UK: port the #1095 fixes that need no new policyengine-uk release

[juaristi22]

This is the third #1095 ports PR. It covers every actionable item on #1095 that waits for no policyengine-uk release, on the locked policyengine-uk 2.122.2.

A calibration of this head passes 6 of the 7 calibrated-seam gates. uk_weight_ratio fails at 1,176 against the fence of 1,151. Main, this branch's base, fails the same gate at 1,213, so the gate was already failing before this PR (below). María ruled on 2026-10-12 that this PR proceeds and that the breach is resolved apart from it. Main's evidence reader refuses real run manifests (#1187), so both calibrations ran with #1188's fix on top, for measurement only.

This branch conflicts with main since #1167 merged on 11 October. The measurement below is on the branch's base, 7639ef8b0. The rebase onto feed 1ee7dfe and a new measurement, with a control on the new main, are still to do.

Before any national build: the renewals of the measure exclusions that expire from 15 October, and a resolution of main's weight-ratio breach. After merge, as for #1100 and #1121: the signed-difference entries for the new and changed columns (D1, D5, D3, D4), whose drafts stay outside the repository until signed, then the national build and the certifier.

Summary

Root columns (D1, D5, D6)

• D1. UC gainful self-employment (Set UC gainful self-employment from the FRS main-job status policyengine-uk-data#525). uc_is_in_gainful_self_employment holds for a self-employed main job, whatever its profit, or for a side trade whose profit is above zero and above pay (UC Regs 2013 reg 64(a)). The engine's own rule read any nonzero self-employment income as gainful, so a trader at a loss never met the minimum income floor.
  • SPI-channel rows keep their donor's employment status but draw new incomes that don't read it. So they take the earnings route alone, until the SPI draw follows employment status (Draw SPI incomes within earnings groups set by FRS employment status policyengine-uk-data#529, UK spine: composition-faithful SPI income draw — bridge predictors and rank-preserving replacement (with-children UC residual) #840).
• D5. Children placed by a local authority. is_looked_after_by_local_authority marks foster children under 16, whom the FRS puts in the carer's benefit unit. The engine counted them as the carer's own children in its child tests. At 16 or over the FRS makes a foster child with a local authority allowance a separate benefit unit, so the rule stops there.
• D6. Loss-making landlords keep the CGT residential signal. UK FRS spine: stop counting lodgers' rent and property losses as property income; sub-letting rent for every tenure #1081 counts a letting loss as zero property income, so those landlords lost one of the CGT split's two residential signs. An internal flag, dropped at export, restores it for FRS rows. A redrawn SPI row drops its twin's flag, since its property income is an SPI donor's.

Take-up and capital (D2, D3, D4)

• D2. The UC take-up population is benefit units with a UC claimant or partner aged 18 or over and under Pension Credit qualifying age, the age half of the engine's is_uc_eligible (Supply is_hbai_dependent_child from the FRS and move targets and imputations off the deprecated is_child/is_adult policyengine-uk-data#486, Supply is_claimant_or_partner from the FRS adult table policyengine-uk-data#524). An engine lockstep test names the two cases it doesn't mirror: reg 8 claimants aged 16–17, and mixed-age couples on the Pension Credit route. The contract's population text is re-signed with the rate unchanged.
• D3. Property in recorded capital. UC and Pension Credit recorded capital add the household's countable property to the financial carrier. The engine replaces every capital source with the recorded figure, and the FRS capital question holds no property.
  • Who holds it. The household reference person's benefit unit holds all of the household's property, and no other unit records any (María's ruling of 2026-10-12). Each means test counts the claimant's and partner's own capital, and WAS records property for the household. The engine's proxies share it over every claimant or partner, which puts a parent's property on an adult child's claim; they apply to fewer than 10 units.
  • At what value. 90% of its value: market value less 10% for the costs of sale. Secured debt is not on the frame and goes with the WAS property draw (Value means-tested capital less 10% sale expenses and secured debt (UC reg 49 and legacy equivalents) policyengine-uk#1969 is the engine side).
  • A unit that reports Universal Credit keeps its receipt and records no UC property. DWP assessed its capital to pay it, and the property is imputed without that receipt.
  • The reporter redraw screens an FRS unit that reports UC on its carrier alone, so that unit stays in the training set.
  • The uk_uc_capital_coherence gate re-derives both capitals with the same rule, and now checks Pension Credit capital.
• D4. Pension Credit take-up (Solve Pension Credit take-up over entitled benefit units policyengine-uk-data#510) is solved over Great Britain, where DWP's rates apply, and drawn UK-wide. Units not entitled are drawn at the 37% rate, so a reform that entitles them pays them. Entitlement is read at the calibration year, on the frame policyengine-uk projects there from its survey year.

Targets (T1–T6)

• T5. The 13 HMRC SPI savings-interest band targets reach 2025 in two legs (Regenerate the uprating factor table from policyengine-uk's load-time uprating policyengine-uk-data#541): ONS household interest income to the dataset's year, 2024, then GDP per head, the index policyengine-uk's dataset projection applies. Their factor moves from 1.182 to 1.192. The SPI donor rebase is unchanged.
• T6. The PIP daily-living caseload rows take obr.pip's England and Wales scope, resolve, and are then held as diagnostics on obr.pip (T6b): fitting them breaks the weight-ratio fence (below).
• T3. obr.council_tax binds OBR table 4.1 row 15, England, Scotland and Wales receipts, over Great Britain's households. Row 19, bound before, adds Northern Ireland domestic rates and two adjustments no household carries (£1.34bn).
• T4. obr.ni becomes a diagnostic beside the three class rows. It's a cash total against accrued class rows, and it carries items no household pays (Stop calibrating OBR total NICs on ni_employee policyengine-uk-data#537).
• T2. The salary-sacrifice relief bands bind by income tax rate class (Restore HMRC salary sacrifice relief calibration targets policyengine-uk-data#533). The measure resolver gains a declared allocation: the rise in tax on earned income, bracket by bracket, under each person's rest-of-UK or Scottish schedule. uk-data's worked examples reproduce exactly. The measure refuses if a person's classes stop adding up to the engine's earned_income_tax. The all-rates total becomes a diagnostic, and the three exclusions retire.
• T1. UC award bands read in whole pence (Calibrate GB universal credit to one OBR total; reach the UC payment top band policyengine-uk-data#530). The "£2,500.01 or over" band stays unbound (below).

