[juaristi22]

The fifth #1095 ports PR. It locks policyengine-uk 2.132.6 (frozen for this PR) and ports the uk-data wave up to that release, with #868's borrower stocks, #1284's council-tax basis and Vahid's points from his review of #1271. #1287 is left out, on María's ruling. Rulings applied are María's of 15 and 16 October; each commit names the one it applies.

Every item is in, and the head passes all 7 calibrated-seam gates. The directors' dividends (O2) are taken out on María's ruling of 16 October: they made the solve fail the weight-ratio gate, and they stay open in #1095 with what we found (see "Not in this PR"). The receipts are in experiments/1095-uk-data-ports-5-receipts.md.

Fixes #868. Fixes #1284.

Summary

The build stopped on main (fix)

• Mobile-home NEED cells start at their mean when they drew no kWh (UK: data fixes to model the High Value Council Tax Surcharge (mansion tax) correctly, and the related housing and council-tax gaps #1240). After Place high-value homes where the Land Registry sees them (#1240) #1270 main's spine build stopped at the energy_rake gate: a NEED cell had a zero current mean, and the rake multiplies, so it cannot move it. price_domestic_energy declares zero_kwh_cells_at_need_mean (accommodation: all dwellings); before the first sweep the rake starts each declared cell in which no household drew kWh at the cell's NEED mean. The gate fails on a started cell the stage does not declare and on any other zero-mean cell, and its details count the started cells without naming them.

Vahid's review of #1271

• The stored birth months are pinned by a test (his second should). Both fixture adults share one stratum; the test checks the stored column against the engine-way spread at each household's GROSS4, that it differs from the equal-weights spread, that no month is zero and that an SPI copy keeps its source person's months. It fails under both of his mutations.
• The engine moves past 2.131.0, and the ETB services stage reads HBAI income before the reset (his third nit): B1 and B2 below.
• CTC and WTC on both arms (his fourth nit): in the receipts, as shares, with the head's results.

Engine (B1–B3)

• B1. The UK lock moves from policyengine-uk 2.130.1 to 2.132.6. Only that package moves; policyengine-core stays at 3.33.3 (UK) and 3.32.5 (US). Nothing the build writes is renamed, removed, re-typed or turned into a formula. 2.131.0 adds trade losses, which join the engine's HBAI subtractions; the survey-concept gross income leaves them subtracted. The uprating pins, concept coverage, cited texts, H2 fixture, coverage manifest digests and APPROVED_UV_LOCK_SHA256 re-derive; the target references and both compile-parity receipts regenerate identically.
• B2. HBAI income before the reset to zero. 2.132.4 resets negative HBAI BHC income to zero (pe-uk#2402). The LCFS predictors and NEED band, the NTS income band, the ETB VAT rate and the ETB services predictors now read hbai_household_net_income_before_reset_to_zero under the predictor name they had.
• B3. The ESA support group from the FRS ESA group. The root stage reads the benefits tab's VAR3 (ESAGROUP) on ESA rows: esa_includes_support_component is false only where a person's ESA rows record the work-related activity group or the assessment phase and none records the support group. SPI-redrawn rows return to true.

Self-employment losses (L1–L3; uk-data#528)

• L1. FRS SEINCAM2 splits by sign: profits to self_employment_income, losses to trading_loss (read by the engine from 2.131.0). Before, every loss was dropped.
• L2. The SPI stage's FRS-only fill draws trading_loss last for the rows it redraws; a row with a drawn profit keeps no loss.
• L3. The tape's losses brought forward (LOSSBF) move out of self_employment_income into loss_relief and ni_class_4_losses_brought_forward, which the engine relieves (ITA 2007 s.83; SSCBA 1992 Sch 2 para 3); they move, they are not copied. On María's ruling the HMRC self-employment rows measure the profit after that relief, so they keep their values at fixed weights. Three follow-ups make every reader agree: the HMRC total-income bands read total_income_after_loss_relief; the CGT stages and the area income read the profit after the relief; and the measure reads the stored profit, as the rows did before.

Childcare (U2; #1126, #1127)

• U2a. Vahid's UK: source the extended-hours usage draw from DfE additional-hours PTEs (#1126) #1127 commit, cherry-picked with his authorship: the extended-hours draw's mean moves from 15.019 to 31.141 (DfE's January 2025 additional-hours PTEs). This supersedes UK: source the extended-hours usage draw from DfE additional-hours PTEs (#1126) #1127.
• U2b. The fitter (tool version 4) follows pe-uk#2177: targeted and universal hours are allocated from working-parent hours, and the universal row counts universal-only children.
• U2c. The four rates are refitted on the 2.132.6 spine: Tax-Free Childcare 0.9797 (was 0.88), extended hours 0.6579 (0.6054), universal 0.4353 (0.4539), targeted 0.7769 (0.597). No rate reached its bound. The two HMRC Tax-Free Childcare rows still cannot both be met (spend 0.86, children 1.11): a residual of the frame's spend per child, not of the rate.

