UK: policyengine-uk 2.130.1 and the fourth batch of #1095 ports, with the release's weight year, identifiers and birth months fixed

[juaristi22]

This is the fourth #1095 ports PR. It moves the engine lock to policyengine-uk 2.130.1, ports every item on #1095 that waits for no later engine release, fixes the release's weight year (#796 A19) and identifiers (#1194), and stores each person's birth month in the spine so the release reads the dates of birth its own solve fitted.

All 7 calibrated-seam gates pass on the head, with dwp.uc.households_children_5_or_more held out until 26 November on María's ruling (below). The head's calibration writes a release file, and the weight year, the identifiers and the stored birth months are checked on it.

Before any national build: the renewals of the measure exclusions that lapse from 15 October. After merge, as for the earlier ports PRs: the signed-difference entries for the new and changed columns, whose drafts stay outside the repository until signed, then the national build and the certifier.

Summary

Engine (B1–B3)

• The lock moves from policyengine-uk 2.123.0 to 2.130.1. The branch first locked 2.127.0. 2.130.1 is the first release published after it (2.128.0 to 2.130.0 never reached PyPI), and María ruled to take it here. 2.131.0, published at 12:29 UTC on 14 October (trading losses), is left for the next bump. policyengine-core moves to 3.33.3 on the UK side, which is what the engine's own floor resolves, and stays at 3.32.5 on the US side; a test pins both.
• Nothing the UK build writes is renamed, removed or re-typed. Six parts of the build follow what the engine now returns:
  • Gift Aid extends the basic rate band instead of reducing allowances (Give Gift Aid relief by raising the rate limits, not by deducting the donation policyengine-uk#2271). The CGT taxable-income measure follows. So does the salary-sacrifice relief by rate class, which now extends a donor's bands by the grossed-up gift as earned_income_tax does (ITA 2007 s.414(2)(b)): on 2.130.1 the first head arm refused at that measure until it did. The three relief rows declare the extension;
  • the universal-only childcare row leaves the extended entitlement out itself (Preserve universal and targeted hours alongside extended childcare policyengine-uk#2177);
  • the UC take-up lockstep test names a pension-age family moved onto Universal Credit at the legacy closure (Move families onto Universal Credit when a legacy benefit they receive closes policyengine-uk#2119);
  • the concept mapping declares would_claim_uc_at_legacy_closure left at the engine's default, true: Move legacy-benefit reporters onto Universal Credit at Move to UC claim rates policyengine-uk-data#534 draws it at DWP's Move to UC claim rates, which are not yet a Chronicle fact, so the 2025 Universal Credit rows rest on that default;
  • the liquid-asset reason names the company-capital inputs of Universal Credit's regulation 77, which default to no company.
• The other changes the review named need nothing from the build, and the bump arm measures them: statutory sick, maternity and paternity pay as Universal Credit earned income (Count statutory sick, maternity and paternity pay as Universal Credit earned income policyengine-uk#2365), the benefit cap's exceptions (2.129.0) and its earnings test with statutory pay (Fix benefit-cap statutory-pay earnings exemption policyengine-uk#2399), and income ESA and JSA ending in remunerative work (2.129.1). Capital gains get their own projection path in 2.130.0, but with no scenario passed it copies GDP per head exactly, so the build's gains are projected as before.
• Re-derived with no change to data: the uprating pins, the concept coverage (288 inputs to 299) and its mapping, the road-fuel ratios in the donor-uprating audit, the H2 fixture, the coverage manifest's digests and the approved lock hash. The target references and the compile-parity receipt regenerate identically.

Spine fixes (R1–R4)

Two of these port merged incumbent code; the others go first, so a divergence shows if the incumbent lands differently.

