UK spine follow-ups: SPI donors funded, WAS coherence, residential split, Child Benefit, release-cut blockers (#1063)

[juaristi22]

Closes #1063. On main (after #1084). Receipts: experiments/1063-uk-spine-followups.md.

The eight items and the fold-ins

• Items 1, 2, 4 (c1): the SPI income band donors become a mass-conserving support channel. Each band seats max(120, ceil(published taxpayers / 250)) donors at equal weights, funded in proportion from the households of each donor's region; no created mass, graph policy declared, an identity-keyed seating (salt spi_income_band_donor_draw) the E8 receipt reruns from ids with the licensed tape.

• Item 3 (c2): mass_increasing_support leaves the coverage validator.

• Items 5, 6 (c3+c4): WAS wealth is coherent with tenure (no mortgage debt off mortgaged tenure, no main-residence value off owner tenures) and with its own totals (property and financial wealth derived from their components, identities asserted at cleaning); new gate uk_stage_was_wealth_coherence.

• Items 7, 8 (c5, c6, c5b): the residential-property flag is carried as weight by a new expansion stage, cgt_residential_split, after the incidence anchor: each liable gainer's solved probability splits the household into arms at p·w and (1 − p)·w (2^k arms for k gainers), so the residential count and gains are identities on HMRC Table 8a in every gain band at design weights, with no draw and no seed; calibration decides between the arms as it does for the incidence clones. The asset-type stage types the arms and fits Table 7 once per claimant status with an error-diffusion walk (seed 554 retired). The per-band walk of c5 is replaced because its gains error in the open band was bounded by one stake several times the band's expectation.

• Fold-ins: salary sacrifice predicted only for people who are paid (c7); Child Benefit claims and opt-outs redrawn after the SPI chain against HMRC's August 2025 statistics in a new child_benefit_take_up stage (c8).

• Release cut (c9): the certifier rehearsal's five blockers cleared at their source (candidate label, export surface read without ids and with the weight, 43 declared outputs allow-listed, incapacity_benefit_reported dropped at the release boundary, FRS 2024-25's TRAIN2 read, the LCFS donor floored at 0 with a receipt, three signed register entries). The plan-2 measure exclusions are renewed to 2026-11-03.

• Identity receipts after the pensions PR: since microcosm#1069 c9 the salary_sacrifice stage lowers a converted record's pay in place, after the CGT support split and the incidence anchor banded on it, so the E8 recomputes banded 2,797 clone carriers on post-sacrifice pay and failed on the final build's spine. The stage now keeps a converted record's pre-conversion pay on salary_sacrifice_pre_conversion_pay (an internal carrier, dropped at the release boundary) and the identity tool reverses the conversion for every receipt of an earlier stage; the residential split walks the liable gainers in person_id order so its arms are bitwise identical for any row order. The certifier's non-negative and QRF tail gates read export-dropped declared outputs from the spine frame.

• Target registry: the SPI 2023-24 Table 3.6 employment, self-employment and private-pension amounts by total-income band are projected to 2025 with HMRC's band-specific Income Tax Liabilities growth, as the count, total-income and tax rows by band already were, instead of a flat national index. The flat average-earnings index raised the £15,000 to £20,000 band's employment amount 10.7% while HMRC projects the band's total income down 2.3%; the three lowest employment targets move from £18.7bn, £53.8bn and £190.4bn to £17.7bn, £47.5bn and £172.5bn, and the frame's design-weight gaps from 20 to 31% to 15 to 22%. The remaining gap is the frame's composition of the low bands, documented in the receipts note as a follow-up.

Measured so far (licensed local arms, nothing released)