Targets

• U1. dwp.uc.households_children_5_or_more becomes a diagnostic beside dwp.uc.two_child_limit.households_5_children, pinned to DWP's April 2025 count, the month of the publication it sits beside. The count includes families outside the two-child limit, and fewer than 10 FRS source households carry that difference. Its exclusion entry retires.
• E1. obr.esa binds every benefit on OBR's incapacity line the engine carries: esa_income + esa_contrib + incapacity_benefit + sda (uk-data#582; OBR table 4.9 footnote 2). The incapacity part of Income Support has no engine variable and stays a residual of the row with no declared difference.
• D1a. The student-loan borrowers below the threshold (UK spine: SLC plan-2 borrower stocks at a third of SLC's — the FRS flag captures repayers, not the liable stock #868). frs_education stores SLOS (reports_outstanding_student_loan) and EDUCLEFT (age_left_full_time_education). A reported borrower repays, or, in England or Wales, reports an outstanding loan and is not in full-time education. A course starts three years before its borrower left full-time education (at 18 without an answer), and its start year sets the plan; the Plan 2 walk takes nobody in full-time education. The four SLC rows bind academic year 2024/25, whose reference date falls in calendar 2025, with a signed in-year-snapshot hold; they now match the incumbent fixture. Its own commit, found on the head's calibration: nobody repays a Plan 5 loan before April 2026, so before 2026 a repayer never takes Plan 5; a repayer whose dates point to a course from 2023/24 takes Plan 2, and SLC's published zero of Plan 5 borrowers above the threshold holds. The cohort operation declares the year (plan_5_repayments_from: 2026); María confirmed the rule on 16 October. The year is fixed, not derived from the cohort, so it is to be revisited when the build year moves (Vahid's review).
• C1. The council-tax stock on the household basis of its area (UK national register: council-tax dwelling stock and the official household controls bind the same household count at different levels #1284). A national-only household_basis_partitions declaration rescales each English region's, Wales's and Scotland's dwelling total onto its household control, and its bands by the same factor: the area's published households per chargeable dwelling, 0.980 to 1.023 on the 735820c feed. A tripwire refuses a factor further than 0.035 from one (the measured 0.0227 with a margin), and the closure check requires each rescaled total to equal its control.
• U3. The sparse-band support floor. See its own section below.
• P1. The parity fixture's Scottish constituency rows sit in their seats (uk-data#589). Five unchanged constituencies keep their 2010 codes in the incumbent's source; the extraction tool maps them to S14000107–S14000111 and takes the rows by code into the PCON24 roster's order. 798 values move, no row is added or lost; the local compile parity is regenerated.

Inputs

• O1. Six owner-manager inputs from the FRS job tab (pe-uk#1965; UC Regs 2013 reg 77). A director owning at least half of the company (DIRPERC 1–3) stands as its sole owner or partner and works in it; the company is the main employment for the first job, and carries on a property business when the main job's SIC division is 68. The income share is the job's dividends after tax; it reads job DIVIDEND only, and dividend_income is unchanged. On SPI-redrawn rows the stage-2 fill draws the six inputs last.
• U4. Private-school attendance at the spine's weights (uk-data#568). A new last spine stage materializes the engine's attends_private_school once, at the calibration year and the spine's design weights, and stores it, so the solve and the release read the same flags. Stored attendance no longer responds to income changes or to reforms of the attendance parameters.

U3. The sparse-band support floor

The measure-exclusion register moves to schema 3 and declares the floor once, signed by María on 16 October (expires 8 January 2027). Over the four dwp.uc.payment_distribution_* targets, a band whose compiled matrix row reaches fewer than 10 distinct source households (household_source_id) with a positive prior weight is dropped from the solve, not pooled. It runs on the compiled problem before the loss weights, families and metadata are drawn: in the national problem node and, after the stitch, in the full build's measures node (so before #1207's family loss-share floors are computed; those floors cover other families). Its receipt gives each family's dropped published total and no support count. The run refuses outside the window, past the ceilings, and when a per-row exclusion names a band the floor would drop. The drops join the problem's measure exclusions as kind: support_floor, and the national gates, the diagnostics, the full build's target-surface and reference-coverage gates and the certifier read them back.

Ceilings: 24 rows and 5% of the bands' published total. On the head spine the floor drops 19 bands, 3.1% of that total, so the ceilings leave a margin of five rows and about two points; which bands sit under the floor can move from one spine to the next. María confirmed the ceilings on 16 October, with a little headroom.

