Bump policyengine-uk to 2.102.3 for the FYE 2025 absolute poverty line

[MaxGhenis]

Fixes #535

Summary

This bumps the bundled policyengine-uk from 2.90.2 to 2.102.3. 2.102.3 rebases the UK absolute poverty line to HBAI's FYE 2025 reference year, as DWP has reported it since the March 2026 HBAI release (PolicyEngine/policyengine-uk#1889). The line is 60% of the unrounded FYE 2025 medians, £431.69 BHC / £373.89 AHC a week at FYE 2025 (HBAI table 2.4ts), with OBR CPI uprating after. FYE 2021 and earlier keep the FYE 2011 line, matching DWP's current split.

absolute_bhc and absolute_ahc, two of the four measures calculate_uk_poverty_rates returns, read in_poverty_bhc / in_poverty_ahc. PolicyEngine headlines absolute_bhc (see #534). UK absolute poverty outputs rise.

Certification

The certified UK data release stays policyengine-uk-data 1.56.16 (enhanced_frs_2024_25), as #491 set it. Only the model moves.

• Publisher claim. policyengine-uk==2.102.3 was appended to the 1.56.16 release manifest's compatible_model_packages, as Bump to policyengine-uk 2.90.2 and revert the UK certified default to enhanced FRS 2024-25 #491 did for 2.90.2: HF commit 78788372. The evidence is in the commit description.
• Certification. scripts/bundle.py certify-data --country uk --data-producer populace --manifest-uri hf://model/policyengine/policyengine-uk-data-private@78788372…/releases/1.56.16/release_manifest.json --model-version 2.102.3. The basis is legacy_compatible_model_package, since the data was built with 2.89.2. Per the certifier's warning, the test suite is the arbiter.
• Evidence. On this exact artifact, for 2026-27, 2.102.3 and 2.90.2 were run side by side:
  • Every aggregate is finite, and the input columns the model doesn't recognise are identical.
  • Totals are within 0.1% for income tax, NI, UC, Child Benefit, Pension Credit, State Pension, council tax, and household and HBAI net income. The UC benefit-unit count (6.35m) is also within 0.1%. Housing Benefit is +0.8%.
  • Absolute poverty rises about 32%, which is the intended line change.
  • Relative poverty rises about 10–12%, mainly from policyengine-uk#1865: the relative line is now 60% of the median over individuals, as in HBAI. Household and HBAI net income move by less than 0.01%.
• Other files. Regenerated: pyproject.toml pins, uv.lock (policyengine-uk entries only), the bundle manifest and both TRO sidecars. The US sidecar changes only because it hashes the shared bundle manifest. The test pins UK_MODEL_VERSION and UK_DATA_RELEASE_REVISION in tests/test_release_manifests.py, and tests/test_models.py, are updated.

What changes for users

All policyengine-uk changes since 2.90.2 come with this bump, not just the poverty line. The PolicyEngine-uk CHANGELOG from 2.90.3 to 2.102.3 lists them.

On the certified enhanced FRS, 2026-27, absolute poverty. This branch computed these end to end:

• ensure_datasets built the year files with 2.102.3;
• the "before" column is policyengine-uk 2.90.2 on its own year files.

Group	BHC 2.90.2 → 2.102.3	AHC 2.90.2 → 2.102.3
All people	11.1% → 14.7%	14.5% → 19.0%
Age 0–17 (calculate_uk_poverty_by_age "child")	11.8% → 16.3%	17.0% → 23.7%

For comparison, DWP projects 17% BHC and 24% AHC for children in FYE 2027 on the same line (DWP, Low income poverty projections for children, FYE 2026 to FYE 2031, 9 July 2026). DWP counts dependent children (under 16, or 16–19 in non-advanced education), not everyone aged 0–17. On HBAI's own definition, policyengine-uk#1889 reports 16.7% BHC and 24.1% AHC.

The UK model-surface snapshot (tests/fixtures/household_calculator_snapshots/uk_model_surface.json) is refreshed with PE_UPDATE_SNAPSHOTS=1, as the certification skill prescribes for a model bump. Parameters rise from ~2,300 to ~2,600 and variables from ~800 to ~900, in 100s buckets. The four UK household snapshots are unchanged.

Year files. The files ensure_datasets builds with 2.102.3 match the 2.90.2 ones in every column except 2026 domestic_energy_consumption, which is 5.5% lower.

• This is policyengine-uk 2.100.1's intended change: energy is no longer projected forward from the data year.
• 1.56.16's build (policyengine-uk-data 12a1e028, datasets/imputations/consumption.py) already prices energy at Ofgem Q2 2026 unit rates, so 2.90.2 was applying the price rise twice.
• 2024 year files are identical.

Inputs this data doesn't have. 34 input variables added to policyengine-uk since 2.90.2 are absent from 1.56.16, so they take their defaults.

• Duties: alcohol volumes and draught shares, tobacco quantities, car attributes (CO2, engine size, registration date, fuel type, list price, VED exemption), and LPG and natural-gas kg. The new alcohol, tobacco, vehicle excise and road-fuel-gas duties therefore compute as about zero on this data (gov_balance +£0.13bn).
• Balance-sheet items: mortgage_debt, consumer_debt, private_pension_wealth, and three capital_gains_* components.
• Other: tax_free_childcare_spend_routed_share, plus the take-up flags would_claim_carers_allowance and would_claim_uc_childcare.

This is a data gap, not an incompatibility. It closes when a data release built with a newer policyengine-uk is certified.

The "Which measure to headline" docs in #534 describe the new line and should merge with or after this PR.

Checks run locally

• Full suite on Python 3.13, with the HF token: 1,187 passed, 9 skipped and 1 failed. The failure was the UK model-surface snapshot, which drifted as expected and was then refreshed. After the refresh, all 11 tests in tests/test_household_calculator_snapshot.py pass.

• Independent review (Opus 5.5): APPROVE. Its findings on body accuracy, the second changelog fragment and unrelated uv.lock marker churn are addressed in the follow-up commit.

• make format, make lint, scripts/bundle.py check and scripts/release_lock.py all pass.

🤖 Generated with Claude Code

[MaxGhenis]

Merged on Max's go (decision d631, 2026-09-29), which carries out his d516 ruling to follow DWP's FYE 2025 absolute line. CI was green on e0d105e. An independent Opus 5.5 review approved, per the description.