What the measurement changed

The first head bound the top UC band and fitted the two PIP rows. It failed two gates, and calibrate-only runs on its spine attributed each failure to one change.

• The PIP rows raise the weight ratio. They fit within 0.1%, but on that spine they pushed the heaviest support family, a single FRS household, by 180 times the positive median. So T6b holds them as diagnostics.
• The top UC band costs the UC total. Unbound, the calibration holds 2.57 times DWP's count of lone parents in it and 1.99 times for couples with children, £13.1bn of UC, and reaches OBR's £76.1bn. Bound to DWP's counts, obr.universal_credit falls to −5.6%. Couples without children also have no household in the band. So the band stays unbound, as on main. The awards below it are too low for DWP's band counts and OBR's total to fit together, which needs its own issue.
• D3 for every unit put UC reporters over the capital limit. On the Arm 2 spine, 90 FRS-channel rows that report UC went above £16,000 through property alone. 70 were in households with no property income, and 45 in households with two or more benefit units. Hence the reporter rule.
• D1 from the donor's status flagged SPI rows with no trade. 1.67m weighted SPI-channel people were flagged with no self-employment income. Hence the earnings route for SPI rows.

María ruled on all four on 2026-10-10. The sparse-band support floor the plan added moves to its own PR.

Review round 1 (11 October)

An adversarial review of cf42d8a7a is in the comments. Five commits answer it.

• D4 read 2024 values against 2025 rules, because the engine adapter stamped the frame with the entitlement year. The frame is now stamped with its survey year and projected. Units entitled to Guarantee Credit fall from 1,354k to 1,247k.
• T5 rebased the 2022-23 SPI donors to 2024 by GDP per head, and so dropped the rise in savings interest of those years (ONS household interest income is ×1.99 from 2022 to 2023). The donor rebase is main's again, and only the projection leg of the band targets uses GDP per head. This was the cause of the pull on the top income cells that the first write-up assigned to T5 as a side effect.
• D3's training screen dropped FRS units that report UC and hold imputed property. It now screens them on their carrier.
• D6 let a redrawn SPI row keep its FRS twin's landlord flag, which turned the CGT residential signal on for 695k weighted SPI-channel people. Redrawn rows now drop it.
• T2 now checks its classes against the engine's earned_income_tax.
• W is reverted. With these fixes the West Midlands £12,570–15,000 income-tax cell is above the bound (+26.9%, and +25.3% after the D3 ruling), so María's deferral is needed. The retirement relied on the T5 error.

María's rulings of 12 October, from the review:

• D3, who holds the property. Shared pro rata, £104bn of property sat on 0.99m benefit units that are not the household reference person's. Now the reference person's unit holds it, at 90% of its value. Measured on the recalibrated head:
  • units over the UC capital limit through property fall from 2.65m to 1.50m, and almost none of them is another unit than the reference person's (1.16m before);
  • in households with other property, units that are not the reference person's are paid £1.16bn of UC across 216k units (£0.60bn across 85k before). The UC total is on target both times.
• Secured debt waits for the WAS property draw. On the WAS donor file, mortgages on other property are 26% of its value.
• Main's weight-ratio breach is resolved apart from this PR.

Still open: the signed differences. The committed uc-reported-capital-net-new-column text predates D3, and its redraft waits for María's signature.

Where this departs from the plan

• D5 reads the foster relationship code alone. The carer's fostering allowance covers fewer than 10 of the 13 looked-after children the FRS records.
• D2 doesn't use Max's test scaffolding, which isn't on main. His uk-take-up-per-person-spa branch solves the State Pension age by date of birth and will need a rebase onto D2.
• D3 gives the property to the household reference person's unit at 90% of its value, where the plan mirrored each engine proxy (María's ruling of 12 October).
• T3 adds an exact-id geography pin: the substring rule would read "scotland" in the row's concept.
• T3 and T4 give their compile-parity rows their own rationales.
• T5 changes the projection leg of the targets only (review round 1).
• T6b lets a diagnostic reference carry a fact dated by month.
• T1 leaves the top band unbound, and the support floor moves out (above).

Measurement

Each arm was built and calibrated locally on Chronicle 825406f. The control is this branch's base, main 7639ef8b0. The control and the head both carry #1188's commit on top, which changes the evidence reader and its tests only.

• Control (main 7639ef8b0). Loss 0.00702, 98.5% of 1,167 targets within 10%, ESS 4,840.
  • uk_weight_ratio fails: 1,213 against 1,151. One FRS household is the heaviest support family. It was at 1,059 on d50220d1f. An arm at Weight UK BRMA assignment by census private-rented households #1087's merge commit gives 1,213 too, so Weight UK BRMA assignment by census private-rented households #1087 moved it.
  • The West Midlands cell is at +26.5%, under its deferral.
• This head. Loss 0.00685, 98.5% of 1,168 targets within 10%, ESS 4,809. Six of the 7 gates and the input-mass gate pass.
  • uk_weight_ratio fails: 1,176 against 1,151, on the same FRS household. It was 1,155 before the D3 ruling: that household's weight fell 2%, and the positive median fell 4%.
  • obr.universal_credit and obr.pip calibrate to within 0.1%, obr.council_tax to −7.2% (control −9.5%), obr.vat to +23.3% (control +24.1%), and the three salary-sacrifice bands to within 0.3%. 108 of 111 UC cells are within 10% (control 106).
  • The £500,000–1,000,000 employment-income band is at −1.5% (control −3.4%).
  • The West Midlands cell is at +25.3% (control +26.5%), under its deferral.
  • 29 FRS-channel benefit units that report UC keep their receipt; 26 of them would have gone above £16,000.
  • On the head's weights, uk_uc_capital_coherence and uk_take_up_signal pass. The arm pipeline leaves both to the certifier, so they were run with their own evaluators.
• Earlier arms, on the older base d50220d1f, are in the receipts. They attributed the first head's failures, and their deltas also carry Weight UK BRMA assignment by census private-rented households #1087's effect.

