[juaristi22]

Summary

The next policyengine-uk bump (plan of 2026-10-22, microcosm#1095 and microcosm#1278). It locks policyengine-uk 2.151.0, the first release past pe-uk#2573 and pe-uk#2580, and writes the inputs those releases read. One PR of focused commits in three phases, then the rows arm 2 re-measures, then Vahid's two reviews (below).

Closes #1340, closes #1320, closes #1222, closes #1349. Part of #1319: Scotland opens in the Carer's Allowance draw, and the back-test and the follow-ups listed there stay open.

Phase A, on the 2.147.2 lock (13 commits; receipts in experiments/1095-uk-next-bump-phase-a-receipts.md)

• the DWP benefit-cap rows bind the UC reduction the award loses (UK: bind the DWP benefit-cap rows to the UC cap the award actually loses #1349);
• the rows held to 2026-11-26: the zero-support floor scopes by unit, rows bind where households carry them, the OBR welfare-cap split and the England postgraduate repayments retire;
• obr.vat bound again, and is held as a diagnostic again since the review (item 10, below);
• HBAI FYE 2025 validation: four statistics inside or outside DWP's 95% intervals (UK scorecard: populace_uk_2023 vs enhanced_frs_2024_25 vs admin benchmarks (2026) #731, report-only);
• the self-employed's missing receipts are signed as an engine-default approximation;
• stage-health gates for the three stages Port the uk-data wave to policyengine-uk 2.134.4: stamp duty, Move to UC, national home values (sixth #1095 PR) #1308 left ungated;
• a marker on the UK outputs that rest on the SPI Public Use Tape (SN 9422);
• the HMRC £1m-and-over cells are held again, since arm 1 measured a distortion; the rows still held are renewed on arm 1's measurement;
• thin local rows' figures are withheld, with the complementary check run over each region's rows of every grain and over each local family with its national and regional parents (review item 1, below);
• every weekly FRS amount annualises at 52 weeks (uk-data#531).

Phase B, the lock (1 commit): policyengine-uk 2.147.2 to 2.151.0, policyengine-core unchanged at 3.33.3. The uprating pins' stamp, the concept coverage (314 to 335 inputs), the coverage manifest, the H2 fixture and APPROVED_UV_LOCK_SHA256 re-derive. The ESA support-group test's legacy tenant no longer claims UC, since 2.149.0 ends income-related ESA on a UC claim (pe-uk#2563). The gambling duty inputs of pe-uk#2590 stay at their defaults (#1359), and so does housing_benefit_residential_capital_exception (2.148.4), signed in UK_ENGINE_DEFAULT_INPUTS: the FRS records no assessment of the accommodation the exception covers, so Housing Benefit is understated only for a qualifying working-age resident, who takes tariff income on capital above £6,000 instead of £10,000.

What the lock alone moves, before any data change (from Vahid's engine-only run on a133f37a):

• income-related ESA falls from £3.85bn to £1.88bn in 2025, as the engine ends income-related ESA on a UC claim (pe-uk#2563). No household of the SPI channel carries that row. Phase C's UC take-up rule keeps a reported award off a UC claim, which offsets it, and DWP's income-related ESA claimant row fits within half a percent;
• Welsh LTT and Scottish LBTT rise by 2.150.0's 1/0.045 factor, into gov_tax, household_tax and net income; no bound row carries them;
• council tax falls under 2.148.5's projection; obr.council_tax binds the total;
• the targeted two-year-old entitlement falls, as 2.148.1 reads UC earnings before the work allowance; dfe.funded_childcare.early_learning_2_year_olds binds the count;
• September CPI uprating (2.148.2) and the Pension Credit floor (2.151.1) move benefits from April 2027; no 2025 row reads them;
• statutory paternity pay is now taxed, inside obr.income_tax, which binds;
• household net income and the poverty rates move a little. They rest in part on the SPI channel's households, so no figure is given here (SN 9422).

Costings and dashboards built on a133f37a, on 2.147.2, need a full rerun on the next release rather than an engine swap.

