Exclude corporate_wealth from Universal Credit countable capital pending a pension split

[MaxGhenis]

Interim fix for the Universal Credit capital test counting imputed pension wealth. Part of the diagnosis in policyengine-uk-data#452 (mechanism M2 in this comment). The proper fix is a dataset-side pension split (separate task); no new uk-data release is involved here.

What changes

corporate_wealth leaves gov.dwp.universal_credit.means_test.capital.sources. Both dataset builds construct it from the WAS Round 8 household file as (total private pension wealth − current defined benefit pension wealth) + employee shares and options + UK shares + investment ISAs + collective investments (policyengine-uk-data imputations/wealth.py:145-154; the microcosm build uses the same arithmetic). The Universal Credit Regulations 2013 disregard the value of pension scheme rights (Schedule 10 paragraph 10, verified against the in-force text); shares, investment ISAs and collective investments are countable under regulation 46 with no general Schedule 10 disregard. Pension wealth is 85% of the bundle's mass in the WAS donor, so counting the bundle fails the £16,000 test for claimants whose only large asset is a pension, and no artifact column isolates the share-like remainder. The parameter description states the approximation and the intent to restore the share-like remainder once the dataset splits it out.

Note the asymmetry this fixes only partially: 96.7% of FRS-reported-UC benefit units failing the capital test have corporate_wealth alone above £16,000, but only 47% of them pass after this change — the rest also carry imputed savings above the limit (see follow-ups).

Verified scope

Every consumer of corporate_wealth and uc_assessable_capital was traced (multi-agent read + adversarial verification, plus two independent full-population diffs):

• The sources list has exactly one reader, uc_assessable_capital. Housing Benefit, JSA, ESA, Income Support and Pension Credit each read their own sources list (all still include corporate_wealth); their assessable capital and awards are bit-identical before/after, as are total_wealth, shareholding, corporate_land_value and the LVT/land inputs, which read the input variable directly.
• Changes propagate only through universal_credit itself: council tax reduction (national schemes count UC in applicable income; the legacy-style local schemes also read uc_assessable_capital when a UC award exists), benefit cap, deductions, poverty rates, MTRs, deciles and government balance all move for UC-award reasons only.
• The dividend gate at uc_unearned_income.py:24 still reads corporate_wealth > 0 directly, deliberately: shares remain countable capital in law, corporate_wealth is the only available signal a benunit holds them, and regulation 72(3) treats dividends as capital once tariff income applies. Worth about £0.25bn of UC in 2025 on the measured artifact; the capital_derived.yaml description now documents the reasoning.
• No runtime randomness confounds any of this: would_claim_uc is a stored dataset column and deduction draws are deterministic hashes of benunit_id; repeated runs are bit-identical.

Measured impact (enhanced_frs_2024_25.h5, uk-data v1.56.16, year 2025, stored weights)

Engine before: main @ c93e1a0. Engine after: this branch. Admin: DWP Apr–Dec 2025 monthly averages (deductions-derived households; childcare Jan–Aug 2025), GB geography.

series	admin avg (m)	before (m)	gap	after (m)	gap
households (GB)	6.759	6.199	−8.3%	6.708	−0.8%
housing element (GB)	4.304	4.868	+13.1%	5.067	+17.7%
LCWRA element (GB)	2.417	2.102	−13.0%	2.195	−9.2%
carer element (GB)	1.134	0.641	−43.4%	0.683	−39.7%
childcare element (GB)	0.184	0.471	+156.1%	0.492	+167.6%
single, no children (GB)	3.447	2.969	−13.9%	3.279	−4.9%
single, with children (GB)	2.226	2.135	−4.1%	2.160	−3.0%
couple, no children (GB)	0.284	0.225	−20.8%	0.330	+16.3%
couple, with children (GB)	0.900	0.871	−3.3%	0.939	+4.3%

Other movements (UK basis): caseload 6.346m → 6.896m; spend £75.30bn → £78.48bn (the #735 target surface's two OBR UC spend references sum to £79.3bn for FY2025-26); eligible benefit units 10.52m → 17.35m; FRS-reported-UC benefit units failing the capital test 2,958 → 1,578 records (1.38m → 0.87m weighted); benefit units over £16,000 across the whole population 22.5m → 13.3m; poverty 11.3% → 11.0% BHC and 14.6% → 14.3% AHC (child AHC 17.7% → 17.2%); benefit-capped benunits 0.134m → 0.148m; council tax reduction £2.290bn → £2.288bn (UC passthrough only).

Regressions this knowingly worsens (they measure other defects, previously masked by the capital test): the housing element overshoot rises to +17.7%, the childcare element overshoot to +167.6%, and couple-no-children flips from −20.8% to +16.3%. Element incidence is formula-driven with no direct calibration target (see #452), and the flat 55% take-up draw is a separate mechanism (M1).

This differs from counterfactual C in the #452 diagnosis (caseload 6.69m GB, spend £78.23bn), which zeroed the corporate_wealth input and therefore also relaxed the HB/JSA/ESA/IS/PC and local-CTR capital tests and disabled the dividend gate. This PR changes none of those.

Tests and CI

• is_UC_eligible.yaml: the corporate-wealth case now asserts exclusion.
• test_stock_capital_periods.py: replaces the corporate-wealth contribution with non_residential_property_value so the stock-proration assertions keep their force, and asserts the exclusion.
• New test_uc_capital_sources.py: pins the list across instants, the exclusion, the still-counted sources, and the deliberate scope limit (HB and total_wealth still count the bundle).
• reforms_config.yaml: the "Reduce Universal Credit taper rate to 20%" regression fixture is re-baselined from −18.1 to −30.3 on the CI dataset (enhanced_frs_2023_24.h5 @ 1.40.3) — the taper reform now operates on the larger eligible caseload. The fixture file states expected values are current-model regression values, not external costings. All other fixtures move by at most £0.1bn; the UC deductions DWP-statistics bands still pass.
• Full local funnel: 1,118 YAML cases, non-microsimulation pytest suite, and the microsimulation-marked suite on the CI-pinned artifact.

Follow-ups (out of scope here)

1. Dataset-side pension split (the proper fix, own task): split private_pension_wealth out of corporate_wealth in the microcosm WAS stage, add a matching model input variable, and return the share-like remainder to this list. The remainder is not DC-only — it keeps retained DB rights and pensions in payment.
2. Legacy means tests: HB, JSA, ESA, IS and PC still count the bundle through their own lists, though the same disregard exists in their regulations (HB 2006 Sch 6 paras 31-32; IS 1987 Sch 10 paras 23-23A; JSA 1996 Sch 8 paras 28-29; ESA 2008 Sch 9 paras 28-29; SPC 2002 Sch V para 22, where pension-age treatment needs its own check). Deliberately unchanged pending the maintainer's call and the build split.
3. The savings tail: 1,578 reporter benunits still fail the £16k test, almost all on imputed savings, whose artifact distribution is far heavier than the WAS donor's (weighted median £4,000 vs £800; share above £16k 38.7% vs 19.3%) — a build-side imputation question, not a model one.

Measurement provenance: artifact sha256 e433e532… (HF revision a9e52499), licensed data, aggregates only; scripts under ~/PolicyEngine/_uk-defects-452-448/measure/ (measure_pr_before_after.py), matched dependency pins across both venvs (policyengine-core 3.31.0).

🤖 Generated with Claude Code