Bump PolicyEngine UK to 2.91.0 and Core to 3.30.4 - #496
Conversation
The UK model's 2.90+ line floors policyengine-core at 3.30.1 (the nested-branch cache fix behind the CGT response hardening in uk#1803), so the bundle's exact core pin moves with it — to 3.30.4, the version the UK model's own lock has been validating since that arc, rather than the newer 3.31.0, to keep the blast radius minimal. Downstream consumers waiting on this pin: PolicyEngine/uk-cgt-paper#7 and PolicyEngine/uk-equalising-cgt#1, both migrating off hand-rolled or conversion-based CGT behavioural responses onto the model's parameters.
|
Closing as wrong-shaped, and the Verify-bundle-metadata gate caught exactly why: pyproject pins are generated from src/policyengine/data/bundle/manifest.json, and a hand-edit bypasses the certification record. Deeper: the bundle's UK dataset entry certifies the June microcosm release against model 2.89.2, and that release's own manifest declares compatible_model_packages ==2.89.2 — so a package-pin-only bump (even via scripts/prepare_package_bundle_update.py) would strand the certified pairing. The correct single move, coming once the new UK data artifact publishes: one bundle update pairing policyengine-uk 2.91.0 + core 3.30.4 with the new microcosm release, certification rerun, artifacts regenerated. Downstream repos (uk-cgt-paper#7, uk-equalising-cgt#1) then bump the bundle and score through the interface — no direct model pins anywhere. |
1 participant
The UK model's 2.90+ line floors policyengine-core at
>=3.30.1(the nested-branch cache fix behind the CGT response hardening, policyengine-uk#1803), so the bundle's exact core pin moves with the model pin — to 3.30.4, the version the UK model's own lock has validated since that arc (not the newer 3.31.0, keeping the blast radius minimal).What 2.91.0 brings through the interface: retention-rate CGT elasticity semantics on hardened branch machinery (#1803), CGT rate effective-date handling and fiscal-year blending (#1806), and the new auxiliary MTR elasticity parameter (#1823).
Downstream consumers blocked on exactly this pin: PolicyEngine/uk-cgt-paper#7 and PolicyEngine/uk-equalising-cgt#1 — both migrating off hand-rolled or conversion-based CGT responses onto model parameters (the second currently carries a sign inversion under ≥2.90 semantics if rerun unmodified, so the sooner the bundle moves, the sooner that trap closes).
Sequencing note: aware release cadence here runs through Vahid (#491's publisher-claim thread) — no urgency on cutting the release itself; landing the pin unblocks the downstream PRs to be authored against a real version.
🤖 Generated with Claude Code