OBR policy costings (mode 2): Policy Measures Database + EFO 3.17 → PE-UK static counterparts

Max's steer (2026-08-16): center PE model development — including PolicyEngine Macro — on the Scorecard. Use it to compute PE counterparts to OBR's published scores first, before adding machinery nothing else uses. This issue is the static half; the macro half is the companion issue linked below.

The population

The 8/2 harvest staged the full OBR Policy Measures Database (~/scorecard-harvest/uk_obr/; access recipe and staging decisions in its NOTES.md): 24,304 rows = every measure × tax head × FY, Budget 1970 → Autumn Budget 2025, certified costings, positive_gain_to_exchequer verbatim; plus EFO March 2026 Table 3.17 (95 rows): OBR's current re-estimated yields of the personal-tax measures (PA/HRT freezes, additional-rate cut, NICs thresholds) — each a one-line policyengine-uk reform.

None of it is in the DB. #48/#51 cover mode 1 (levels/baseline moments; OBR enters as 259 EFO welfare rows). This lane is mode 2 (#4): reform deltas — OBR's certified costing of a measure vs PE-UK's static costing of the same measure.

Precedent to extend, not reinvent

sources/campaign-20260802/uk/obr_measures.jsonl — 10 rows (SB2024 HICBC, employee NICs 2p/4p): reversal legs on the certified 2026 world, with annotations naming the axes — OBR = announcement-baseline vintage, behavioural-adjusted, FY accrual; PE = certified world, static, CY. Those axes are the assumptions-registry entries (#10) and the explained-share material.

Design

• Measure → reform registry (e.g. data/uk/obr_measure_reforms.yaml): per measure — PMD row key, fiscal event, verbatim OBR description, PE reform spec (parametric dict + start FY), computability ∈ {expressible, partial (annotate what's missing), not_expressible (why)}, construction (reversal-on-certified-world, or forward from the announcement baseline when the baseline registry (#13) has it), notes. Human-readable provenance on every row; the registry is itself a Scorecard artifact.
• Compute: policyengine.py (latest) managed sims on the certified populace-uk bundle, one reform per measure, per FY in the OBR window; emit GBP aggregates by head where PE can split (income tax / NICs / benefits) so head-level and total rows both join. Every pe_value traces to a saved per-run artifact.
• Stage under the campaign ingest contract (results/uk/staged/*.jsonl: external_claim_match descriptor, engine_version, full bundle id, status/construction/run_id/annotations). DB ingest of the PMD claims themselves rides #48's vocabulary (gbp, fy, basis) once it merges — stage first, ingest second, exactly as the US campaign did.
• Compare descriptively: ratio bins, never pass/fail; benchmark_class = different_model (HMRC/HMT costing models) unless assumption-matched; baseline mismatch surfaced per #13; behavioural vs static named on every row.

First families (in order)

1. EFO Mar 2026 3.17 — 95 rows, current re-estimates, cleanest joins.
2. PMD 2023–2025 events, income tax + NICs + welfare heads (~1,000 rows before computability filtering): AS2023/SB2024 NICs cuts, HICBC, PA/HRT freeze extensions, AB2024 employer NICs, AB2025 measures, two-child limit (joins the pentagon).
3. Backfill earlier events as the registry grows.

Gate

Sol-reviewed PR(s), org-side branch (private repo — fork PRs don't run Actions). Descriptive register throughout (#9): no "replication rate", no wins.

Start with sources/campaign-20260802/uk/obr_measures.jsonl and the staged harvest's NOTES.md to understand the existing precedent and data. Read policyengine.py and the proposed data/uk/obr_measure_reforms.yaml structure, then trace how results/uk/staged/*.jsonl feeds ingestion. Done means the first EFO 3.17 family is computed with saved run artifacts, provenance, and descriptive comparisons under the stated campaign contract.