[MaxGhenis]

Draft, hold. Which regional route the Budget-week UK pairs take is María's call (microcosm cut runbook Q13). This change is ready in case she picks the accepted-mismatch route.

Why

• Each rolling Microcosm UK national re-cut before the 28 Oct Budget lands on a newer policyengine-uk.
• The only local-area (dense) cut, microcosm-uk-2024-25-dense-20261019T000432Z-ee5c2dbe, was built on 2.132.6 and claims ==2.132.6. So certify-data refuses to keep routing constituency and local-authority regions to it.
• A new dense cut takes 2–13 h to build, and its two-build sign-off last took about 30 h.
• Dropping the routing loses area regions from sim-api's UK default.
• Area receipts are uncertified already: the certificate covers the national dataset only (Certify the Microcosm UK national and local-area pair with policyengine-uk 2.132.6 #685).

What

bundle.py certify-data --accept-regional-model-mismatch "<who accepted it and why>":

• Scope. UK only.
• Routing. Keeps the regional release routed when it does not cover the pinned model, and records this in data_releases.uk:

  "regional_certification": {
    "compatibility_basis": "accepted_model_mismatch",
    "certified_for_model_version": "<X>",
    "built_with_model_version": "<Y>",
    "accepted_reason": "<reason>",
    "area_receipts": "uncertified"
  }

• Typing. The record is a typed field on CountryReleaseManifest, so consumers see it rather than Pydantic silently dropping it.
• Warning. Certification emits a warning naming both versions.
• Refusals. It refuses:
  • an empty reason;
  • a regional release that already covers the version, so a certificate never records a mismatch that isn't there;
  • a run without a regional release.
• Core. The core must still be covered; a core mismatch is never accepted.
• National certificate. The national certificate and its compatibility_basis are unchanged.

The cost, stated in the docs

• The local-area file was built and calibrated under the older model.
• policyengine-core ignores, with a logged warning, any column the newer model no longer defines (simulation.py, "column(s) that do not match any variable … were ignored").
• Inputs that the newer model added take their defaults in area runs.
• A certification PR that uses the flag should list that input-surface difference.

Tests

tests/test_certify_data_release.py (66 pass locally on Python 3.13):

• the refusal without the flag;
• the routed and recorded case: record contents, the typed model, sim-api still routing both area types, and the warning;
• the refusal when the release already covers the version;
• empty and blank reasons;
• the core still enforced;
• no regional release;
• bundle.py forwarding the flag.

Companion: PolicyEngine/microcosm#1348 adds tools/certify_uk_cut_in_policyengine_py.py, which passes this flag only when asked.

axiom: n/a: certification tooling, no policy encoding

🤖 Generated with Claude Code