• R1. FRS EMPSTATI 11 is OTHER_INACTIVE. Ports Map FRS EMPSTATI 11 to OTHER_INACTIVE, not LONG_TERM_DISABLED policyengine-uk-data#526 (merged). These are Max Ghenis's commits from Map FRS EMPSTATI 11 to OTHER_INACTIVE and refuse unknown adult codes (uk-data#526 parity) #1097, cherry-picked with their authorship: the mapping, one label per code, and a refusal for a blank or unknown adult code. Map FRS EMPSTATI 11 to OTHER_INACTIVE and refuse unknown adult codes (uk-data#526 parity) #1097's experiments/ folder is not carried.
• R2. Approved training. Goes first: Read approved training from the FRS TRAINEE flag for dependants policyengine-uk-data#561 is an open PR, and the rule is its rule. is_in_approved_training is read from the FRS TRAINEE flag on child-table records. The stage read the adult table's scheme codes, which flagged adults on employment schemes and no dependant, so dependants aged 16 to 19 in training failed the qualifying-young-person test.
• R3. Statutory paternity pay. Goes first: Construct statutory paternity pay from the FRS (sppadj) policyengine-uk-data#575 is an open issue. It is set from the FRS adult table's SPPADJ, beside sick and maternity pay; nothing set it before. On 2.130.1 the engine counts it as Universal Credit earned income and in household benefits and HBAI income, so it raises the net income of the households that report it (on 2.127.0 it was only a National Insurance base and lowered it). Whether SPPADJ times the weeks in a year is the right annual amount for a payment of at most two weeks is still open (calibration_diagnostics.json has no declared schema and is not validated on write #575). Fewer than 10 source households report it, so no weighted figure is given anywhere.
• R4. Salary sacrifice within pay. Ports Limit salary sacrifice to the pay it comes out of policyengine-uk-data#518 (merged on 14 October). Every amount is limited to the pay it comes out of, the forest trains on respondents with pay, and the headcount conversion moves at most half the pay. The stage refuses a spine where any sacrifice exceeds the pay that remains.

Targets (T1)

• T1. The national 15–19 age bands split into 15, 16–17 and 18–19 by sex. Goes first: Calibration leaves the teenage age mix and education status without direct targets: 35% too many 16-year-olds, half the 18- and 19-year-olds policyengine-uk-data#563 is an open issue. A fit to a five-year band leaves the ages inside it free, and 16 and 18 are where the child and young-person rules change.
  • Measured first on main's file, which fits each band within 0.1%: it holds 19% too many females and 14% too many males aged 16–17, and 18% and 28% too few aged 18–19.
  • On the head the six rows fit within 0.1%.
  • The contract has 400 targets (396) and 1,283 active references (1,279).

Export (E1–E3) and the birth months (M)

• E1. Weights at the dataset's year (WS-E spine deferrals: known bugs and signed inconsistencies awaiting their increment #796 A19). Goes first: Runtime calibration-year weights are 0.44% below the calibrated ones (base-year materialization and policyengine-uk's uprating use different weight indices) policyengine-uk-data#538 is the incumbent's open issue. policyengine-uk multiplies household_weight by ONS population growth for each year after the dataset's own, so weights solved at 2025 and stored on the 2024-25 dataset read 0.72% high at 2025 on every weighted count and total. The release write now divides by the engine's own factor. The rule-1 scorer reads each side with the weights the engine gives it at the scored year, so the incumbent's estimates are read 0.72% higher than before. Scores are not comparable with earlier ones.
• E2. Identifiers inside 32 bits (UK clone stages can number identifiers past the 32-bit range policyengine-uk stores #1194). Goes first: OA clones number identifiers past the 32-bit range policyengine-uk stores policyengine-uk-data#558 is an open issue. The clone stages number copies past 2,147,483,647 and the engine stores identifiers as 32-bit integers, so on main's file more than half of them wrap. The release boundary renumbers, one plus the rank of the build identifier per entity, and keeps the build identifiers as person_build_id, benunit_build_id and household_build_id.
• E3. The calibration measures the engine on the release's identifiers. Microcosm only. Measured on 2.127.0, E2 alone made the release differ from its solve on two-child-limit and Universal Credit award rows, because the engine dealt birth months by a hash of person_id.
• M. The birth months are stored in the spine. Microcosm only; María asked for it in this PR. policyengine-uk places each date of birth from the age and months_since_last_birthday, and left unset it spreads each age and sex over the year by weight, in an order set by a hash of person_id. So a release, at the calibrated weights, read different dates of birth from those its solve fitted at the spine's weights, on main too: the two-child limit, the State Pension scheme and the Pension Credit qualifying age depend on them. The root stage now draws the months once, the engine's way, at the FRS weights, and every later read takes the stored value.
  • Read as the engine reads it, a release file without M differs from its own solve on 55 of 1,217 rows, 8 of them by more than a percentage point; the State Pension rows by scheme and amount move most. With M, 3 rows differ, and no State Pension, two-child-limit or Pension Credit row moves. What is left (private-school pupils and two land rows) depends on the weights themselves, and is main's too.

