[MaxGhenis]

Summary

Bumps the bundle to policyengine-uk 2.90.2 + policyengine-core 3.30.1, and — per the UC caseload postmortem (policyengine-uk-data discussion #464) — reverts the UK certified default from Microcosm's populace_uk_2023 to policyengine-uk-data's enhanced_frs_2024_25 (release 1.56.16).

Why the revert

Measured at 2025 (annual basis, weighted benefit units with universal_credit > 0):

artifact	UC benefit units	UC spend	deducting benefit units
populace_uk_2023 @ dd68c73 (what this PR previously certified)	3.26m	£36.9bn	—
enhanced_frs_2024_25 @ 1.56.16 (this PR)	6.35m	£75.3bn	3.02m / £1.86bn / 47.6%
administrative / OBR	~6.7m	£75.8bn	~3.1m / ~£2.0bn / 46–47%

populace_uk_2023 won its June certification rotation on a 148-target surface containing no benefit targets (PolicyEngine/microcosm#701); the certified default should not ship the least accurate artifact on the second-largest working-age benefit.

Re-flip condition (coverage floor): populace_uk returns as the certified default when the microcosm UC chain (PolicyEngine/chronicle#184, PolicyEngine/microcosm#727, PolicyEngine/microcosm#729) lands, the rebuild passes a benefits-inclusive surface, and the rotation is scored with UC caseload and spend within 10% of the OBR/DWP counterparts. Proposal: make that floor a standing certification requirement, so no future default can be selected on a surface missing a major program.

Certification provenance

• Publisher compatibility claim policyengine-uk==2.90.2 appended to the 1.56.16 release manifest (HF commit 2966541f), with the measurement evidence recorded in the commit message and policyengine-uk-data#452.
• Certification basis: compatible_model_packages (the data was built with 2.89.2 — disclosed in the bundle certification block; per the certifier's own warning, the test suite is the arbiter, and the full battery runs in this PR).
• Manifest URI pinned at the claim commit; all seven datasets (enhanced/base FRS 2024-25 + tiny variants + LA/constituency/local-geography weights) carry sha256s and are verified under the strict runtime contract. Note: data_producer: "populace" in the bundle block names that verification contract (mandatory per-dataset sha256), not the producing repo — the legacy certification path is an unimplemented stub (certification.py), a follow-up candidate.
• populace-uk-2023-dd68c73 remains published on HF for explicit-URI consumers; it is no longer the certified default and needs no publisher claim.

Also in this PR (unchanged from the original bump)

• policyengine-uk 2.89.2 → 2.90.2 (brings the UC deductions module, policyengine-uk#1815), policyengine-core → 3.30.1.
• TRO sidecars and the household-calculator snapshot regenerated against the new default dataset.

🤖 Generated with Claude Code

[MaxGhenis]

CI failure is the certification gate working as designed: the UK data release (populace-uk-2023-dd68c73-4aa4b14-20260619T023711Z) declares compatible_model_packages: [policyengine-uk==2.89.2] only, so certifying for 2.90.2 fails on both bases (not built-with, no publisher claim). Everything else here is done — pins, lockfile, TRO sidecars, version-surface snapshot, and the full suite passes locally on the pinned pair (783 tests), which per the certification doctrine is the arbiter once a claim exists.

Remaining sequence:

1. Publisher side (microcosm-data tooling): append a policyengine-uk==2.90.2 compatibility claim to the dd68c73 release manifest — the same pattern as the earlier release's post-hoc ==2.88.20 claim.
2. Then on this branch: uv run python scripts/certify_data_release.py --country uk --manifest-uri "hf://dataset/policyengine/populace-uk-private@<new-revision>/releases/.../release_manifest.json" --model-version 2.90.2 (in the pre-bump environment, per the script's ordering), commit the certification block, and CI goes green.

Ordering note for future bumps: certify before moving pins — the import-time guard makes the post-bump environment unable to run the certifier.

[MaxGhenis]

Data point for the certification decision here (not a hold request — sequencing is yours):

The release this PR certifies (populace-uk-2023-dd68c73) measures 3.26m UC benefit units / £36.9bn UC spend at 2025 against ~6.7m households (DWP, 2025 average) and £75.8bn (OBR) — its calibration surface does not bind UC. Full diagnosis and cross-artifact table: PolicyEngine/microcosm#701 and policyengine-uk-data#452 (diagnosis comments). By comparison, policyengine-uk-data's enhanced FRS at v1.56.14 measures 6.40m / £77.6bn.

Implication: UC-touching aggregates through this bundle (including the new deductions module's counts and costs) run at less than half of administrative levels, while per-household statistics remain valid. Options are to certify as-is and note the caveat (unblocks the WPI beat; per-household UC deductions results are unaffected), or hold for a data-side fix per microcosm#701. Either way the caveat belongs in the bundle notes.