Full numbers are in experiments/1095-uk-data-ports-3-receipts.md.

Testing

• Each commit: touched-surface tests.
• This head:
  • the engine-free lane's 534 test files, with policyengine-uk blocked as in CI: 16,152 passed and 141 skipped. The three failures are environmental: the live Hugging Face pointer test fails on main too, the telemetry token exchange needs the hosted service, and main's US head-to-head scorer test needs Python 3.14.7;
  • the engine-uk lane's 41 test files, run in full: 179 passed;
  • against the pinned feed, at 78ea6b695: the national compile-parity receipts regenerate identically, the committed references regenerate, and the NICs, PIP and salary-sacrifice diagnostics resolve their exact facts (45 passed). The D3 commit after it changes no target, reference or parity file.
• Lint and CI plan: ruff check and ruff format --check pass on the 66 changed Python files, and so do tools/ci_test_plan.py verify and tools/graph_acceptance_burndown.py --verify.

Overlaps and what follows

• Before any national build: the renewals of the measure exclusions that expire from 15 October, and Bound native evidence by bytes, not by the schema item budget #1188.
• UK landlords: bind the SPI net property amounts, leave property_wealth to the engine, add receipts and finance costs #1145 (UK landlords) also edits the target contract, measure_simulation.py, spi_income.py and the reference generator, and it deletes the West Midlands deferral this PR keeps. Its engine-input change drops this PR's internal landlord flag from the calibration input, which no measure reads.
• One consistent UK target surface: enforced cross-grain reconciliation, declared uprating holds, nation household controls and the 942a1fa re-pin (#1123 PR A) #1198 holds hmrc.salary_sacrifice.it_relief_total in its uprating register. This PR turns that row into a diagnostic, and One consistent UK target surface: enforced cross-grain reconciliation, declared uprating holds, nation household controls and the 942a1fa re-pin (#1123 PR A) #1198's compile refuses a hold that matches no held row. Whichever merges second removes the id. One consistent UK target surface: enforced cross-grain reconciliation, declared uprating holds, nation household controls and the 942a1fa re-pin (#1123 PR A) #1198 also rescales the West Midlands cell and adds a filter to the savings-interest binding T5 re-indexes.
• Set UK road-fuel and energy levels in the LCFS stage on a calendar-2024 basis (#1113) #1167 merged on 11 October. It excludes obr.vat, this head's other cell near the bound, changes the spine and moves main to feed 1ee7dfe. This branch regenerates its references, membership and parity receipts on that feed and measures again.
• Feeds: this PR is on 825406f, main on 1ee7dfe and One consistent UK target surface: enforced cross-grain reconciliation, declared uprating holds, nation household controls and the 942a1fa re-pin (#1123 PR A) #1198 on 942a1fa. Generated surfaces (contract.py, the coverage manifest, the H2 fixture, the export-surface list) regenerate; they are not merged by hand.
• Left for later PRs: the sparse-band support floor, SPI incomes by employment status (Draw SPI incomes within earnings groups set by FRS employment status policyengine-uk-data#529, with UK spine: composition-faithful SPI income draw — bridge predictors and rank-preserving replacement (with-children UC residual) #840), the WAS property draw, the UC awards below the top band, and the weight year.

Receipts: experiments/1095-uk-data-ports-3-receipts.md.

🤖 Generated with Claude Code

[juaristi22]

Independent adversarial review (Claude Code) at cf42d8a7a

Static review only. I read the code, the committed registers and experiments/1095-uk-data-ports-3-receipts.md against main 7639ef8b0, and the installed policyengine-uk 2.122.2 source for engine semantics. I ran nothing. CI is green.

Verdict. Most of the ports are sound:

• D1 and D2: the rules are as stated.
• D2 direction: mixed-age couples are correctly inside the UC population.
• Determinism: no new RNG streams.
• D6: the flag is dropped at export.
• T2: the rate-class arithmetic reproduces uk-data's examples by hand.
• T1, T3, T4 and T6: they do what the body says.
• Fixture: synthetic.

Three things should change before this leaves draft: D4 reads survey-year values against calibration-year rules, D3 adds gross property to recorded capital, and T5 changes the donor rebase in a way the targets can't absorb. One cross-PR collision with #1198 breaks the national compile whichever PR merges second.

Should fix

1. D4 reads 2024 values against 2025 rules.

   • Where: pension_credit_take_up.py:232 calls engine.materialize(with_saving, requested, entitlement_year). with_saving is the survey-year frame (time_period=period), and the adapter builds UKSingleYearDataset(fiscal_year=int(period)) from the requested period (microcosm-frame/.../policyengine_uk.py:172).
   • Effect: the frame's 2024 incomes, capital and State Pension are relabelled as 2025 inputs and never projected. On main, both the frame and the call were at the survey year.
   • Example: a single pensioner on the full new State Pension and nothing else is read at about £221.20 a week against the 2025 guarantee of £221.85, so is entitled. On the released file at 2025 the pension is £230.25 and the unit is not.
   • Consequences: the entitled set E is inflated. The residual rates, and the stage-gate receipt, describe a population the released file doesn't reproduce.
   • Fix: materialize at the survey year with the engine computing period 2025, so policyengine-uk projects the frame forward. If that's what uk-data does, cite it.

2. D3 adds gross property, not countable capital. It does so by mirroring the engine proxy, but the engine treats a recorded figure as already countable.