The row held out (X)

dwp.uc.households_children_5_or_more is held out of the calibration until 26 November 2026, on María's ruling. DWP counts about 21,000 more Universal Credit households with five or more children than its two-child-limit counts of families with five and six or more. On main the frame tells the two counts apart only through fewer than 10 source households, where a dependant in training counted as a child in the first row but not in the two-child-limit rows. R2 fixes that, the two counts then agree on every family, and the calibration cannot meet both. The two-child-limit rows stay bound. With the row held out, the head passes all 7 gates.

Measurement

Four arms on feed 942a1fa, each a spine build and a national calibration, all with --review-date 2026-10-14 for measurement only. Effective sample sizes are not given.

• control (main 65d7d153a, policyengine-uk 2.123.0): loss 0.00583; 98.8% of 1,214 rows within 10%; all 7 calibrated-seam gates pass.
• bump (main plus only the three lock commits, policyengine-uk 2.130.1; measurement only, not in this PR): loss 0.00593; 98.6% of 1,214; all 7 gates pass. Before calibration 169 rows move by more than 1%. On the rows the review named, at the spine's weights: Universal Credit +0.4%, 0.8% fewer capped households, ESA −3.0% with 5.1% fewer income-related claimants; Pension Credit is +6.5% and Carer's Allowance −5.5%, as on 2.127.0.
• head without M (the head less the stored birth months; measurement only): loss 0.00588; 98.7% of 1,217; all 7 gates pass.
• head (1da8d681f, the branch before its rebases): loss 0.00590; 98.8% of 1,217; all 7 gates pass. The input-mass gate passes on its calibrated weights.

On the current main. The branch is now rebased onto main 121fd4dd6 (#1256, #1279). #1256 draws the SPI channel's incomes afresh, which changes the spine, so a control and the head were rebuilt on it:

• control (121fd4dd6): loss 0.00584; 98.4% of 1,214 rows within 10%; all 7 calibrated-seam gates pass.
• head (0912388b6): loss 0.00599; 98.6% of 1,217; all 7 gates pass. On its written file the weight year, the identifiers and the stored birth months check out as before, the input-mass gate passes, and the file differs from its own solve on 3 rows (private-school pupils and two land rows).

The decomposition into the bump, the spine fixes and the birth months was measured on 65d7d153a and is not repeated. Rebasing onto #1242 changed one test, which now tells this PR's 2026-10-14 exclusion apart from #1242's housing-element renewal of the same date.

Receipts: experiments/1095-uk-data-ports-4-receipts.md. Survey support is given in bands of distinct source households (<10, 10–29, 30+). A cell under 10 has no weighted figure, and row figures are given only for rows with 30+ households behind them.

Not in this PR

• Additional-dwelling purchases and obr.sdlt: they wait for HMRC's stamp tax tables in Chronicle and for Stop charging property stocks as purchases when a household buys its home policyengine-uk#2241 and Update corporate SDLT and business rates statistics and project them with OBR receipts policyengine-uk#2322. The receipts measure household SDLT by part.
• DWP's Move to UC claim rates in Chronicle, for would_claim_uc_at_legacy_closure.
• policyengine-uk 2.131.0.
• The sparse-band support floor, which is its own PR.
• The renewals of the measure exclusions that lapse from 15 October.