Phase C, on 2.151.0 (12 commits; receipts in experiments/1095-uk-next-bump-phase-bc-receipts.md)

• the energy spending's price period, energy_price_basis (pe-uk#2571): on a file that declares it, the engine moves each household's electricity, gas and domestic energy spending to a later year by its energy price index from 1 January 2024, every household alike, where an undeclared file keeps its stored spending; an engine test runs it end to end through Microsimulation, and releases built before it stay undeclared;
• the legacy closure draw reads the claimant's and partner's reports (pe-uk#2549);
• the UC take-up redraw keeps a reported income-related ESA award off a UC claim (pe-uk#2563); Housing Benefit keeps the stored flag, pending the HB and JSA PR (UK spine: housing benefit £2.7bn against OBR £12.1bn — pensioner and supported-accommodation HB missing after the UC migration #867, UK spine: JSA claimants 2.8× the DWP caseload — legacy-benefit reporting after the UC migration #869);
• Scotland's carers enter the Carer's Allowance take-up (UK Carer's Allowance: would_claim follows reported receipt, so earnings-limit reforms reach 4,600 new carers against an official 61,000 #1319);
• Widowed Parent's Allowance leaves FRS code 6 by a positive rate rule: an amount at its 2024-25 rate of £148.40 a week, or below it as a reduced award (Split War Widow's Pension (FRS code 9) out of bsp_reported #1222);
• spousal maintenance gets its own column: ALIAMT or ALUAMT, and APAMT (ALLPAY2 is counted in two UK income columns; spousal maintenance column (mirror of uk-data) #1320). ALIAMT is an approximation: the FRS asks for payments "from a previous partner", so it also holds a former cohabitant's;
• bonus sacrifice is imputed to OBR's 2027-28 base at the stage's weights (UK: bonus-sacrificed pension contributions are missing, so the salary-sacrifice cap scores about half the OBR's yield #1340): added beside pay, which stays as it is, and an SPI-channel adult draws its bonus share of pay from a base-FRS adult in its pay cell (María's rulings of 2026-10-24);
• liquid and solid heating fuels are drawn (UK local surface, consumption and high-value homes: remaining items (consolidation) #1278), with the bottled gas in the liquid fuels reported apart from ONS's 04.5.3, which leaves it out; neither is levelled to ONS, so a costing on either needs levelling first;
• the LCFS divisions' excess over their shared ratio is levelled once, each division's drawn spend against the donor's same spend, so the levelled parts are no over-draw (UK local surface, consumption and high-value homes: remaining items (consolidation) #1278; María's ruling of 2026-10-24);
• the WAS share split and the debt secured on other property, netted in the recorded UC and Pension Credit capitals (UK: every open data item: uk-data parity, engine locks, the local surface, consumption and high-value homes #1095);
• England's single adult council tax discount binds at the national grain (UK: supply the engine's local_authority input from the ladder's local_authority_code, plus council tax follow-ups from policyengine-uk#1855/#1864 and uk-data#482/#484 #953);
• an engine test pins the engine's 2025 relabel of student loan plans.

The rows arm 2 re-measures (3 commits, with the Phase B and C receipts)

• four of the twelve UC payment-distribution bands, which UK #1334: draw UC take-up by award band, and measure the HMRC income rows over each person's year #1344 now draws by award band, come within 10 percent unbound and bind; the other eight keep their renewals;
• the four Housing Benefit rows and DWP's JSA claimants are renewed, pending the HB and JSA PR (UK spine: housing benefit £2.7bn against OBR £12.1bn — pensioner and supported-accommodation HB missing after the UC migration #867, UK spine: JSA claimants 2.8× the DWP caseload — legacy-benefit reporting after the UC migration #869);
• the two ONS energy rows' reasons add arm 2's consumption drift block;
• the ORR rail-support row stays held: bound on table 7271, the plan's default on the open 7270-versus-7271 ruling, it fails uk_weight_ratio.

Review

Vahid's first review (round 1, at 74a2a37) and second (round 2, at 3ce1fca). Item numbers are his.

• Item 1, withholding across grains. 3ce1fca withheld every constituency, local-authority and BRMA model figure from what ships. At María's ruling of 2026-10-24 it is reverted (e3c3ecd), with d48652b, its fixture follow-up: the fit evaluation ships. A row that ten or more distinct source households carry publishes its figures at every grain; a thin row, and any published row that would give it back by difference, is withheld. f1cec40 runs the complementary check over overlapping blocks: the national, regional and BRMA rows; each region's rows of every grain; and each local family with the national and regional targets the cross-grain declarations or a nested measurement tie it to. A row withheld in one block is protected in every other, and the blocks run until none withholds another row. BRMA rows publish from ten households under the national block's check (her rulings of 2026-10-17 and 2026-10-24). With the revert, round 2's items 25 to 31 no longer apply.
• Items 3, 4, 8 to 12: the energy declaration's end-to-end test, the Housing Benefit engine default, the bottled gas and the pinned fuel parts, the levelling on remainders, obr.vat held as a diagnostic again, the drift flags carried as warn-only diagnostic_warnings, and WPA's positive rate rule.
• Items 5 to 7: bonus sacrifice added beside pay, the SPI channel's bonus drawn from FRS donors, and the anchor's growth basis stated; its check at the final weights stays with the licensed data.
• Items 13, 15, 17 and the nits: the ALIAMT approximation, the bound rows that absorb the 52-week rule, the benefit cap's scope, the single adult discount count's vintage (as at September 2025, revised on 21 January 2026), schema 3's refusal, the re-signed energy rationale, Scotland's carers' rate, and the support floor's units checked against the contract.
• Arm 3: its receipt and the engine checks for items 7, 9, 14 and 17 are in the Phase B and C receipts. The bonus sacrifice base holds at the stage's weights only: by María's ruling of 2026-10-24 this PR records it, and UK bonus sacrifice: meet OBR's base at the final weights, not the stage's #1390 re-picks the sacrificers at the final weights.
• pe-uk: item 8 is liquid_fuel_consumption: COICOP 04.5.3 does not include LPG policyengine-uk#2642; item 12 is noted on bsp_reported is CPI-uprated but BSP rates are frozen, and the variable is mostly War Widow's Pension policyengine-uk#2381, whose step 3 the split unblocks; item 16 is uprate_student_loan_plans relabels existing holders in 2025 against SLC plan rules (Plan 4, NI Plan 1, Postgraduate Loans, pre-2006 write-off) policyengine-uk#2636, which the student-loan engine test links.