• Arm C (the walk): all spine gates pass; residential gains gap 0.3% of target; calibration loss 0.00853 against main's 0.00836, ESS 5,217 (5,198), max-to-median weight 1,115 (1,079; bound 1,151), projection entrants 6,715 (6,711).
• Arm B (WAS): all spine gates pass including the coherence gate; the rule drops £127bn of £1,266bn off-tenure mortgage mass; renters with property wealth fall to 16.5% private and 4.6% social.
• Child Benefit trial: 86.7% of eligible children claimed for (HMRC 86.6%), 9.0% of claiming families opted out, 11.1m children in payment against 13.5m before.
• Final national build (two runs, 2fe7d8b and 64c460b; the second differs only by the carrier column): all spine gates, 7/7 seam gates, calibration loss 0.00713 over 1,090 targets (main's 2026-09-30 build 0.00836); evaluation against the enhanced-FRS incumbent passed (loss 0.00817 vs 0.327, 1,020 target wins vs 30); E7 and E8 pass with the licensed tape. The release-cut certifier reaches all 20 gates: 19 pass; uk_input_mass_parity fails on three columns against the uk-data 1.56.16 enhanced FRS (access_fund +535%, jsa_income_reported +1,058%, working_tax_credit_reported +747%; see the open calls). The compile-parity receipt was re-signed for the band-aware Table 3.6 amounts and the stale adult_ema exclusion retired.

Tests

Stacked PRs get no CI here. At the stack head: 983 targeted tests, the engine-free lane (6,294 passed; the two failures were the extras-free-venv and importlib-mode environment checks), the engine-UK lane (137), the integration smoke build and the parity fixture tests pass; the c9 commit was checked on its touched files plus the engine-UK lane and the smoke build.

Open calls

• Child Benefit opt-outs are drawn from fully charged families first, then the taper; or one pool over every charged family.
• other_residential_property_value (WAS DVHseVal) excludes buy-to-let, which sits in the drawn remainder.
• The Table 8a rows stay fenced; promoting them to bound targets is the UK CGT: sub-exempt gains capped at exactly the annual exempt amount become 10.8m taxpayers under uprating; interpolate the remainder below the threshold and anchor incidence #970 leverage question after the final build.
• The certifier's input-mass findings (not new to this stack: the 2026-09-30 main candidate sat at +421% and +428% on the same JSA and tax-credit columns): access_fund is one FRS household whose 19-year-old reports an access fund the FRS weeklyises to £1,035.62 (sernum 14443, ACCSSAMT), annualised to £54,037 and taken to the calibration's 10x weight bound; jsa_income_reported and working_tax_credit_reported sit in the SPI synthetic channel, where the stage-2 QRF imputes these legacy-benefit leaves onto taxpayer rows. An FRS outlier fence, the SPI channel's treatment of the two leaves, or reviewed exclusions: her call.

🤖 Generated with Claude Code

[vahid-ahmadi]

Automated review pass (Claude Code, high effort) — round 1 at ebff8412

Verdict: one blocker and one urgent split-out. The rest is design-sound. c9 allow-lists the five internal disability carriers on the export surface, and a test that was green on #1084 now fails. Separately, the plan-2 exclusion renewal sits behind unmerged #1084, so main's exclusions lapse on 3 October. The spine changes (c1, c3+c4, c5b, c8) read correctly. C2, the final build and the certifier run are still owed, so this stays draft.

All diffs are against #1084's head 004b5c9c, so they cover only this PR. I ran 13 targeted engine-free files: 356 passed, 1 failed (item 1). These are the new-stage files, plus stage health, terminal gates, battery bindings, diagnostics, source stages, release coverage, salary sacrifice and LCFS.

---

1. Blocking: c9 puts the internal disability carriers on the export allow-list.

• 48eb73bf adds person.attendance_allowance_reported, dla_sc_reported, dla_m_reported, pip_m_reported and pip_dl_reported to uk_export_surface's allowed_extra_columns (gates.json:926 and the four lines around it).
• Those are UK_INTERNAL_DISABILITY_REPORTED_COLUMNS (frs_disability.py:25). test_uk_source_stages.py::TestE3ManifestLockstep::test_internal_disability_carriers_stay_out_of_export_registers exists to keep them off the release.
• That test fails at this head. On 004b5c9c none of the five is listed, so the failure is new.
• R2's "every declared stage output the incumbent lacks is allow-listed" caught the internal carriers. Drop them from the allow-list, or drop them at the release boundary as incapacity_benefit_reported is. Stacked PRs get no CI here, so this would only have shown up after the retarget.

2. Urgent, outside the stack: ship the plan-2 renewal to main on its own.

• On main 5187fce2, slc.borrowers.plan_2_liable and plan_2_above_threshold expire on 2026-10-03 (calibration_measure_exclusions.json:362,371).
• c5368635 renews them to 2026-11-03, but only on this branch, which is stacked on UK pensions: resident State Pension, Pension Credit, contributions and salary-sacrifice relief (#1069) #1084, itself still a draft.
• Until a one-line PR to main lands, every UK build on main fails on those two exclusions from tomorrow. I'd cherry-pick c5368635 onto main today. It is self-contained, apart from the gate-digest re-pin it needs.