Tests

• CI's engine-uk job list on 1da8d681f: 205 tests, all pass.
• CI's engine-free job list on 1da8d681f: 17,475 tests. 17,318 pass and 155 skip; two fail on this machine only (a US scorer test that needs Python 3.14.7, and a test that reads the live Hugging Face pointer).
• The feed-gated tests on feed 942a1fa, on 1da8d681f: 99 tests, 97 pass and 2 skip.
• tools/ci_test_plan.py verify, the graph burndown verify, the pins tool's --check, the concept-coverage --check and ruff check . pass.
• These ran on 1da8d681f, before the rebases onto UK local release: nation-grain doctrine at K=15 and 2,000 epochs, a 100,000-household release path, line-scoped gate exclusions, and the QRF tail surface #1242 and Draw SPI-channel incomes at the FRS twin's conditional rank (#840) #1256. On the rebased head 0912388b6: CI's engine-uk job list, 207 tests, all pass; the engine-free test files the rebases or the last commits touch, Draw SPI-channel incomes at the FRS twin's conditional rank (#840) #1256's SPI tests among them, 966 tests, 964 pass and two skip.
• The receipts commit on top, fcb6ff23e, adds only the receipts file.

🤖 Generated with Claude Code

[vahid-ahmadi]

@juaristi22

Needs changes before merge. The ports themselves check out and every source change is pinned by a test that fails when it is reverted. Two things block: the receipts commit effective sample sizes, and the head fails its own uk_target_fit gate until you rule on the row. I also think the 2.127.0 lock deserves a decision before merge, because the engine releases after it fix rules this build calibrates on.

1. must. experiments/1095-uk-data-ports-4-receipts.md:62, 64, 65, 67, 70, 123, 174 give each arm's effective sample size. That is on the never-publish list for this microdata. Please remove them or replace them with a band or a ratio to the control. The PR body correctly leaves them out, so the receipts should match.

2. should: the 2.127.0 lock misses engine fixes that reach this calibration. policyengine-uk 2.130.1 has been on PyPI since 01:58 UTC on 14 Oct. The 2.128.x to 2.130.0 releases never reached PyPI because the release job failed on the benefit-cap earnings property from 12 Oct, so 2.130.1 is the first published version after 2.127.0. I ran synthetic households on 2.127.0 (the lock) and on 2.130.1:

   • Benefit cap earnings test (pe-uk#2399, 2.130.1). In 2.127.0, is_benefit_cap_exempt_earnings takes pay less all income tax and NI, so it deducts the tax and NI on statutory pay without counting the pay. Take a London lone parent with three children, £10,600 pay and £4,000 SSP. On 2.127.0 she loses the exemption and is capped by about £10,250 a year. On 2.130.1 she is exempt. The spine stores SSP and SMP, and dwp.benefit_cap.capped_households and the up-to-£100 band stay in the objective (calibration_measure_exclusions.json). So part of the fit to DWP's capped count comes from families that 2.127.0 caps wrongly. Main's 2.123.0 has the same formula, so this PR does not make it worse. But it is the fix that unblocked the release, and this PR is the point where the lock moves.
   • SSP and SMP in Universal Credit (pe-uk#2365, 2.128.0). In 2.127.0 the NI and tax on statutory pay come off UC earnings but the pay itself is not counted. In the same household with £13,000 pay, adding £3,000 SSP raises UC by about £460 on 2.127.0 and lowers it by about £1,190 on 2.130.1. That affects every SSP and SMP record on the spine and the UC rows they feed.
   • 2.129.0 rewrites the cap's ESA and disability exceptions (claimant and partner only, ESA support component). 2.128.0 moves legacy tax-credit families onto UC from 2025-26 (pe-uk#2119), and later versions uprate capital gains by their own index instead of GDP per head. All three will move the UC, tax-credit, benefit-cap and CGT rows, so a bump needs its own arm, not a quiet lock edit.