Measurement

All runs are local and licensed; their outputs stay with the spines. Details are in the two receipts.

• Control: b2572-c0main, main at 6bdb800 on policyengine-uk 2.147.2, one spine and one calibration (review date 2026-10-23).
• Arm 2: one spine and one calibration at the bump head. All seven terminal gates pass; 1,265 rows bound against 1,259. Over the 1,258 shared rows with a nonzero target, the share within 10 percent goes from 99.21 to 99.05 percent, the mean absolute error from 0.56 to 0.55 percent, the effective sample size from about 7,400 to 7,500, and the share of design weight stretched 9.5 times or more from 0.67 to 0.59 percent. Two shared rows cross the 10 percent line outward: a row of the SPI channel and a UC band whose figures are withheld.
• Arm 2b: calibrate-only on arm 2's spine with the four UC bands bound (the head's calibration; review date 2026-10-24). All seven gates pass; over the 1,264 shared rows, the share within 10 percent goes from 99.05 to 99.13 percent, and no shared row leaves the band.
• Arm 2c: arm 2b with the ORR row bound on table 7271. The row reaches its target, but uk_weight_ratio fails and the effective sample size falls by about a tenth, so the row stays held on table 7270.
• Arm 3: one spine at the review's head for its stage changes (bonus sacrifice, the levelling, WPA, the fuels) and a calibrate-only run on it. All seven gates pass; against arm 2b the share within 10 percent goes from 99.3 to 99.2 percent, the effective sample size is about the same, and one UC household count leaves the band. Its engine checks are in the Phase B and C receipts, as classes.
• Phase A: arm 1 and arm 1b on Lock policyengine-uk 2.145.0, follow its bus fares and bind thin target rows #1333's 2.145.0 spine, in the Phase A receipts.

Disclosure

No figure in this PR, its commits or its receipts rests on fewer than ten source households or names a record. No row's model figure is given where the row is of the SPI channel or rests in large part or in part on the SPI channel's households: SN 9422's Special Conditions send analyses of the tape to HMRC ten working days before they are distributed, and that ruling is open (#1278). The surface-wide fit and weight statistics are given as in #1308. Publisher figures appear as published. The private checks behind C1, C2 and C10 stay with the licensed data, outside the repository.

Signatures

Signed by María on 2026-10-24:

• obr.vat's entry, held again, 2026-10-24 to 2027-01-24 (review item 10).
• The restated ons.household_energy_expenditure rationale.
• housing_benefit_residential_capital_exception in UK_ENGINE_DEFAULT_INPUTS (review item 4).

For María to confirm:

• The new uprating hold on mhclg.council_tax_stock.single_adult_discount (in-year snapshot, signed 2026-10-23 to 2026-12-11), drafted as the band rows' hold is.
• The Housing Benefit and JSA renewals, signed 2026-10-24 to 2027-01-24, and the eight UC bands' reasons restated on arm 2.
• The two ONS energy rows' reasons and the ORR row's reason and adjudication, each with an arm 2 or arm 2c measurement added.
• The Phase A drafts listed in the Phase A receipts.