   • Gross value: UC_PROPERTY_CAPITAL_SOURCES sums other_residential_property_value and non_residential_property_value (Pension Credit adds owned_land) at market value (uc_capital_coherence.py:89-93).
     • UC Regs 2013 reg 49 and SPC Regs 2002 reg 19 value capital at market value less 10% for sale costs and less any debt secured on it.
     • A £120k buy-to-let with a £110k mortgage counts as about £0; this records £120k, so the unit fails the UC limit and Pension Credit deems tariff income on it.
     • was_wealth already draws other_mortgage, but only mortgage_debt reaches the frame.
   • Owners: _household_property_share splits a household's property over every claimant and partner in it (UC), or every member at or over qualifying age (Pension Credit). So an adult child's unit takes a third of the parents' buy-to-let; tests/engine/uk/test_uk_uc_capital_coherence.py:136 asserts the £66,667/£33,333 split. UC counts only the claimants' own capital (reg 18); Pension Credit counts the claimant's and partner's (SPCA 2002 s.5).
   • Size (receipts:108-112): £1,247bn of UC capital over 5.18m units. With the reporter rule, UC falls £0.80bn and 179k non-reporters lose UC.
   • What the reporter rule tells us: about 157k weighted UC reporters would have crossed £16,000 on imputed property, although DWP assessed their capital and paid them. That is evidence the imputed figure overstates countable capital. The rule protects reporters, but the same overstatement still applies to non-reporters.
   • Fix: net secured debt and the 10%, and assign property to its owning unit (the household reference person's unit, or WAS ownership where available) rather than pro rata.

3. T5's donor rebase drops a known surge that the targets already contain.

   • Where: SPI_INCOME_UPRATING_INDEX_OVERRIDES (spi_income.py:93) rebases the 2022-23 SPI donors' savings interest to 2024 by GDP per head (×1.09) instead of household interest income (×2.27). The commit says donors fall about 52% and the band targets 8.4%.
   • The problem: the engine-alignment argument (uk-data#541, uprating_indices.yaml) is about how policyengine-uk projects a dataset forward from its own year. The 2022 → 2024 donor rebase is history, and savings interest roughly doubled over it as rates rose. The 2023-24 SPI target base contains that rise; the rebased donors don't. So the solve has to make up about half of SPI savings interest through the weights, concentrated where savings sit.
   • What it explains: T5 alone moves the £500k–1m employment band from −2.5% to −13.3%, which an 8.4% cut to savings targets can't do directly. The head still has top income cells at −11.5% and −15.2%.
   • Fix: keep the household-interest index for the historical rebase to the base year and use GDP per head only for the projection beyond it. Or record why not.
   • Index pairing: the amount rows also stay on a flat index while their count rows grow by band. The implied mean per saver falls 4–13% above £30k, against about 1.00 for employment under the UK spine follow-ups after #1045: SPI band donors as a mass-conserving support channel, WAS tenure coherence, the residential draw's bound and Table 7 composition, then a national build #1063 rule. Dividends and property (UK landlords: bind the SPI net property amounts, leave property_wealth to the engine, add receipts and finance costs #1145) have the same pattern.

4. Cross-PR: One consistent UK target surface: enforced cross-grain reconciliation, declared uprating holds, nation household controls and the 942a1fa re-pin (#1123 PR A) #1198's holds register refuses this PR's T2 in either merge order.

   • One consistent UK target surface: enforced cross-grain reconciliation, declared uprating holds, nation household controls and the 942a1fa re-pin (#1123 PR A) #1198's uk/uprating_holds.json:224 declares a hold on hmrc.salary_sacrifice.it_relief_total, which this PR turns into a diagnostic.
   • One consistent UK target surface: enforced cross-grain reconciliation, declared uprating holds, nation household controls and the 942a1fa re-pin (#1123 PR A) #1198's compile refuses a declaration that matches no held row (uprating_holds.py:177, "stale"), in both the full and national loaders. The two PRs touch different files, so git merges them cleanly and only the compile fails.
   • The "Overlaps" section mentions only UK landlords: bind the SPI net property amounts, leave property_wealth to the engine, add receipts and finance costs #1145. One consistent UK target surface: enforced cross-grain reconciliation, declared uprating holds, nation household controls and the 942a1fa re-pin (#1123 PR A) #1198 also rescales the West Midlands cell W depends on (item 7).

5. The after-merge sequence needs the exclusion renewals first.

   • dwp.uc.households_housing_element expires 2026-10-15, ons.savings_interest_income 2026-10-22, and the two slc.borrowers.plan_2_* entries 2026-11-03.
   • apply_uk_calibration_measure_exclusions raises outside the window (measure_simulation.py:919-931), so any national build after 15 October fails until they're renewed.
   • This PR's own measurement relied on both near-term entries. T5 moves the model further from the D.41 row that ons.savings_interest_income excludes, so that entry needs renewing, not dropping.
   • Treat the renewal PR as a prerequisite, not a follow-up.