   Which way to go is your scope call. One option is to move to 2.130.1 here and re-derive the pins, concept coverage, H2 fixture, lock hash and arms. The other is to merge at 2.127.0 and say in the body and the changelog fragment that the benefit-cap and UC rows are calibrated on an engine with these known gaps, with 2.130.1 as the next bump. Either way, the body's "the latest release on PyPI" (Summary, Engine) is no longer true and should change.

3. should: R3 lowers the income of the households it touches on the locked engine. In 2.127.0, statutory_paternity_pay is only a Class 1 NI base. It is not in household_benefits, HBAI income or UC earned income. Those were added in 2.128.0 by pe-uk#2365, which is not an ancestor of the 2.127.0 tag. In a synthetic household, £5,000 of SPP raises NI by £400 and lowers both household_net_income and hbai_household_net_income by £400. On 2.130.1 the same SPP raises net income by £4,600. So at frs_spine.py:700, the new column moves these households' income the wrong way until the engine catches up. Support is under 10 households, so the effect is small, but the column ships in the release. Either hold R3 until the lock reaches 2.128.0 or later, or say this in changelog.d/1095-statutory-paternity-pay.added.md. The fragment says now which engine variables read SPP, but not that it reduces net income. Separately, uk-data#575 leaves open whether SPPADJ × 52 is the right annual amount for a payment that lasts at most two weeks. Matching SSP and SMP is a fair default, but the receipts should record that the question is still open.

4. should: the failing uk_target_fit row needs a recorded ruling, not only an arm. Until you rule, the head cannot write a release file. My view, as a calibration judgement rather than a code point: the R2 fix is right, and it matches uk-data#561 (trainee == 1 on a child-table record is the same rule as is_trainee & ~is_adult_record). Under 10 source households were what separated DWP's two counts, so the frame has no real support for the roughly 21,000 large UC families outside the two-child limit. That makes holding dwp.uc.households_children_5_or_more as a diagnostic defensible, as is a lower loss weight. Whichever you choose, it should go into calibration_measure_exclusions.json with a reason and an expiry in this PR, so that the head passes its gates as merged rather than only in the measurement-only local commit.

5. nit. The body calls every R and E item a port, but only uk-data#526 and Rebase stale HT2 taxable-interest rows onto the Pub 1304 national actual (#489 adjudication) #518 are merged. Finish #515: per-record E19200 decomposition by published SOI 2.1 AGI-band components #561 is an open PR, and calibration_diagnostics.json has no declared schema and is not validated on write #575, Staging telemetry silently disabled since 2026-07-02: empty POPULACE_STAGING_REPO_ID defeats the on-by-default staging repo #563, QBI v3 wiring: simulation consumes SCF/SOI/JCT evidence (shape from SCF, level from JCT) #538 and Read HMRC's joint distribution of capital gains by taxable income #558 are open issues. It would help to say which items port merged incumbent code and which go first, so any divergence is visible if the incumbent lands differently.

6. nit. The test counts differ between the body and the receipts. The body has the engine-free job list as 17,384 tests on 9629176f0. The receipts (Tests) have 17,381 on ffa13ec29.

What I checked

• Worktree at ed33feb76 with a fresh uv venv (uv sync --all-packages --locked --extra uk, which resolves policyengine-uk 2.127.0).
• ruff check . passes. ruff format --check would reformat 45 files, none of them touched here, and CI does not run it. tools/ci_test_plan.py verify and tools/graph_acceptance_burndown.py --verify pass.
• All 29 touched test files, engine-free and engine-uk, pass. Five tests skip because the Chronicle feed is not present here.
• Reverts and mutations, each of which fails the new tests:
  • R1: code 11 back to LONG_TERM_DISABLED.
  • R2: the base frs_education.py.
  • R3: the base frs_spine.py.
  • R4: three mutations (the pay limit dropped, the half-pay cap dropped, and training on unpaid respondents).
  • E1: the factor set to 1, and the full-build write without the division, which test_uk_rowwise_national_role catches.
  • E2: no renumbering at the export.
  • E3: the scratch file left on build identifiers, which the engine birth-month test catches.
• R4 against the merged uk-data#518 diff, and R2 against the Finish #515: per-record E19200 decomposition by published SOI 2.1 AGI-band components #561 diff.
• policyengine-uk 2.127.0 against 2.130.1 and main: the changelog, is_benefit_cap_exempt_earnings, where statutory_paternity_pay is read, uprating_indices.yaml, and the ONS population growth that E1 divides by, which is unchanged after 2.127.0, so the factor holds on 2.130.1. I ran household probes on both versions.
• Disclosure scan of the body, the receipts, the changelog fragments, the signed differences and the spine fixture. The fixture is generated by tools/graph_uk_spine_fixture.py and is synthetic.
• The PR is mergeable. main's commits since the base are US and NZ only.