Not in this PR

• C8: main's UK: childcare take-up at the spring term, under-2 take-up and single-age targets behind the £100k childcare costing (policyengine-uk 2.147.2) #1338 already fits the childcare rates on DfE's January basis.
• C14: the Welsh fare receipts row is a difference of two facts, and a diagnostic reference holds one; it needs its own design.
• The regional tenure, rent and UC-by-children rows through the cross-grain bridge: no national contract row carries them yet (Place UK private rent and BRMA at the assigned area; bind Scotland's rents at BRMA grain (#1262, #1090) #1305, the local surface).
• The Housing Benefit and JSA repairs (UK spine: housing benefit £2.7bn against OBR £12.1bn — pensioner and supported-accommodation HB missing after the UC migration #867, UK spine: JSA claimants 2.8× the DWP caseload — legacy-benefit reporting after the UC migration #869).

Tests

Counted with --junitxml, on this machine.

• At the review head, only the touched surfaces ran locally: the support floor, the take-up contract and the carers stage (203 tests); the registers, signed exclusions, frame and spec readers (770); the withholding, the gates, the gate kernel, the Orrery summaries, the calibration graph, the sign-off draft and the release readers after the revert (342 and 425). All pass, after one gate test frame gained a household region; two are skipped. CI runs the full suites.
• Before the review: tools/ci_test_plan.py verify ok; tools/spec_engine_coverage.py --check 42,184 of 42,184 fields; the UK engine tier and integration-uk passed at the Phase C head.

🤖 Generated with Claude Code

[juaristi22]

@vahid-ahmadi thanks for both rounds. This replies to both, at 0640d051e. Item numbers are yours. The body now has a Review section and the "what the lock moves" list.

Round 2 (items 25 to 31)

At María's ruling of 2026-10-24, the blanket withholding is reverted: the fit evaluation ships. e3c3ecde0 reverts 3ce1fca6c and its fixture follow-up d48652beb. f1cec40d5 puts back the joint check for item 1 (below).

• 25 to 27, 30 and 31 no longer apply. The gate report, targets.csv, the local gate report and the size evaluation again read one withheld set: the thin rows and their difference partners. The fit gates count the same set, as before 3ce1fca6c. So the sign-off tool's checks hold again, the manifest keeps the signed local exclusions, and the size evaluation sees the red local targets.
• 28, BRMA rows: María's ruling. A BRMA row that ten or more distinct source households carry publishes its figures, under the national block's difference check (her rulings of 2026-10-17 and 2026-10-24). Where a run also carries Scotland's local rows, the BRMA rows sit in Scotland's region block beside them too.
• 29: the fixture patches full_gates.uk_withheld_rows again, which exists again; the file's 15 tests pass.

Round 1

1 (must), withholding across grains. f1cec40d5 makes uk_withheld_row_sets span overlapping blocks:

• the national, regional and BRMA rows, as before;
• each region's rows of every grain: the rows whose positive-weight carriers all live in that region. So a region row less its published local cells is caught, and so is a region's constituencies less its local authorities;
• each local family with its higher-grain parents: the targets that the cross-grain declarations (bridges, fan-out, sum and band bridges, partitions) or a shared measurement with nested filters tie it to. So a national row less a family's published cells is caught too.

A row that one block withholds is protected in every other block that holds it, and the blocks run again until none withholds another row. Local rows need the frame's household region, so the check refuses without it. The dense gates and the Orrery summary pass each row's contract target.

Tests cover a region row less its cells, constituencies less authorities, a national parent less its family's cells, a row withheld by one block and kept hidden in another, and the refusal; each mechanism is mutant-checked. On a synthetic region of 2,218 rows, the check takes under two seconds.

The limit: a higher-grain row that neither a declaration nor a shared measurement ties to a family is not paired with it.

2. The body's Phase B now lists what the lock alone moves. ESA keeps your figure, since no household of the SPI channel carries it. The other items are qualitative, each with the bound row that absorbs it, where there is one. Net income and the poverty rates get no figure (SN 9422).

3. f995397cd: an engine-tier test writes a file, loads it through Microsimulation, and checks each energy column in 2025 and 2026 against the engine's energy price index over its 1 January 2024 value. An undeclared copy keeps its stored spending. The changelog says what the declaration changes, and that releases built before it stay undeclared.

4. ff8dfdd85: housing_benefit_residential_capital_exception joins UK_ENGINE_DEFAULT_INPUTS with the error it can make. The body states it beside the gambling inputs.

5 and 6. 9bc164c6c, on María's rulings of 2026-10-24:

• The sacrifice is added beside pay, which stays as it is. The UGRSPAY comment now cites SN 9563's derived-variable notes.
• An SPI-channel adult draws its bonus share of pay from a base-FRS adult with pay in the same pay cell. The cells are weighted pay quantiles, and a cell with fewer than 20 donor source households joins the one below. The draw is keyed on person_id, and FRS rows keep their own bonus.

