Minimum wage coverage and bite (mode 1): the first external check on PE-UK's minimum-wage machinery (#88)

[vahid-ahmadi]

Closes #88. Branches from main — independent of #91/#92.

PolicyEngine-UK carries minimum_wage, minimum_wage_category and the whole gov.hmrc.minimum_wage parameter tree, and there was no external claim about the minimum wage anywhere in the 15,858 UK rows. 119 claims from the Low Pay Commission Report 2025 data annexes:

metric	n
minimum_wage_bite	66	per cent of median hourly pay
minimum_wage_coverage	13	jobs at or below the rate
minimum_wage_coverage_rate	40	per cent of jobs

A JOBS unit concept

Coverage counts jobs. One person can hold two and one household several, and ASHE — the survey behind it — is an employer survey of jobs. Mapping that onto persons because the unit already existed would have misstated what the number is. uk_aliases records lpc:jobs DISTINCT from dwp_hbai:persons and uk_hmrc:individuals.

Count and rate stay separate metrics for the same reason poverty_count and poverty_rate do.

The bite denominator rides on the claim

Bite is the rate over ASHE's median hourly wage of full-time workers. The certified PE-UK world is FRS-based — a household survey whose hourly-pay distribution is not ASHE's. So a PE-vs-LPC bite gap is a survey-population difference before it is an engine question.

That axis is in conditions["denominator"] on all 66 bite claims, a bite claim without one is refused at staging, and source.json states the axis must be sized before any engine defect is named. I suspect this is the family's most interesting finding rather than a problem with it.

What is deliberately not read, tallied rather than implied

• 191 of the 193 main-annex sheets — the report's other figures (earnings distributions, employment effects, international comparisons). Each is its own population decision, not something to sweep in because the file was already open.
• The 349-row local-authority coverage table. LA geography is finer than any UK geography this repo registers, and the certified engine cannot resolve it either; opening an LA vocabulary for one table would create identities nothing else can join to.

Both are declared drops with reasons and row counts, asserted by a test.

Every claim also says which rate it is about — adult rate, an age-band rate, or all NMW/NLW rates together are three different questions. LPC's "East of England" is aliased once onto the repo's registered "East".

Verification

• suite 282 passed
• two builds agree on content_hash
• no-drift gate clean
• ruff format --check clean

Reviewers

@MaxGhenis @DTrim99 — two calls worth challenging. (1) JOBS as a new unit concept: the alternative is to refuse coverage counts entirely until a jobs population exists elsewhere, but I think an honestly-labelled job count beats no claim. (2) Reading only 2 of 193 sheets: I'd rather land a small, fully-identified population than a large half-triaged one, but the split is a judgement and the tally is there so you can see exactly what I left.

[DTrim99]

Reviewed on tip f934ddc — pytest tests/test_lpc_minimum_wage.py = 20 passed. Clean; same lane discipline as the rest of the stack.

Verified:

• Provenance — both .xlsx SHA-256 pinned and recompute-checked at build + in a test; edition pinned ("Low Pay Commission Report 2025"), source.json uses versioned GOV.UK asset URLs (not "latest").
• Adapter — parses by sheet name + title-anchor (a reshuffled workbook raises), no cell-ref fragility. Coverage(jobs) vs bite(percent) modeled as distinct Metrics from distinct columns; spot-checked UK 2025 raw cells (2,020,000 / 6.6) stage verbatim with no re-derivation or %/count confusion.
• Accounting — ingest asserts counts == _EXPECTED and finish raises on claim_id collision; 119 = 66 + 13 + 40 pinned, and the 191 unread sheets + 349 LA rows are declared drops with reasons (tallied and asserted), so the identity holds.
• New metrics MINIMUM_WAGE_BITE/_COVERAGE/_COVERAGE_RATE + UnitConcept.JOBS keep count vs rate apart (poverty_count/rate precedent) and are non-duplicative.
• Fail-loud on unmapped geography / age-band / rate_scope / unit, and — the strongest design choice — bite refuses to stage without its denominator. All tested.
• Identity — lpc:jobs DISTINCT from dwp_hbai:persons and uk_hmrc:individuals; rate_scope keeps adult-rate vs age-band vs regional claims from colliding.
• held_out — all 119 → HELD_OUT with an ASHE-jobs-vs-FRS-survey basis; never_calibrate permanent-holdout precedence is checked before the lpc branch.
• Determinism/mirror — data/lanes.json byte-identical to the app/public mirror; the only rebuild "drift" was CRLF-vs-LF from running under Windows Python (CI produces LF), an environment artifact, not a lane defect.

One cosmetic nit: a uk_aliases comment says the region is "aliased here, once," but the "East of England"→"East" alias actually lives in the adapter REGIONS map (functionally correct — staged output never contains "East of England"; the comment is just slightly misleading). No partner test files touched. Approving.

[vahid-ahmadi]

Thank you — and the nit is a fair catch, fixed in the latest push.

You're right that the alias doesn't live where the comment said. "East of England" → "East" is applied on the harvest side, in the adapter's REGIONS map, so the staged rows never carry the LPC spelling and uk_aliases only ever accepts the repo's own region names. The comment now says that instead of claiming the alias is registered there.

On your CRLF observation — that's the same thing David hit on #92, and there's now a .gitattributes on that branch forcing -text for the vendored .csv/.csv-metadata.json/.jsonl artifacts. It doesn't cover .xlsx because those are binary and git won't touch them, so this lane's pins are safe either way; but if the LPC lane ever grows a text-shaped artifact it should inherit that rule.

Noted on the rest — particularly that you spot-checked the UK 2025 cells (2,020,000 / 6.6) staging verbatim. That pair is the one I'd most want independently eyeballed, since it's where a count-vs-rate confusion would be invisible.