I couldn't build a spine, run a calibration or check any figure in the receipts, because I have no FRS or feed access. I also couldn't run the full engine-uk or engine-free CI job lists or the feed-gated tests here.

🤖 Generated with Claude Code

[juaristi22]

@vahid-ahmadi thanks, all six are addressed. The branch is force-pushed: it is rebased onto main 121fd4dd6 (#1256, #1279), so every SHA has changed, and the old receipts commit is replaced rather than edited, so no commit on the branch carries an effective sample size. Head: fcb6ff23e. The four arms in the receipts ran on main 65d7d153a; #1256 changes the SPI channel's draws, so a control and the head were rebuilt on 121fd4dd6: the control fits at loss 0.00584 (98.4% of rows within 10%) and the head at 0.00599 (98.6%), both with all 7 gates, and the head's written file checks out as before.

1. Effective sample sizes. Removed. The new receipts (fcb6ff23e) compare the arms on loss, rows within 10% and gates only, and give no effective sample size, absolute or relative.

2. The lock. María ruled to move to 2.130.1 here, with its own arm. 3623ca57d locks it and re-derives the pins, concept coverage (288 to 299 inputs), H2 fixture, lock hash, stage texts and coverage manifest; the target references and compile-parity receipt regenerate identically. The bump arm is main plus only the three lock commits, so the engine's effect is measured on its own. Against main it fits alike (loss 0.00593 against 0.00583, 98.6% of rows within 10% against 98.8%, all 7 gates on both). Before calibration 169 rows move by more than 1%; on the rows you named, at the spine's weights, Universal Credit is +0.4%, capped households −0.8%, and ESA −3.0% with 5.1% fewer income-related claimants. The legacy closure and statutory pay as earned income both reach Universal Credit, and the arm measures them together. The body no longer says "the latest release on PyPI"; 2.131.0 (trading losses, published at 12:29 UTC today) is left for the next bump.

   • The bump surfaced one thing the build had to follow. With Gift Aid extending the bands (pe-uk#2271), the salary-sacrifice relief by rate class no longer added up to earned_income_tax for donors, and the first head calibration on 2.130.1 refused. The allocation now extends a donor's bands by the grossed-up gift as calc_income_tax_with_gift_aid does, the three relief rows declare it, and hand-computed donor cases hold it against the engine (test_a_donors_relief_bands_follow_the_engines_extended_bands).
   • Legacy closure (pe-uk#2119): the build leaves would_claim_uc_at_legacy_closure at the engine's default, so every family moved off a legacy benefit claims Universal Credit. uk-data#534 draws it at DWP's Move to UC claim rates, which are not a Chronicle fact yet; the concept mapping, the changelog and the receipts say the 2025 Universal Credit rows rest on that default.
   • The CGT index: 2.130.0 gives capital gains their own projection path, but with no scenario passed it copies GDP per head exactly (legacy_gdp_per_capita), so the build's gains are projected as before.

3. R3. Kept, on María's ruling, and documented. On 2.130.1 statutory paternity pay is UC earned income and in household benefits and HBAI income, so the column now raises net income. changelog.d/1095-statutory-paternity-pay.added.md says that before 2.128.0 it was only a National Insurance base and lowered net income, and the fragment and the receipts record the SPPADJ annualisation (weekly amount × 365.25/7) as open on uk-data#575.