Minor

6. The measurement isn't of this tree, and the control is confounded.
   • Not this tree: "this head" solved on a7e5bc56f plus Bound native evidence by bytes, not by the schema item budget #1188's commit, with the spine from 0324e9e84. The W pass is a gate replay after the commit that retires the deferral.
   • Not in the repo: 0324e9e84, 341a02bd4, ae9795d7c, a42edaa9e and c7fe76e8a, and the scripts certifier_gates_check.py and uc_top_band_check.py.
   • Confounded control: the control is UK: port the remaining uk-data fixes tracked in #1095, engine items included #1121's head-r3, before Weight UK BRMA assignment by census private-rented households #1087's BRMA change merged. So every control-to-head delta also carries Weight UK BRMA assignment by census private-rented households #1087's unmeasured effect: loss, council tax −9.8 → −7.0%, minimum income floor 1.84m → 1.63m, and West Midlands +28.9 → +23.9%. The absolute head figures stand, but the attribution doesn't.
7. W's 1.1-point margin across the open PRs.
   • Retirement is required: the gate fails stale entries (terminal_gates.py:1029-1042).
   • Under One consistent UK target surface: enforced cross-grain reconciliation, declared uprating holds, nation household controls and the 942a1fa re-pin (#1123 PR A) #1198: the band bridge itl_income_tax_12570_vs_spi_region_12570 scales this cell's target by ×1.057 (55.15m → 58.31m), which on these weights would put it near +17%.
   • Under UK landlords: bind the SPI net property amounts, leave property_wealth to the engine, add receipts and finance costs #1145: that PR deletes the same entry and measured the cell at +23.4% (and +26.8% on one arm). The combination is unmeasured.
   • Root cause: the design value is still 6.2× its target. Banding the regional cells on total income including other investment income, as the old deferral reason suggests, would fix the cause instead of the register.
8. The signed-difference step is not enforced, and one committed signature is now false.
   • Not enforced: spine_swap_signed_differences.json is read only by manual tools (verify_uk_spine_parity.py, compare_uk_h5_payload.py). The release cut already lists uc_is_in_gainful_self_employment and is_looked_after_by_local_authority as reviewed candidate-only fields.
   • False signature: the uc-reported-capital-net-new-column entry (:525) says uc_reported_capital comes "from the same frs_benunit_capital carrier". After D3 it doesn't, and its only check (that the column exists) still passes.
   • Question: why aren't D4 (would_claim_pc now drawn for 37% of non-entitled units) and T5 (SPI donor savings −52%) on the D1/D5/D3 signature list?
9. D3 reporter rule.
   • The share doesn't stay put: a reporter records the carrier alone, but policyengine-uk's household residual sums only recorded figures. So a co-resident unit with the −1 sentinel absorbs the reporter's withheld share through the proxy, which contradicts "the other units of its household keep their own shares".
   • Training screen: uc_reporter_redraw.py:280/381 applies carrier plus property to every training unit, including the reporters the stage protects, so real UC reporters drop out of the QRF training set.
10. D4's 37% draw has an Income Support side effect. policyengine-uk's income_support_eligible bars Income Support where meets_pension_credit_age_conditions & would_claim_pc, with no means test. So mixed-age couples on the Pension Credit route who report Income Support and aren't entitled to Pension Credit now lose Income Support 37% of the time. The population is small, but it's undeclared.
11. D5's under-16 cutoff (frs_relationships.py:193, below_age: 16) can only drop flags. It misses 16–19-year-old foster children who are qualifying young persons in the carer's unit, who are still looked after (Children Act 1989 ss.22, 105; UC Regs reg 4(6)(a)). test_uk_frs_relationships.py:437 pins a 16-year-old as not looked after.
12. D6 on SPI rows: the landlord flag is written from the FRS twin (frs_spine.py:661), and no SPI refresh rewrites it. So SPI rows whose drawn property income is zero keep the twin's flag and turn on the residential CGT signal. That's the "donor status says nothing about drawn incomes" problem D1 rules against.
13. T2.
    • Wrong comparison: the receipt compares the class sum with Δincome_tax, which differs legitimately through savings, dividends and HICBC. The engine variable the allocation re-implements is earned_income_tax, and comparing against that would catch schedule drift.
    • Untested diagnostic: no test checks that the it_relief_total diagnostic resolves. _diagnostic_reference_metadata turns a mismatch into unavailable without failing.
    • Question: does HMRC Table 6.1 classify relief bracket by bracket or by marginal rate? At design weights, basic reads 0.66×, higher 0.53× and additional 1.37×, which takes strong opposite reweighting to fit.
14. Stale text.
    • tools/build_uk_ledger_compile_parity_signed_differences.py:76 still says the savings rows are "uprated by the household interest index", and that's regenerated into the 13 rows' signed reason, whose values moved ×0.916. docs/uk-income-anchors-280.md says the same.
    • The changelog fragments carry intermediate counts (391/1,232, 390/1,231, "stays at 1,231") where the head is 388/1,229.
    • The W fragment's "+28.9% on main" is pre-Weight UK BRMA assignment by census private-rented households #1087.
15. Nits.
    • The D1 earnings route silently skips SPI rows if the column is missing (spi_income.py:2082), falling back to the engine default.
    • The D2 lockstep test narrowed from ages 0–100 to 14–90 without saying why.
    • T1 compares annual pence, not DWP's monthly pence (negligible).
    • income_tax_rate_classes.py:65-71 files a schedule's top rate as "additional" even at 40%.
    • The UC skipped-fact count changed from 12 to 8 on the same feed pin with no note.

Merge order with #1145, #1198 and #1167

• Feeds: UK: port the #1095 fixes that need no new policyengine-uk release #1200 is on 825406f, Set UK road-fuel and energy levels in the LCFS stage on a calendar-2024 basis (#1113) #1167 on 1ee7dfe and One consistent UK target surface: enforced cross-grain reconciliation, declared uprating holds, nation household controls and the 942a1fa re-pin (#1123 PR A) #1198 on 942a1fa, all on one line. Whichever lands first, the others regenerate their references, membership, parity receipts and uk_data_target_parity on the new feed, and re-measure.
• Regenerate, don't hand-merge: contract.py and its test pins, the coverage manifest, the H2 fixture, and the gates.json export-surface list (UK landlords: bind the SPI net property amounts, leave property_wealth to the engine, add receipts and finance costs #1145's two property columns plus this PR's two).
• UK landlords: bind the SPI net property amounts, leave property_wealth to the engine, add receipts and finance costs #1145: its _engine_scratch_frame change will also drop this PR's reports_rent_from_other_property from the calibration engine input. That's harmless, because no engine measure reads it.
• One consistent UK target surface: enforced cross-grain reconciliation, declared uprating holds, nation household controls and the 942a1fa re-pin (#1123 PR A) #1198: remove it_relief_total from its hold (item 4). It also adds income_tax > 0 to the same hmrc.spi.savings_interest_income binding T5 re-indexes. That combination, with the −52% donors, has never been calibrated.
• Set UK road-fuel and energy levels in the LCFS stage on a calendar-2024 basis (#1113) #1167: it excludes obr.vat, which removes this head's other near-bound cell (+22.7%), and its LCFS levels change the spine. So a combined calibration is needed.

Checked and fine:

• Counts: 1,168 targets = 1,229 active − 61 exclusions.
• Values: T3's £1.34bn (50.116 − 48.774), and T4's total at 1.0182× the class rows.
• T5's ratios: 0.916 on all 13 rows.
• T6: a month-dated fact reaches only diagnostics.
• Exclusions: the three salary-sacrifice exclusions were genuinely redundant.
• Mirrors: the coverage-manifest and gate digests move consistently.
• Take-up contract: its rate is unchanged at 0.55.

[juaristi22]

Reply to the adversarial review (Claude Code) at bcbe274b2

Thank you. I checked each item against the code, the engine and the data, and rebuilt the measurement. The review was right on its three main items. One of them, T5, was an error in my port that also explained a result I had reported as a side effect.

