UC residual after the pension-wealth split: support 4.24m vs 6.76m — re-draw undershoot, the 0.55 take-up freeze, and composition shape #809

Activity

1. juaristi22 commented on Aug 28, 2026

   

   juaristi22

   on Aug 28, 2026

   CollaboratorAuthor

   More actions

   Mechanism 2 measured — the three-point would_claim_uc probe on post-split spines. Three full spine builds (0.55 control / 0.70 / 0.85) on identical code (the #813 re-bind branch, policyengine-uk 2.92.1), each through the seam at 1500 epochs / family_equal. Receipts: data/ukds/acceptance/809-lever-probes/ (spines + sidecars), runs/uk-623-first-calibrated/spine-assessment-l{055,070,085}/. Full-rebuild probes, not column patches — the g-vs-h precedent showed the flag feeds engine-materialised predictors downstream.

   cell	0.55	0.70	0.85
   dwp.uc.households (init support)	−18.3% (4.26m)	−15.7% (4.80m)	−14.0% (5.36m)
   households_single_no_children	−7.0%	−2.4%	−0.1%
   households_couple_no_children	+0.0%	−0.0%	+0.0%
   households_couple_with_children	−5.9%	−2.0%	−1.5%
   households_single_with_children	−35.6%	−35.6%	−34.2%
   households_children_1	−29.6%	−29.0%	−27.6%
   households_children_2	−35.7%	−32.7%	−32.4%
   households_children_5_or_more	−29.1%	−29.0%	−29.0%
   obr.housing_benefit	−0.1%	+0.1%	−44.9%
   loss	0.0254	0.0257	0.0265

   Findings:

   1. The lever's marginal claimants are overwhelmingly childless. Raising 0.55 → 0.85 adds +0.73m single-childless initial claims but only +0.07m lone-parent claims — the eligible-non-claiming pool is ~10:1 childless (FRS-reported family claims are already anchored true; eligible-but-unclaiming families barely exist on the spine). No take-up rate — uniform or composition-conditional — can manufacture the missing ~0.8m lone-parent claims. Mechanism 3 is not a take-up shape problem; it is a support/eligibility gap in with-children claims, now cleanly isolated by the Re-bind the DWP UC composition and TCL disability-children targets to the publisher's grain (#808, #809) #813 grain fix.
   2. The legacy trade-off reappears at 0.85: housing benefit's initial support drains £2.67bn → £0.90bn as claimants move off legacy benefits, and the solver can no longer recover it (−44.9%, the same signature as the pre-split probe). At 0.70 the solver still holds HB green.
   3. The blocking cells do not clear at any rate — the four with-children composition cells stay −27.6% to −35.6% across the sweep. Per the freeze's own standard (re-raise only on evidence it moves the failing cells), 0.55 stands. Noted for adjudication: 0.70 is a strict fidelity improvement everywhere else (headline −15.7%, childless singles −2.4%, HB green, loss flat) — but choosing a rate should come from DWP take-up evidence, not from tuning to fit, so this is recorded rather than proposed.
   4. The UK spine: child disability under-identification blocks the TCL disability cells (−56.5% / −40.0% across three spine vintages) #808 cells stay green at every rate (+0.1%/−0.4%), and the marginal hmrc drifts (state_pension 50–70k ~+25.5%, private_pension count ~+25.3%, CGT ~−25%) ride just at the band on all three runs — solver-pressure marginals independent of the lever.

   Where this leaves the residual: with-children claims are short ~0.8m at the support level — either the FRS under-carries eligible lone-parent benunits at DWP's volume, or entitlement modelling zeroes claims that should be positive (childcare/element interactions), or the wealth re-draw pushes family claims over the capital threshold (mechanism 1, the remaining unmeasured decomposition). Mechanism 1's controlled re-draw comparison is the next measurement.
2. juaristi22 commented on Aug 28, 2026

   

   juaristi22

   on Aug 28, 2026

   CollaboratorAuthor

   More actions

   Mechanism 1 measured — the £16k threshold-crossing share against WAS donor truth (the E5-consistent benchmark: donor faithfulness, not incumbent-matching). Countable capital = the engine's UC assessable set (savings + other_residential_property_value + non_residential_property_value + corporate_wealth); donor cleaned with the stage's own clean_was_household_table, donor weights; spine-l055 (current main + #813) at design weights.

   surface	households with countable capital > £16,000
   WAS round-8 donor, post-split arithmetic	28.6% (n=15,128)
   spine-l055 re-draw	38.3% (n=52,846)
   WAS donor, pre-split arithmetic (pension folded in)	63.7%

   The re-draw puts ~10pp more households (~2.9m) over the capital threshold than the donor it draws from — a ~34% relative overshoot on exactly the statistic the UC caseload depends on. This is the concrete, donor-anchored defect signature behind the "~1.15m below incumbent" observation, now separated from the incumbent's own inflation (E5) and from the predictor-surface confounds of the 452 receipt: the continuous full-leaf draw spreads mid-mass — it rarely reproduces the donor's zeros and small balances, so the >£16k share inflates even while aggregates stay donor-faithful.

   The pre-split row also sizes M2 cleanly: the pension fold alone had 63.7% of donor households over the threshold.

   Where this points: the fix is in the draw's zero/small-balance behaviour (regime-gate the zero class per column, or donor-atom faithfulness in the mid-range), which is the #145 draw-quality workstream — promoting it from "registered revisit" to a measured defect with an acceptance statistic: the spine's countable->£16k share should land near the donor's 28.6%. Closing that ~10pp gap frees capital-test-failing claims across all family types — including the with-children claims the take-up lever cannot reach (see the lever readout above).

   Caveats recorded: donor is GB round-8 (2018–20) households vs the 2024 FRS spine (population drift is real but far smaller than 10pp); nominal-pound threshold applied to both sides consistently since draws are donor-scale.
3. juaristi22

   added a commit that references this issue on Aug 28, 2026

   Merge pull request #813 from PolicyEngine/808-809-rebind

   ...

   e0de620
4. juaristi22

   closed this as completedon Aug 28, 2026