3. Should-fix: the WAS tenure rule drops £127bn of mortgages but keeps the property they're secured on.

• mortgage_debt is HMortGR8, all of the household's mortgages (was_wealth.py:251; the comment at :712 says so).
• The rule zeroes it off a mortgaged tenure, so outright owners' and renters' buy-to-let and second-home mortgages go: £127bn of £1,266bn, all of it OthMortR8, per the receipts.
• The property those loans are secured on stays, in other_residential_property_value and the drawn remainder. So property and net wealth rise by £127bn on exactly the landlord households the WAS undercount already hits.
• Mortgaged-tenure households also keep their other-property mortgages inside mortgage_debt. So the column means "all mortgages" on one tenure and "main-residence mortgage only" on the others.
• Suggested fix: apply the tenure rule to the main-residence mortgage (TotMortR8), which you already read for diagnostics. Carry the other-property mortgage (OthMortR8_sum) unrestricted, as its own column or added back to mortgage_debt. Then the coherence gate tests the thing that is actually structural.

4. Should-fix: the residential split's band identities hold at design weights only, and nothing reports them after calibration.

• cgt_residential_split makes Table 8a's count and gains identities in every gain band before calibration.
• Calibration binds only the two national totals: hmrc.cgt.residential_property_taxpayers and _gains in uk_population_targets.json. The by-band rows stay fenced, as the receipts note says.
• The arms differ only in capital_gains_residential_property, and main and residential rates are both 18/24 from 2025-26. So no other target distinguishes the arms, and the post-calibration split within each household is set by the solver's update rule, not by data.
• That's fine as a design, and it's the incidence-clone precedent. But C2 should report the by-band residential count and gains on the calibrated weights next to the design-weight identities, so the "identity in every band" claim isn't read as a release property.
• Also worth recording in C2: the row-level minimum weight and the positive-weight median with the arms in place. A 0.04-household arm next to a 20.76 one is fine on the fold, but it is new for any row-level reader.

5. Question: Child Benefit children in payment are about 5% below HMRC.

• The vendored August 2025 statistics give 12.73m children registered and 1.00m opted out (all ages), which implies 11.73m in payment.
• The trial reads 11.1m in payment, while the claim rate (86.7% against 86.6%) and the opted-out family share (9.0% against 684,635 / 7,552,330 = 9.07%) both match.
• So the gap is in the eligible-child base or in the age mix of opt-outs, not in the rates. HMRC's opt-outs are concentrated at ages 10–16 (63,975 to 94,905 a year against 9,770 at age 0).
• Is the gap in the frame's eligible children, or in opt-out families having more children than the frame's charged families? One line in the receipts would settle it.

6. Nit: the stage notes overstate what calibration binds. cgt_residential_split.py:23 and the source_stages.json / sources.yaml notes say calibration "binds" Table 8a. It binds the two national totals only (see item 4). Reword it to "binds the national residential taxpayers and gains".

---

Your open calls

• Child Benefit opt-out pool: I'd use one pool over every charged family, with selection probability proportional to the charge as a share of the benefit (1 for fully charged, tapering below). That keeps fully charged families most likely, as now, without making taper families structurally ineligible. Opting out has a real cost for them (a Self Assessment return) and HMRC doesn't split opt-outs by income. Either way, check the age profile of opt-outs against the vendored by-age rows, since that's where item 5's gap may sit.
• other_residential_property_value excludes buy-to-let (DVBltVal): fine as long as the buy-to-let value lands somewhere in property_wealth. Item 3's fix matters more here, because buy-to-let is exactly where the dropped mortgages sit.
• Table 8a rows fenced: agree. Leave promotion to UK CGT: sub-exempt gains capped at exactly the annual exempt amount become 10.8m taxpayers under uprating; interpolate the remainder below the threshold and anchor incidence #970 after the final build, with item 4's post-calibration readings as the evidence for it.