In short

• Fixed: 1, 3, 6, 9, 12, 13, 14, and parts of 15.
• Changed on María's ruling of 12 October: 2 (who holds the property, and the 10%). Secured debt goes with the WAS property draw.
• Not changed, with a reason: 10, 11, and two nits.
• Open: 8, the signatures.
• Changed result: the head no longer passes all 7 gates. Main does not pass them either. See item 6.
• Main has moved. Set UK road-fuel and energy levels in the LCFS stage on a calendar-2024 basis (#1113) #1167 merged on 11 October, after these arms started, and this branch now conflicts with main. Every figure below is on the branch's base, 7639ef8b0. The rebase onto the new feed and a new measurement are still to do.

Should fix

1. D4 read 2024 values against 2025 rules. Fixed.

• The adapter stamped the frame with the entitlement year, so nothing was projected. The UK adapter now takes a dataset_period. The stage stamps the frame with its survey year, and policyengine-uk projects it to 2025 as it does the released file.
• Solve Pension Credit take-up over entitled benefit units policyengine-uk-data#510 reads it the same way: Microsimulation(dataset) on the survey-year dataset, then calculate(..., calibration_year).
• An engine test pins the review's example. Read as 2025 inputs, a single pensioner on the 2024-25 full new State Pension is entitled to Guarantee Credit. Projected, the pension is £230.25 a week and the pensioner is not.
• Effect on the stage's receipt: units entitled to Guarantee Credit fall from 1,354k to 1,247k, and the Guarantee Credit residual rate from 0.213 to 0.176. Savings Credit-only units are 371k (369k before).

2. D3 added gross property, shared pro rata. Changed, on María's ruling of 12 October.

• What I measured first. Fewer than 10 benefit units are on the −1 sentinel, so the recorded figure decides alone, and "mirror the proxy so a household adds up" carries little weight. Shared pro rata, £104bn of property sat on 0.99m benefit units that are not the household reference person's.
• Owners: changed (cf91835c9). The household reference person's benefit unit holds all of the household's property, and no other unit records any. The WAS draw is at household level, so ownership within the household is not on the frame, and the reference person is the best single reading of it.
• The 10%: applied. Property counts at 90% of its value. Alone it moved UC by £0.003bn.
• Secured debt: not in this PR. On the WAS donor file, mortgages on other property are 26% of its value, and 47% of the households that hold other property have one. Less 10% and secured debt, countable value is 64% of gross, and the share of those households at or under £16,000 rises from 2.3% to 8.1%. was_wealth draws that mortgage but does not emit it. It goes with the WAS property draw; the engine side is Value means-tested capital less 10% sale expenses and secured debt (UC reg 49 and legacy equivalents) policyengine-uk#1969.
• Measured on the recalibrated head (item 6), at calibrated weights in 2025:
  • units over the UC capital limit through property fall from 2.65m to 1.50m; of those, units that are not the reference person's fall from 1.16m to fewer than 10;
  • in households with other property, units that are not the reference person's are paid £1.16bn of UC across 216k units (£0.60bn across 85k before). The UC total is on target both times;
  • 29 FRS-channel units that report UC keep their receipt (51 before).
• The reporter rule stands. The review's point holds there in a smaller form: those 29 are reference-person units that DWP pays and that WAS-imputed property would bar. That is the WAS draw's defect.

3. T5's donor rebase dropped a real rise. Fixed. The review is right.

• The engine projects from the dataset's year. The years before it are history, and ONS household interest income is ×1.99 from 2022 to 2023 against GDP per head's ×1.05.
• My port moved the 2022-to-2024 donor rebase to GDP per head as well. Evidence in the arms: at prior weights, mean interest per saver against the target's mean was 1.02 to 2.66 in the control's top four bands and 0.40 to 0.76 after T5. Between £15,000 and £100,000 it fell from between 0.91 and 1.03 to between 0.82 and 0.95.
• Now the donor rebase is main's again. The 13 band targets take two legs: household interest income from 2023 to 2024, then GDP per head from 2024 to 2025. Their factor is 1.192 (main 1.182, T5 as built 1.083).
• It explains what I reported as T5's side effect. On the new head the £500,000–1,000,000 employment band is at −1.5% (control −3.4%, old head −11.5%), and Scotland's £200,000-and-over total income at −8.3% (control −8.2%, old head −15.2%).
• Index pairing. With the fix, the amount factor over the count factor is 0.96 to 1.06 for the bands above £30,000 (0.87 to 0.96 before). Dividends and property have the same flat-amount pattern on main; that is outside this PR.

4. #1198's hold on it_relief_total. Confirmed. Its uprating_holds.json names the row this PR turns into a diagnostic, and its compile refuses a hold that matches no held row. The PR that merges second removes that id from the hold. The Overlaps section now says so. No other row this PR removes from the fit is in that register.

5. Renewals first. Agreed. The description now lists the renewals PR as a prerequisite of any national build after 15 October. The entries need María's signature.

Minor

6. The measurement. Redone, and the result changed.

• New control: main 7639ef8b0, the branch's base, with Bound native evidence by bytes, not by the schema item budget #1188's commit on top. Loss 0.00702, 98.5% of 1,167 targets within 10%, ESS 4,840.
• Main fails uk_weight_ratio alone: 1,213 against the fence of 1,151. One FRS household is the heaviest support family. It was at 1,059 on d50220d1f.
• Weight UK BRMA assignment by census private-rented households #1087 is the cause. An arm at its merge commit, a469ebe3d, gives the control's figures to every digit: loss 0.00702 and a weight ratio of 1,213. No later merge moves the calibration. No national build has run from main since, because of Native evidence reader rejects real run manifests by item count #1187, so this had not shown up.
• New head (cf91835c9 plus Bound native evidence by bytes, not by the schema item budget #1188's commit, spine and calibration from the same tree): loss 0.00685, 98.5% of 1,168 targets within 10%, ESS 4,809.
  • Six of the 7 gates pass, and so does the certifier's input-mass gate.
  • uk_weight_ratio fails: 1,176 against 1,151, by 2.1%, where main fails by 5.4%. It was 1,155 before the D3 ruling. The old head passed at 1,067 because of the T5 error.
  • obr.council_tax −7.2% (control −9.5%), obr.vat +23.3% (control +24.1%), 108 of 111 UC cells within 10% (control 106).
• So the corrected deltas against the right control are: loss −0.00017, council tax +2.3 points, the weight ratio −37, and the West Midlands cell −1.2 points.
• María's ruling on the breach: this PR proceeds, and main's breach is resolved apart from it.
• Not this tree: Bound native evidence by bytes, not by the schema item budget #1188 changes the evidence reader and its tests only. The scripts are outside the repository beside the licensed data, as for UK: port the unblocked uk-data fixes tracked in #1095 #1100 and UK: port the remaining uk-data fixes tracked in #1095, engine items included #1121. The receipts now cite commits on the branch.