4. The failing row. Recorded: e3342783a holds dwp.uc.households_children_5_or_more out in calibration_measure_exclusions.json with the reason, tracking microcosm#1095 and the sparse-band expiry, 26 November 2026, signed as María's ruling of today; the packaged-exclusions test pins it. The two-child-limit rows stay bound, and the head passes all 7 gates as merged and writes a release file.

5. Ports and go-firsts. The body and the receipts now say which: R1 (uk-data#526) and R4 (uk-data#518, merged on 14 October) port merged code; R2 (Finish #515: per-record E19200 decomposition by published SOI 2.1 AGI-band components #561, an open PR), R3 (calibration_diagnostics.json has no declared schema and is not validated on write #575), T1 (Staging telemetry silently disabled since 2026-07-02: empty POPULACE_STAGING_REPO_ID defeats the on-by-default staging repo #563), E1 (QBI v3 wiring: simulation consumes SCF/SOI/JCT evidence (shape from SCF, level from JCT) #538) and E2 (Read HMRC's joint distribution of capital gains by taxable income #558) go first; E3 and the stored birth months have no incumbent counterpart.

6. Test counts. The body and the receipts now quote the same runs. On 1da8d681f (X before the rebases): engine-uk 205 tests, all pass; engine-free 17,475 tests, 17,318 pass, 155 skip and two fail on this machine only (the US scorer test that needs Python 3.14.7, and the live Hugging Face pointer); feed-gated 99 tests, 97 pass, 2 skip. On the head rebased onto Draw SPI-channel incomes at the FRS twin's conditional rank (#840) #1256 (0912388b6): engine-uk 207, all pass, and the engine-free files the rebases touch, Draw SPI-channel incomes at the FRS twin's conditional rank (#840) #1256's SPI tests among them, 966 tests, 964 pass and two skip.

Also new: 84ce97a2a stores months_since_last_birthday in the spine, which María asked to fold in here. The engine spreads each age and sex over the year by weight when the column is unset, so a release at calibrated weights read different dates of birth from those its solve fitted, on main too. Read as the engine reads it, a file without the fix differs from its own solve on 55 of 1,217 rows, 8 by more than a point (the State Pension rows by scheme and amount most); with it, on 3, and no State Pension, two-child-limit or Pension Credit row moves. What is left (private-school pupils and two land rows) depends on the weights themselves. With the months stored, E3 is the backstop for any other engine read keyed on an identifier.

🤖 Generated with Claude Code

[vahid-ahmadi]

@juaristi22

Good to merge. CI is green, and the two shoulds below can land here or in a follow-up. All six round-1 points are fixed. The engine bump has its own measured arm, and on 2.130.1 R3 now raises the income it adds.

Round 1, at fcb6ff23e

1. Effective sample sizes: fixed on the branch. No effective sample size appears in the receipts, the PR body, any of the 20 commit messages or any added line on the branch. The old commit is still public, though (new item 1).
2. The lock: fixed. uv sync --locked --extra uk resolves policyengine-uk 2.130.1 and core 3.33.3. The bump arm is reported as shares and rates, with no counts. The rows I named move the expected way before calibration: 0.8% fewer capped households, with ESA down 3.0%, which is what the 2.129.1 remunerative-work rule should do. After calibration the bump fits about as well as the control. The CGT claim holds. With no projection passed, 2.130.1's add_capital_gains_projection copies the GDP-per-head series exactly, and nothing in microcosm passes one. The new cgt_taxable_income definition (net_income less allowances, with Gift Aid moved to the band) matches 2.130.1's capital_gains_tax. "Latest release on PyPI" is gone from the body.
3. R3: fixed. I re-ran the probe on the locked 2.130.1. £5,000 of SPP now raises household_benefits by £5,000, NI by £400, and both household_net_income and hbai_household_net_income by £4,600. The fragment says what happened before 2.128.0, and the fragment and the receipts both record the SPPADJ annualisation as open on uk-data#575.
4. The failing row: fixed. calibration_measure_exclusions.json:590 has the reason, the tracking issue, the ruling and the expiry (2026-11-26). If I delete the entry, test_packaged_exclusions_load fails.
5. Ports against go-firsts: fixed, in both the body and the receipts.
6. Test counts: fixed. The body and the receipts quote the same runs, and the parts add up.