Checked and correct

• c1, mass and seating:
  • Donors are funded from each donor's region stratum. Every stratum's and the total household mass are conserved, and the stage refuses if funding would empty an incumbent row (spi_band_donors.py:694-702).
  • Seating is an Efraimidis–Spirakis draw without replacement on identity-keyed uniforms, salted per band, and households already taken are excluded across bands. So no source household seats twice.
• c1, identity keys: donors carry their own support clone index (2), so geography identity keys stay unique.
• c5b, identity and the fold:
  • Arms take id + j × multiplier and carry cgt_residential_clone_index. The geography identity kernel reads both and enforces an int64 index (graph_population.py:84-85, 402).
  • The family fold now folds residential arms onto their household as well as support copies (diagnostics.py:492), so the weight-ratio and ESS gates read the folded weights.
  • k > 3 liable gainers is refused rather than grown.
• c8, rates: the claim rate and opted-out family share match the vendored HMRC rows (item 5 aside).
• c3+c4: the coherence gate and the derived-total identities are asserted at cleaning.
• Tests: all the new-stage test files pass.

[juaristi22]

Thanks, all six addressed in the latest commit; the branch is now on main (#1084 merged) with the same twelve commits plus this one.

1. Carriers (blocker). The five internal disability carriers are off the allow-list and leave at the release boundary with incapacity_benefit_reported (UK_RELEASE_EXPORT_DROPPED_COLUMNS); test_internal_disability_carriers_stay_out_of_export_registers passes, and the terminal-gates test now asserts the boundary drops exactly the reviewed exclusions plus the carriers.
2. Plan-2 renewal. María's call: it stays on this PR, which she expects to merge today or tomorrow, and no dataset is built without it.
3. Mortgages. Taken as suggested, with one difference: the other-property mortgage is HMortGR8 less TotMortR8 on the donor rather than OthMortR8_sum, so the identity mortgage_debt = main_residence_mortgage + other_mortgage holds on every donor row and no mass leaves (the £15bn of HMortGR8 that is neither TotMortR8 nor OthMortR8_sum stays). Only the main-residence mortgage is tenure-stratified; the other mortgages are drawn on every tenure, conditioned on the property the chain already placed, and mortgage_debt is the derived sum. Eight segments (the first seven roots unchanged), the coherence gate holds main_residence_mortgage_off_mortgaged_tenure_rows at zero and mortgage_debt_violation_rows at zero, and the receipt reports the other-mortgage mass held off the mortgaged tenure on each side instead of a dropped mass.
4. By-band identities after calibration. scripts/residential_readout.py on C2's calibrated weights: the two bound national totals hold (count 202,628 vs 202,630; gains £12.246bn vs £12.242bn) and within them the solver moves mass between bands (£250–500k 3,680 → 4,256 taxpayers, £500k–1m 2,004 → 2,173, £1–2m 730 → 552, £2–5m 427 → 274, open band 134 → 139; under £250k within 3%). Row level with the arms in place: smallest positive calibrated weight 0.012 (an arm, 0.050; smallest design arm 0.144), positive-weight median 26.6 (arms 11.8), no row at zero. Both are in the receipts note, with the by-band identities now described as a design-weight property.
5. Children in payment. The 11.1m is at design weights on the 1 October spine, where the eligible-child base is 14.04m; the rates match HMRC and opted-out families carry 1.82 children each, so claimed children are 12.17m and 1.09m of them opted out. On the 30 September build's calibrated weights the base is 14.79m, which at the same rates gives 11.7m against HMRC's 11.73m. The gap is the design-weight child base, not the rates or the opt-out ages.
6. Wording. Done in the docstring; the manifest notes already named the two rows.

On the open calls: the opt-out pool and the Table 8a promotion are María's; the buy-to-let value sits in the drawn remainder of property_wealth, and with item 3 its mortgage now sits beside it.

One more thing the stack arm found: dwp/uc_payment_dist/SINGLE_annual_payment_28_800_to_30_000 has a single supporting household on every main-based spine (a disabled single council tenant, design weight 760 against 1,654). On the stack its award moved £678 lower, out of the band, when the larger donor block reshuffled the rows every later non-identity-keyed draw reads; the cell joins the measure exclusions beside its 27,600–28,800 sibling of the same uk-data#452 class, for María's signature.

🤖 Generated with Claude Code