7. 865dda866: the stage comment and the changelog now say that the anchor grows amounts with OBR average earnings and the weights with the ONS population, while the OBR grows its base with employment and earnings. The check of both OBR lines at the final weights is a private receipt, and I post its outcome as a threshold class with arm 3's receipt.

8 and nit 20. 94cb840b6:

• The receipt now gives the donor's liquid fuels without bottled gas beside ONS's 04.5.3, with the concept note.
• The column comments warn that a costing on either fuel would come out near twice ONS's spend.
• A test pins each fuel's diary parts, with c45511 in the remainder.

The pe-uk docstring fix is drafted for María to file.

9. a94fc7a05: a division's ratio is now the spend the chain drew for it against the donor's mean of the same spend. For housing and transport, that is the remainder against the donor's total less the same parts. The cut is still one pass, by María's ruling: measured again after each cut, the shared ratio falls until every division sits at the lowest one's. On arm 3's spine, six divisions come down; communication now keeps its draw. The change in total consumption and in the VAT base follows with arm 3's receipt, as threshold classes.

10. 94d4102fd: obr.vat is held as a diagnostic again, signed 2026-10-24 to 2027-01-24. Its reason states the circularity and the private-school residual. It rebinds when pe-uk re-derives the factor on a file that does not bind it.

11. 2ee07b894: each drift flag becomes a warn-only diagnostic_warnings entry (uk_consumption_drift). The entry names the total or share and the threshold, but none of its figures. Nothing is gated on it.

12. 3f120c452: a code-6 amount is now WPA only at its 2024-25 rate (£148.40 a week, within 1%) or below it, as a reduced award. Any other off-rate amount stays bsp_reported. The pe-uk issue on bsp_reported's CPI uprating is drafted for María to file.

13. 7174e821d: the FRS 2024-25 question AliMny asks for payments "from a previous partner", so ALIAMT also holds a former cohabitant's. The survey does not say whether the partner was a spouse or civil partner, so the stage comment and the changelog state the approximation.

14. Open. It needs the engine on arm 3's calibrated file, with and without the secured debt. The result follows with arm 3's receipt, as a threshold class, since UC and Pension Credit rest in part on the SPI channel's households.

15. 32b2ee545 fixes the docstring. The receipt (0640d051e) names the bound rows that absorb the fall: the HMRC SPI income-band rows, the OBR income tax and National Insurance receipts, and the OBR and DWP benefit spending and caseload rows.

16. The pe-uk issue is drafted for María to file. The SLC rows are re-measured on the first arm after the engine fix, and at the latest before the Plan 2 and Plan 5 holds renew on 2027-01-23.

17. The receipt states that Housing Benefit-capped households have no row, and how the complementary check covers the total. The run check follows with arm 3's receipt.

Nits:

• 18. 7174e821d: the feed vendors the revision of 21 January 2026, whose counts are as at September 2025. The row's note, the changelog and the receipt said "October", after the feed's 2025-10 period label.
• 19 and 21. 32b2ee545: the changelog says a schema 3 register is refused, and the restated energy rationale is signed on 2026-10-24.
• 22. e59a53d30: the take-up contract had still applied the rate "outside Scotland", so its citation and population now say Scotland's carers take the same rate.
• 23. The room rule went with item 5. The docstring now says that a person's place in the order is their own, but where the order stops depends on everyone's contributions.
• 24. 8ed7835c1: each scope's unit is now checked against its targets' contract measurement, so the run refuses a mislabelled scope. Band H says households.

Still to come

Arm 3's calibration receipt. Arm 3 is one spine at the review's stage changes plus a calibrate-only run, and its receipt brings items 7, 9, 14 and 17's run checks.

🤖 Generated with Claude Code

[vahid-ahmadi]

@juaristi22 thanks, this is my third review, at f3bd0b42b.

Verdict. No must is left open. The revert and the joint check across grains are a better answer to item 1 than the blanket withholding was. The local fit evaluation ships again, and items 25 to 31 go with the revert. From round 2, I would have treated 25 as release-blocking for the local line. Blanket withholding would have been acceptable to us if the body and changelog said so. Neither point matters now. In the new check, I could not fault the overlapping blocks, the protection carried from one block to another, or the fixpoint loop, and I enumerated the contract's local families to check the limit you state. Two new shoulds: the region blocks may be slow at real scale (32), and the family check fails open when a row has no contract target (33). Items 7, 9, 14 and 17 wait for arm 3's receipt. The other round 1 items are resolved. From our side, this can merge once CI is green at the head, with 33 fixed or accepted. Do you plan to land arm 3's receipt in this PR, or in a follow-up? Either works for me if the body says which. CI at 0640d051e had engine-uk red on the H2 parity test. At f3bd0b42b that test passes locally, and CI was still running when I posted.

