Run the £100,000 childcare lane on policyengine.py's certified Microcosm pair (6cba5d71, policyengine-uk 2.132.6) #228

Activity

1. MaxGhenis

   added a commit that references this issue on Oct 25, 2026

   Scope each engine note to the pairs it was checked on; record the pub…

   ...

   371f1cf
2. MaxGhenis commented on Oct 25, 2026

   

   MaxGhenis

   on Oct 25, 2026

   ContributorAuthor

   More actions

   Leads for re-checking the pair-scoped notes, read on 2026-10-25. The sources are the policyengine-uk 2.132.6 wheel from PyPI (sha256 cc54025f…) and the policyengine 6.10.0 wheel. None of this has been run on the pair yet.

   • centax-static-vs-static: still holds by the same mechanism, if the run goes through the managed route as it does today.

     • employment_income still adds employment_income_behavioral_response.
     • That variable returns 0 when gov.simulation.labour_supply_responses.income_elasticity and substitution_elasticity are both 0. Both are 0 from 2020.
     • gov.dynamic.obr_labour_supply_assumptions defaults to true, and 2.132.6's own deprecation text says the OBR elasticities "are applied by Simulation.apply_dynamics … when gov.dynamic.obr_labour_supply_assumptions is true, which it is by default". However, nothing in policyengine-uk 2.132.6 or policyengine 6.10.0 calls apply_dynamics; only policyengine-uk's tests do. The same is true of 2.89.2 and policyengine 5.0.2.
     • The verified_pairs entry should say so. It should not rest on the elasticities alone.

   • pe-universal-switch-off: does not hold. 2.132.6 contains policyengine-uk#2177 ("Preserve universal and targeted hours alongside extended childcare", merged 2026-10-12): the 2.132.6 tag is 197 commits ahead of its merge commit and 0 behind.

     • The note needs a replacement for this pair.
     • The compute test's "universal hours fall" invariant becomes pair-specific.

   • pe-under-ones: needs a measurement. extended_childcare_entitlement_qualifying_child requires age >= young_child_minimum_age, which is 0.75 from 2024-09-01 (SI 2023/1330 reg 2(5)(b)).

     • Whether any under-1 gets hours depends on whether the Microcosm cut's age is fractional. The artifact's age measurement answers that.

   • pe-england-only: the extended entitlement still tests country == ENGLAND (extended_childcare_entitlement_qualifying_child). The universal and targeted heads still need a check.

   • The dataset is already in the local Hugging Face cache. Snapshot bb2258482ef8… of policyengine/populace-uk-private holds microcosm_uk_2024_25.h5, the record's sha256 810ba980… (322 MB).

     • The population record lists only the Modal executor.
     • The requirements file is resolved for x86_64 manylinux.
     • The three policyengine wheels are pure Python, so their RECORD hashes should match on a local venv. Verify this rather than assume it.

   • The record says "Policy experiments only; never a replay variant", and experiment results carry the modern_population_sensitivity label ("not a replay or a costing of {event}"). Running this lane on the pair is therefore a design choice:

     • read the experiment population's identity in the lane's own preflight, or
     • add a lane-specific certified-pair registry.

     It is not a reuse of the experiment label. Add the two-child limit child-poverty benchmarks, recomputed on the certified Microcosm pair #225 takes the managed route for the same pair.
3. MaxGhenis

   added 6 commits that reference this issue on Oct 26, 2026

   Document certified pairs alongside replay variants

   ...

   44174a1

   Refuse a stripped pair id and a registry ingest would read differently

   ...

   1f2ef02

   Re-export the feeds: two CBP rows cite the certified pairs file

   ...

   58e286f

   Close the round-2 review nits on the pair registry

   ...

   466df5a

   Record policyengine.py 6.11.0's certified pair for the childcare lane (…

   ...

   3a20920

   Answer the second review of the 6.11.0 childcare run

   ...

   58faa0b
4. MaxGhenis

   closed this as completedin 7591e3con Oct 30, 2026