7. W. Reverted. With the T5 fix the West Midlands cell was at +26.9%, and on the final head it is at +25.3%. Both are above the bound, so the deferral is needed and stays. With it restored, uk_target_fit passes. Across this PR's arms the cell ranged from +23.9% to +29.7%. A cell that close to the bound makes both a deferral and its retirement unstable.

8. Signed differences.

• Not enforced: agreed. I ran verify_uk_spine_parity.py on the head's spine. That spine holds household copies the pinned reference does not, so its 40 unsigned differences are not a verdict. The instrument has not run as a gate since the swap.
• The false text at :525. The redraft is ready. Signature text is not committed unsigned, so it needs María's signature to land here.
• D4: a mechanism_change draft for would_claim_pc is added to the drafts (nonzero share 0.386 against the reference's 0.700).
• T5: after the fix the donors do not move, so it needs no entry.

9. D3 reporter rule. Fixed, both parts.

• The share: with item 2's rule no other unit takes the property a reporter does not record. A unit on the sentinel still takes the engine's residual, and the text now says so. Fewer than 10 units are on the sentinel.
• Training screen: an FRS unit that reports UC is now screened on its carrier alone, as uc_capital_coherence records it. Training units rise from 3,991 to 4,029, and to 4,104 with item 2's rule.

10. D4 and Income Support. Not changed: the population is empty by construction. A mixed-age couple meets the Pension Credit age conditions only with the SI 2019/37 saving. A couple that reports Income Support has the saving only if it also reports Pension Credit, and then it is a reporter, anchored as on main. An engine test now pins this.

11. D5's age limit. Not changed, and now explained in the code. The FRS instructions (Benefit Unit Allocation, rule d) make a foster child aged 16 or over a separate benefit unit where a local authority maintenance allowance covers them. So one who stays in the carer's unit has no such allowance. FRS 2024-25 has fewer than 10 such child records.

12. D6 on SPI rows. Fixed. The twin's flag alone turned the residential signal on for 695k weighted SPI-channel people, against 35k FRS-channel people, the intended group. A redrawn SPI row now drops the flag.

13. T2.

• Comparison: fixed. Each person's classes are compared with the change in earned_income_tax, and the measure refuses beyond £1. The receipt keeps the gap to income_tax.
• Diagnostic: test added. A feed-gated test holds that it_relief_total resolves its exact 2024-25 fact (£8.8bn).
• Table 6.1: HMRC's methodology applies income tax rates to employees' pay and models contributions that straddle bands, so the allocation is bracket by bracket, as Restore HMRC salary sacrifice relief calibration targets policyengine-uk-data#533 reads it. The design values are at prior weights.

14. Stale text. Fixed: the parity rationale and its 13 regenerated reasons, docs/uk-income-anchors-280.md, and the changelog counts. The W fragment goes with W.

15. Nits.

• D1's skip: not changed. The root stage always writes the column, and the graph checks its cells. The skip serves partial frames in unit tests, as the start-up refresh beside it does.
• D2 lockstep: back to ages 0 to 100. They all agree.
• T1's annual pence: not changed. The model value is annual.
• Top rate: fixed. A schedule with one rate above the basic class files it as higher.
• 12 to 8: the pinned feed leaves one "No payment" and one "or over" fact for each of four family types. The declared 12 counted eight overlapping "or over" facts from an earlier feed.

Merge order

The description now carries the review's notes on #1145, #1198 and #1167. #1167 landed first, so this branch is the one that regenerates its references, membership and parity receipts on feed 1ee7dfe and measures again, with a control on the new main.

Tests

On cf91835c9:

• the engine-free lane's 534 test files, with policyengine-uk blocked: 16,152 passed and 141 skipped, with the same three environmental failures as before;
• the engine-uk lane's 41 test files: 179 passed.

On 78ea6b695, against the pinned feed: 45 passed, the parity receipts regenerate identically, and the three diagnostics resolve. The D3 commit after it changes no target, reference or parity file.

[juaristi22]

Re-review (Claude Code) at bcbe274b2

Static review of cf42d8a7a..bcbe274b2 against the review and the reply. I read the code and policyengine-uk 2.122.2's source, and ran nothing.

Verdict.

• Fixed: the three main items. D4 now projects the survey-year frame, T5's donor rebase is back on history, and D3 follows the ruling.
• Answered with reasons that hold up: item 10 (Income Support) and item 11 (D5 age).
• No regression, and no test edited to match a count.
• Not ready to leave draft: CI hasn't run on the D3 rewrite, the branch conflicts with main, and every figure is on the pre-Set UK road-fuel and energy levels in the LCFS stage on a calendar-2024 basis (#1113) #1167 base, while the head fails uk_weight_ratio and W sits just over the bound.
• Disclosure: a few disclosure-control items went to the author separately.