What I ran

• CI at 0640d051e: lint, wheels, select-countries and integration-uk passed. Both engine-uk jobs failed on test_acceptance_h_parity.py::test_h2_uk_spine_parity (1 failed, 255 passed): the H2 golden's frs_person_draws stage_contract_sha256 did not match the head's. At f3bd0b42b, lint, wheels, select-countries and integration-uk pass; both engine-free jobs, both engine-uk jobs and both engine-us jobs were still running when I posted. tools/ci_test_plan.py verify: ok. The head merges cleanly with main (git merge-tree; main is 83 commits ahead).
• Engine-free, with --extra uk, at 0640d051e: the diagnostics, full gates, graph terminal, full calibration graph, Orrery contract, dense sign-off draft, salary sacrifice, LCFS, support floor, FRS spine, measure simulation, size evaluation and both release assembler files. 839 passed. At f3bd0b42b: the H2 parity test, LCFS and diagnostics files, 178 passed.
• Engine: test_uk_energy_price_basis.py, 3 passed.
• I ran uk_local_family_parents() on the committed contract: 435 contract targets, 59 local (36 local authority only, 18 at both grains, 4 constituency only, 1 BRMA), forming 34 families. 27 have higher-grain parents. For the 7 without, I looked for national or regional rows of the same measurement (item 33 has the detail).
• A timing of numpy.linalg.eigh on dense symmetric matrices of the block sizes in item 32, on my laptop.
• Primary source: DWP, Benefit and pension rates 2024 to 2025, for Widowed Parent's Allowance (£139.10 in 2023-24, £148.40 in 2024-25), which matches WPA_WEEKLY_RATE.
• Not run: no build of any kind and no data, so the joint check has not been run on a real local surface. No £ figures from the frame appear here, so your SN 9422 rule is not engaged.

Status of round 2

25. Sign-off tool refused every local build: resolved by the revert. The kernel writes the unredacted gate report again, uk_published_fit_gate_report and the uncertified-local sets are gone, and test_uk_dense_signoff_draft.py passes.
26. Signed local exclusions left the manifest: resolved by the revert. reviewed_exclusions is back in the details the assembler reads.
27. Size evaluation's red ledger: resolved by the revert.
28. BRMA withheld everywhere: resolved by your ruling, now stated in the body. BRMA rows publish from ten households under the national block's check. Where a run carries Scotland's local rows, they sit in Scotland's region block too, since all their carriers live there.
29. engine-free fixture: resolved. test_uk_full_calibration_graph.py:67-68 patches uk_withheld_rows, which exists again.
    30, 31. No longer apply.

Status of round 1

1. Must, withholding across grains: resolved in code, checked on synthetic tests only. _row_blocks (diagnostics.py:1127) now adds one block per region, holding every row whose positive-weight carriers all live in that region (_region_blocks). It also adds one block per local family with its higher-grain parents (_family_blocks, uk_local_family_parents). Rows withheld beside a thin row in one block get an infinite recovery floor in every other block that holds them (_withheld_in_block), and blocks re-run until nothing changes. Rows are only ever added, so the loop ends. The checks I made:

   • local clones keep their donor's region (constrain_to_region defaults to true in rowwise_dataset.py), so region blocks do hold the local rows;
   • Scotland's BRMA rows join Scotland's block beside its local-authority rows, which also covers the case where a local-authority private rent cell overlaps a BRMA;
   • a national row of one nation, such as Scotland's band rows, falls in that nation's region block;
   • the five new tests cover the region-less-cells, constituencies-less-authorities and national-less-family cases, plus protection across blocks and the refusal without regions.

   The limit you state (an untied higher-grain row) does not bite on today's contract, as far as I can see (item 33).

2. What the lock moves: resolved. The body's Phase B list covers each engine change and the bound row that absorbs it, and gives a figure only for ESA.

3. energy_price_basis: resolved. The changelog says what the declaration changes, and the new engine test loads a written file through Microsimulation and checks 2025 and 2026. It passes locally.

4. housing_benefit_residential_capital_exception: resolved. It is in UK_ENGINE_DEFAULT_INPUTS with the error it can make, and the body states it.

5. Bonus sacrifice and pay: resolved by your ruling. The sacrifice is added beside pay, and pay is unchanged (salary_sacrifice.py, BONUS_SACRIFICE_MOVE). The candidate test now reads pay itself against the limit, and the UGRSPAY comment cites SN 9563.

6. SPI rows' bonus: resolved. draw_spi_bonuses gives each paid SPI-channel adult a bonus share of pay drawn from base-FRS adults in its weighted pay quantile cell. A cell with fewer than 20 donor source households joins the one below, and the draw is keyed on person_id. FRS rows keep their own bonus.

