horizon pin, fixed in 36b4af0 (policyengine-uk 2.130.0 projects to 2032), which now passes.

axiom: n/a: data-release certification; no policy encoding changes

🤖 Generated with Claude Code

[MaxGhenis]

Merged at 36b4af0 on the standard gates. This certifies policyengine-uk 2.130.1 with policyengine-uk-data 1.58.1 as the UK default pair, on the publisher's claim at HF 89be6b73.

• Checks: all 12 passed on the head.
• Independent review (GPT-6.1 Sol): APPROVE. Its three low documentation findings are fixed in the body.
• Evidence: reviewed by the UK hub (~/reviews/uk-hub/reviews/claim-2.130.1/review-evidence.md).
• Data release: Max approved it as d1360.

Next: the release, sim-api's update, and the production check.

[MaxGhenis]

Traceability note. The qualification evidence for this pair ran W1's contract with the restated UC allowance check from #614, which is not yet on main.

• The difference: main's population_uc_standard_allowance_plus_10 still requires that no UC payment fall when the standard allowance rises. On policyengine-uk 2.130.1, one benefit unit in the seeded subsample fails that check. It is a capped award whose deductions rise with the allowance, which is lawful (SI 2013/376 regs 80(2), 81; SI 2013/380 Sch 6). The UK hub's independent legal check agreed.
• Until Check the UC allowance response on the award before deductions #614 merges: a rerun of main's contract on this pair would report that one failure. Nothing on main runs the contract today: it needs the private dataset and is not in PR CI, and the qualification workflow (Qualify UK model and data candidates automatically before promotion #574) has not merged.
• Evidence: the full run with Check the UC allowance response on the award before deductions #614's check holds (673 pass, 0 fail, 0 error; all 19 population checks pass): ~/reviews/w4-cert-2.130.1/contract-report-2.130.1-d1.58.1-final.json.

#614 lands as soon as its round-3 review agrees.