New

1. should: the old receipts commit is still public. ed33feb76 is still served by GitHub by its SHA. I fetched it through the API today. Replacing the commit on the branch does not remove it, so the effective sample sizes from round 1 can still be read. Whether that needs a removal request to GitHub Support (their sensitive-data process clears unreferenced commits and cached PR views) is a disclosure judgement for you, not a code point.

2. should: no test pins how frs_spine stores the birth months. frs_spine.py:513 is the only place the spine gets the engine-way spread at FRS weights. Storing 0 for everyone there passes all 34 touched test files and the H2 parity test, which builds both sides from frs_raw and so moves with the mutation. Spreading with equal weights instead of gross4 also passes both test_uk_birth_month.py files, test_uk_frs_spine.py, test_uk_release_identifiers.py and H2. The engine tests cover uk_months_since_last_birthday itself but not the call. A small test on the synthetic tabs would close the gap: check that the spine's column equals the function at the FRS weights, and that a clone keeps its source person's months.

3. nit: PyPI has moved past 2.131.0. 2.131.1, 2.131.2 and 2.132.0 to 2.132.5 are all out (2.132.5 at 07:53 UTC today). None of them changes an input the build writes:

   • the build writes no trading_loss and clips self-employment income at zero (frs_spine.py:542), so 2.131.0 has little to act on;
   • the HBAI-based rows in uk_population_targets.json are pending.

   But the ETB services stage reads hbai_household_net_income from the engine (etb_services.py:59). So 2.132.4, which resets negative BHC income to zero, would move that matching for the households it resets. 2.131.2 (AHC now deducts ground rent and service charges) and 2.132.4 also change what users read from a release, including the bottom decile. "Not in this PR" could name 2.132.x rather than only 2.131.0.

4. nit: the receipts don't show the legacy closure landing. On the bump, UC spending rises 0.4% while UC households fall 0.1%. That is the one move I couldn't square with tax-credit families moving onto UC from 2025-26. There is no tax-credit target row, so the calibration can't see it either. A line giving CTC and WTC on both arms would show the closure took effect, at the same share-level detail as the rest.

What I checked

• New worktree at fcb6ff23e with a fresh uv venv (uv sync --all-packages --locked --extra uk).
• ruff check . passes. All 34 touched test files pass, with 6 skips. CI is green on all 10 checks.
• git range-diff from ed33feb76 to fcb6ff23e. I also diffed every non-test source file whose change differs from round 1.
• Read the three new commits (B3, M, X), the regenerated-surfaces commit, the receipts, the body, every commit message and the nine changelog fragments.
• Mutations:
  • Gift Aid band extension switched off: test_a_donors_classes_add_up_to_the_engines_extended_schedule fails.
  • X entry removed: fails as above.
  • Birth months: two mutations, both pass (item 2).
• Checked the band-extension arithmetic in tax_by_rate_class by hand. It stays exact when a gift spans more than one band.
• Engine probes on 2.130.1 (SPP, and a legacy tax-credit family). Read the pe-uk changelog from 2.130.1 to 2.132.5, and grepped microcosm for the inputs those releases touch.
• Disclosure scan of the body, the receipts, the fragments, the exclusion entry and all commit messages. The changed H2 fixture files are generated by tools/graph_uk_spine_fixture.py.

I couldn't build a spine or run a calibration, because I have no FRS or feed access. So none of the arm figures is checked, including the 55-to-3 rows result for the stored birth months. I also couldn't run the full CI job lists or the feed-gated tests.

🤖 Generated with Claude Code