7. Bonus anchor after calibration: partly. The comment and changelog now say how the anchor and the OBR grow the base. The check of both OBR lines at the final weights comes with arm 3's receipt.

8. LPG and fuel levels: resolved here, but the pe-uk docstring fix is not yet filed. The receipt reports bottled gas apart from ONS 04.5.3, and f3bd0b42b now says the frame's draws reach neither ONS nor the donor, so a costing needs levelling to ONS first. That wording is better than the "near twice" it replaces. Please link the pe-uk issue or PR once filed.

9. Division levelling: partly. Each division's ratio is now the spend the chain drew against the donor's same spend (lcfs_donor_drawn_means), so levelled parts no longer count as over-draw. The single pass is your ruling, with its reason in the comment. The change in total consumption and in the VAT base comes with arm 3's receipt.

10. obr.vat: resolved. It is held as a diagnostic, signed 2026-10-24 to 2027-01-24, with the circularity and the private-school residual in its reason.

11. Drift flags: resolved. uk_consumption_drift_warnings feeds the diagnostics payload's diagnostic_warnings (diagnostics.py:2793) as warn-only entries without figures.

12. WPA rule: resolved. A code-6 amount is now WPA only at the 2024-25 rate within 1%, or below it; the rate matches DWP's table. The pe-uk issue on bsp_reported's CPI uprating is still to be filed. See nit 34.

13. Spousal maintenance: resolved. The approximation (ALIAMT holds a former cohabitant's payments) is stated in the stage comment and changelog.

14. Secured debt netting: open, waiting for arm 3.

15. 52-week rule: resolved. The docstring now says 52, and the receipt names the bound rows that absorb the fall.

16. Student-loan pin: resolved. The test links policyengine-uk#2636, and your reply says when the SLC rows are re-measured.

17. Benefit-cap total: partly. The receipt states that Housing Benefit-capped households have no row. The run check comes with arm 3.

18 to 24 (nits): resolved. The single adult discount note now says September 2025, revision of 21 January 2026 (18). The changelog notes schema 3's refusal (19). A test pins the fuel parts (20). The energy rationale is re-signed 2026-10-24 (21). The Scotland take-up reason is fixed (22). The room rule went with item 5 and the order docstring is right (23). Each support-floor scope's unit is checked against its targets' contract (24).

New findings in the delta

Should

32. The region blocks may be slow at real scale (diagnostics.py, _region_blocks and _small_recovery). A region block holds every local row in the region. Counting local_target_references.json against local_area_crosswalk.json, the South East has 3,755 local rows (64 local authorities, 91 constituencies) before its own region rows, against the 2,218 of your timing. _small_recovery takes a dense eigh of the published rows' Gram matrix on every pass, and _withheld_in_block makes one pass per row it withholds, plus one. The fixpoint loop then re-runs any block that shares a newly withheld row. On my laptop, eigh takes about 3 seconds at 3,760 rows, so a region with a hundred complementary withholds takes several minutes, more with re-runs, across 12 region blocks. The set is also computed separately in full_gates.py:248, in the Orrery summary (orrery_contract.py:66), and in write_uk_calibration_diagnostics from full_build_cli.py:3160, which passes no withheld. Please time it on the first real local surface before the next local release candidate. And compute the set once per build, then pass it to the other callers, as graph_terminal.py already does.
33. The family check fails open when a row has no contract target (_family_blocks). Without regions, the check refuses (_row_blocks). But a row with no metadata.contract_target_id silently joins no family block. test_a_national_row_less_a_familys_published_cells_withholds_one_more asserts exactly that: without targets, nothing is withheld beside the thin cell. The full gates and the Orrery summary pass the id explicitly. full_build_cli.py:3160 relies on the shared payload's rows carrying it from target.metadata, which the reference authoring does set, but nothing enforces it. Please refuse a local row without its contract target. A parent the build does not bind is not published, so it needs no refusal. On today's contract the limit you state looks empty. The 7 parentless families are the two DESNZ kWh families, the three equivalised income and housing cost families, and the two private rent families. I found no national row of the same measurement for any of them: the national energy rows are spend, not kWh. A contract test asserting that would keep it so as rows are added.

Nit

34. A WPA award with inherited additional pension above the base rate now reads as bsp_reported, untaxed and not means-tested, as the comment says. A private count of code-6 amounts above £148.40 × 1.01 that match no BSP rate would show whether that is rare.