Previous items

#	Status	Note
1 D4	Fixed	materialize(..., dataset_period=) is keyword-only with default None, so other callers are unchanged, and a dataset year after the period is refused (policyengine_uk.py:85-109). The stage and the calibration resolver both load the survey year and compute at 2025. The engine test pins the pensioner case.
2 D3	Changed by ruling	The owner is FRS hrpid == 1, checked as exactly one per household (uc_capital_coherence.py:426-467), and property counts at 0.9 (:104). Pension Credit follows the same rule. The gate re-derives with the same functions and refuses a drift in the declared rule. See N3 and N4.
3 T5	Fixed	A declared composed index (hmrc_uprating.py:455-525): household interest to the dataset year, then GDP per head. The 13 amount rows move ×1.0086 against main, and count rows don't move.
4 #1198 hold	Fixed (text)	#1198 holds all four it_relief_* rows, so whichever lands second removes it_relief_total from it.
5 Renewals	Fixed	The description lists them first. Nit: main's W deferral (expiring 2026-11-03) isn't on the list.
6 Measurement	Partly	The control is now on the base, and the #1087 attribution holds: between d50220d1f and a469ebe3d the only first-parent change is #1087. But see N1 and N2.
7 W	Reverted	The register and its test are byte-identical to main.
8 Signed differences	Open, acknowledged	spine_swap_signed_differences.json:513-525 is now stale twice over (carrier and owner rule), and still nothing enforces it.
9 Reporter rule	Fixed	The redraw screens a UC-reporting FRS unit on its carrier alone (uc_reporter_redraw.py:386-399). The only remaining leak is to sentinel units, which is documented.
10 D4 and Income Support	Not changed, justified	In 2.122.2, meets_pension_credit_age_conditions gives a mixed-age couple the Pension Credit route only with the SI 2019/37 saving. That saving needs reported Pension Credit or pension-age Housing Benefit without an income-related benefit, so an Income Support reporter who isn't a Pension Credit reporter can't have it. The engine test pins it.
11 D5 age	Not changed, explained	frs_relationships.py:186-192.
12 D6 on SPI rows	Fixed	spi_income.py:2106-2128, graph.py:1023-1026.
13 T2	Fixed / partly	The per-person earned_income_tax check refuses beyond £1 (measure_simulation.py:741-763), and the top rate is fixed. The it_relief_total test is feed-gated and skips in CI, where the compile still turns a mismatch into unavailable.
14 Stale text	Fixed
15 Nits	Partly	D2 is back to 0–100, the top rate and 12 → 8 are explained, and D1's skip and T1's pence are kept with reasons.

New

Before leaving draft

• N1. CI has not run on the fix round.
  • 887d36de4, 78ea6b695, cf91835c9 and bcbe274b2 have no check runs. The last run was on 7b17c1d20, and the PR has conflicted with main since Set UK road-fuel and energy levels in the LCFS stage on a calendar-2024 basis (#1113) #1167 merged.
  • So the 20-file D3 change has been tested only locally. The rebase will trigger CI anyway, but it's worth knowing that this head has never been through it.
• N2. The merge gate picture is on a stale base.
  • The head fails uk_weight_ratio, and W is just over the bound, both measured on 7639ef8b0.
  • Set UK road-fuel and energy levels in the LCFS stage on a calendar-2024 basis (#1113) #1167 changes the solve: four new measure exclusions including obr.vat, the LCFS stage, and feed 1ee7dfe. So after the rebase W may go stale (the gate then fails the deferral) or stay over, and the weight ratio may move either way.
  • The ruling that this PR proceeds while main's breach is resolved separately exists only as text. No issue tracks main's uk_weight_ratio breach (I searched).
  • Open one and link it, and re-measure on the rebased tree before leaving draft.
• Conflicts with main. These 13 files conflict:
  • the two compile-parity receipts;
  • release_input_coverage_manifest.json;
  • target_reference_membership.json;
  • contract.py and its test;
  • uk_spine.json;
  • test_country_spec.py, test_uk_measure_simulation.py, test_uk_national_calibration.py, test_uk_population_targets.py and test_uk_target_references.py.
• Merged without conflict but inconsistent:
  • calibration_measure_exclusions.json: Set UK road-fuel and energy levels in the LCFS stage on a calendar-2024 basis (#1113) #1167 adds four entries and this PR removes three, so the 1,168 count moves.
  • gates.json: regenerate the contract.py digests from the merged file rather than hand-merging them.

Minor

• N3. The reference-person rule moves pensioners' property away from Pension Credit.
  • The FRS reference person is the householder with the highest income, not necessarily the owner of the other property. Pension-age parents living in a household headed by their adult child now record none; it lands on a working-age unit, where Pension Credit never reads it and UC may.
  • The receipts show Pension Credit property on pension-age units falling by roughly £53bn, while design Pension Credit moves only +£0.03bn.
  • It's a consequence of the ruling, not a bug, but it belongs in the D3 signed-difference draft.
• N4. Sentinel units no longer match recorded units.
  • Units on the −1 sentinel stay on the engine's pro-rata proxies at 100% of gross.
  • The Pension Credit proxy doesn't subtract recorded figures, so a pension-age sentinel unit living with a recording reference person counts the property twice.
  • The engine test's mixed recorded/sentinel case was deleted rather than replaced, and the docstring covers UC only.
  • The population is tiny, but the test is worth keeping.
• N5. The £1 refusal in T2 is absolute, while the engine computes in float32 (measure_simulation.py:755-756). For very high-pay salary-sacrifice earners, float32 rounding approaches that size, so a correct schedule could stop the build. It fails closed and passes on the current arms. Add a tolerance of a few float32 steps on top of the £1.
• N6. Merge order with UK landlords: bind the SPI net property amounts, leave property_wealth to the engine, add receipts and finance costs #1145 on W is a semantic collision git won't flag. UK landlords: bind the SPI net property amounts, leave property_wealth to the engine, add receipts and finance costs #1145 empties the register and pins it empty. If UK landlords: bind the SPI net property amounts, leave property_wealth to the engine, add receipts and finance costs #1145 lands first, this PR inherits an empty register with its cell over the bound, and uk_target_fit fails. The description notes the deletion but not that consequence.
• N7. Recorded property still overstates countable capital. By the receipts' own WAS figures it's about 64% of gross after debt, against the 90% recorded. The only tracking is pe-uk#1969, which is the engine side, so a microcosm issue for the WAS mortgage draw would keep it visible.
• Nits.
  • dataset_period sits outside the RulesEngine Protocol (microcosm-frame/rules.py:80-92).
  • The declared dataset_year_rule: survey_year is never resolved; the code uses uk_time_period(frame).
  • _refresh_landlord_carrier silently skips a missing column (spi_income.py:2118-2119), the same pattern as D1.