UK release input-coverage contract: 145-required gate, effective-mass semantics, and the real-donor HMRC replay

[MaxGhenis]

What this is

The UK arm of the coverage-contract program: every input column the UK national build exports is now under an enforced release contract — present with weighted signal, or carrying an evidence-backed reviewed exclusion — plus the first real-donor HMRC replay of the SPI channel, run under the UKDS-licensed 2022–23 donor.

The contract

• 145 required columns / 0 reviewed exclusions (final state after the two charitable columns' exclusions went stale — the weighted-support wiring gave gift_aid and charitable_investment_gifts real signal above the 1 ppm floor, so the gate itself demanded their promotion to hard requirements).
• Effective-mass semantics: coverage is weighted support (1 ppm person-mass floor), not raw nonzero counts — a column carried only by zero-weight rows is not covered. Per-column and per-family effective-mass diagnostics ship in the gate report.
• Fail-closed everywhere: degenerate-required, wrong-entity, dormant/stale-exclusion, and missing-column classes each fail the gate with named columns.

The real-donor HMRC replay (adjudicated, fenced)

Run against the licensed SPI donor (identity re-verified: sha256 5ef82946…, 141,323,762 bytes — never committed, held out via git exclude; only aggregate artifacts are tracked) and the pinned HMRC ODS (ad063b06…):

• SPI-side machinery is exact: deterministic TEI + TII = TI holds for all 432,779 donor predictions; the ODS parse contract passes against the real sheets.
• All 208 published-table facts ship as evidence-fenced exclusions — zero exact, zero directional. The FRS instrument cannot express five of the ten SPI employment-measure constituents (taxable benefits-in-kind, deductible employment expenses, taxable termination pay, both miscellaneous legs — collected as receipt flags or gross amounts, not taxable monetary measures), and two more only partially (per-constituent evidence table in UK_COVERAGE_PROGRESS.md). Because the missing legs include a subtraction (expenses), omission bias is not one-directional, so no directional facts survive either. HMRC's distribution tables cannot oracle FRS-based microdata like-for-like; the report says so per fact instead of shipping biased comparisons.
• Wired where source-faithful: PAY, JSA/IS legs, the structurally-expressible incapacity leg; OSSBEN and SRP as explicitly-named subsets (ossben_identifiable_subset, srp_regular_code5), never presented as the full SPI concepts.

Staging build

populace_uk_2023_hmrc_staging.h5 builds through the seam with the gate green; the committed national_staging_build_record.json pins every input artifact by sha256 (FRS tables, HMRC surface = the pinned ODS, replay report). No row-level donor data in any tracked path.

Verification

63 UK/HMRC tests pass; ruff (CI version) clean; manifest-regeneration check green; gate report committed at 145/0 passing.

Conductor adjudications recorded in UK_COVERAGE_PROGRESS.md: the Q1 deterministic-TI rule, the Q2 identical-crosswalk rule and its two fail-closed stops, per-constituent verdicts, and the exclusion promotions.

🤖 Generated with Claude Code