What looks right

• The revert brings back one withheld set for the gate report, targets.csv, the local gate report and the size evaluation, so the gates count what the writers withhold.
• Protection across blocks is conservative in the right direction: any combination holding a row another block withheld counts as too small, so a block cannot hand back what another hid.
• The region refusal fails closed, and region blocks take any concept, so a region row of one family less another family's cells is covered too.
• The fuel wording in f3bd0b42b is more accurate than before, and the frame's fuel figures stay with the licensed data.
• The WPA rule now needs a positive match, and the tests pin the edge cases, including the lump-sum case it cannot tell apart.

Merge order

Unchanged: this PR, then #1381 (stacked on this branch, still draft), then my #1373. Both conflict with this head, mostly in generated files (the coverage manifest, source_stages.json, sources.yaml and the H2 fixture); #1373 also conflicts in contract.py and its test. Both need a rebase once this lands.

🤖 Generated with Claude Code

[juaristi22]

@vahid-ahmadi arm 3's results, which my last reply said would follow. Arm 3 is one spine at the review's stage changes (b1b315ae0) and a calibrate-only run on it, against arm 2b. The receipt is in experiments/1095-uk-next-bump-phase-bc-receipts.md (e570d65ff). The engine figures rest in part on the SPI channel's households, so they are given as classes.

• Fit. All seven terminal gates pass. The run binds 1,268 rows, which is arm 2b's rows less obr.vat. The share of rows within 10 percent is 99.3 percent in arm 2b and 99.2 percent in arm 3. The effective sample size is about the same, no shared row moves by more than 5 points, and one UC household count leaves the 10 percent band.
• Item 7. At the final weights, the bonus base above £2,000 at 2027-28 falls short of OBR's line by more than a quarter, while the stage meets it at its own weights. So your concern holds: the anchor does not survive calibration. The salary line comes out within 25 percent of OBR's. Calibration to the HMRC income-band rows lowers the weights of the highest earners, where the sacrificers sit, and no row binds the base. At María's ruling of 2026-10-24, this PR records it, and UK bonus sacrifice: meet OBR's base at the final weights, not the stage's #1390 re-picks the sacrificers at the final weights.
• Item 9. Against arm 2b, total consumption rises by under 3 percent and VAT by under 5 percent. Both changes mix the levelling on remainders, the bonus no longer taken out of pay, and obr.vat leaving the objective.
• Item 14. Netting the secured debt raises UC spend by under 0.1 percent and Pension Credit spend by under 1 percent at 2025. It takes many benefit units' assessed capital below £6,000 or £16,000, but few of those units are entitled either way.
• Item 17. On the run, the thin bands are withheld, and the band that no household carries leaves the solve under the zero-support floor. The complementary check finds no published combination that gives a thin band back, so it withholds no further row.
• A correction to my last reply (item 8). I wrote that the column comments warn a costing on either fuel would come out near twice ONS's spend. That is the donor's level at diary prices, not the frame's. f3bd0b42b corrects the comments, the notes, the changelog and the receipt: the stage draws both fuels without levelling, and the frame's draws reach neither the donor's totals nor ONS's, so a costing on either column needs levelling to ONS first.
• pe-uk. Item 8 is liquid_fuel_consumption: COICOP 04.5.3 does not include LPG policyengine-uk#2642, and item 12 is noted on bsp_reported is CPI-uprated but BSP rates are frozen, and the variable is mostly War Widow's Pension policyengine-uk#2381. For item 16, your uprate_student_loan_plans relabels existing holders in 2025 against SLC plan rules (Plan 4, NI Plan 1, Postgraduate Loans, pre-2006 write-off) policyengine-uk#2636 is now linked from the student-loan test.

[juaristi22]

@vahid-ahmadi #1383 merged with items 32 to 34 of your third review open, so they're in #1381, now rebased onto main (98a09827a).

• 33. The withholding check now refuses a local row that names no contract target, instead of leaving it out of every family block. A contract test keeps the parentless local families free of a higher-grain row of their measurement. Those families are the two DESNZ kWh, the three equivalised income and housing cost, and the two private rent families.
• 32. A build now computes the withheld rows once.
  • The national path's diagnostics take the set that the national gates flagged in their calibration evidence.
  • The Orrery result summary takes the set recorded by a gate on the same result. It counts again only if that gate did not see every summary row.
  • A local surface's check logs its blocks, block checks and seconds, so the first real local surface gives the timing you asked for.
• 34. On the FRS 2024-25, fewer than ten code-6 amounts sit above Widowed Parent's Allowance's rate at no Bereavement Support Payment rate, so the case is rare. The stage comment says so.

Tests on the touched surfaces pass: the withholding, gates, Orrery and build CLI files, 631 of 631.

🤖 Generated with Claude Code