The 21 per-row band entries, sorted on the head spine (no support count is given here or anywhere):

• (a) Below the floor, so the floor drops them and the entry retires (9): COUPLE_NO_CHILDREN 22,800–24,000, 24,000–25,200, 26,400–27,600, 27,600–28,800 and 28,800–30,000; SINGLE 21,600–22,800, 26,400–27,600, 27,600–28,800 and 28,800–30,000.
• (b) At or above the floor but out of reach at the 10× weight cap: none.
• (c) Neither, so the entry stays and needs your renewal before it expires on 26 November (12): LONE_PARENT 12,000–13,200, 13,200–14,400, 14,400–15,600, 15,600–16,800, 16,800–18,000, 18,000–19,200, 19,200–20,400 and 21,600–22,800; SINGLE 8,400–9,600, 9,600–10,800, 13,200–14,400 and 14,400–15,600.

The floor also drops ten bands the solve bound before: COUPLE_NO_CHILDREN 10,800–12,000, 19,200–20,400, 20,400–21,600, 21,600–22,800 and 25,200–26,400; COUPLE_WITH_CHILDREN 28,800–30,000; SINGLE 20,400–21,600, 22,800–24,000, 24,000–25,200 and 25,200–26,400.

Vahid's review of this PR

• The ESA group test now pins each code on its own (1 and 3 outside the support group, 2 the support group, any other code no group), and an engine test pins the SPI stage's ESA reset and trading-loss clearing on the stage's output; his mutations fail them.
• The ESA group codes cite the FRS 2024-25 variable listing (UK Data Service SN 9563, sheet BENEFITS_table): on an ESA row VAR3 holds the questionnaire's ESAGroup, 1 work-related activity group, 2 support group, 3 assessment phase.
• The receipts say why the Tax-Free Childcare rate (0.98) is not a bound hit and why the universal rate is low; the fitter's docstring on the universal row's allocated term matches the code; the UK: data fixes to model the High Value Council Tax Surcharge (mansion tax) correctly, and the related housing and council-tax gaps #1240 fragment says a started NEED cell replaces the spend its households drew; a comment marks the Plan 5 year for revisiting.
• His two O2 follow-ups are in UK: track the defects uk-data is fixing that microcosm still has #1095's directors' dividends item.
• The band names under the support floor are María's call under the UK Data Service rules, before merge.

Deviations and open points

• L3 uses two named measures (self_employment_income_after_loss_relief, total_income_after_loss_relief) that the UK resolver computes, not a new term in the evaluator grammar the plan named. They give the same values; the grammar change can follow.
• E1 also binds incapacity_benefit, beyond the recommendation: OBR's footnote names it, and it closed to new claims in 2008, so the term carries only what the FRS still reports.
• U1 is pinned to April 2025, the month of the two-child-limit publication it sits beside.
• U4's pupil series stays ISC's member-school census. DfE's count of all independent-school pupils, the series María chose, is not yet a Chronicle fact; switching needs that fact first.
• D1b is out of reach. The head spine's Plan 2 England stock is about 41% of SLC's at design weights, and the walk's whole pool would give about 47%, so the plan's 25% acceptance cannot be met by the cohort rule. The two UK spine: SLC plan-2 borrower stocks at a third of SLC's — the FRS flag captures repayers, not the liable stock #868 exclusions (slc.borrowers.plan_2_liable, plan_2_above_threshold, expiring 3 November) need a ruling: renew, or turn them into diagnostics.
• The per-row refusal runs on the national line only. The full build drops below-floor bands too, but does not compile the per-row-excluded bands to measure them; the national run on the same spine does.

Measurement

Each arm builds its spine locally on feed 735820c and runs the national role's calibration on it. Details are in the receipts.

• control (main 5d4eb37b3 plus the NEED fix, policyengine-uk 2.130.1): loss 0.29148 to 0.00607; 98.6% of 1,226 rows within 10%; all 7 calibrated-seam gates pass.
• head with O2 (before O2 left): loss 0.26546 to 0.00583; 98.4% of 1,217 rows within 10%; 6 of 7 gates pass, uk_weight_ratio fails.
• head: loss 0.26817 to 0.00584; 98.4% of 1,216 rows within 10%; all 7 gates pass. On its written file the weight year, the identifiers and the stored birth months check out, the input-mass gate passes, and of the 1,216 solved rows only the two ONS land rows (declared weight-share formulas) read differently from the solve; the private-school pupils row now matches it.
• C1 on the head: the 97 council-tax stock rows miss by 0.13% on average (control 0.64%) and none by more than 2% (control: 10 rows); the household controls miss by 0.06% (control 0.07%).
• Tax credits (Vahid's fourth nit): child and working tax credit are zero from 2025 on both the control and the head, so the legacy closure takes effect on both engines; Universal Credit at 2025 is £77.7bn on both at their calibrated weights.

Not in this PR

• O2, the directors' dividends (María's ruling of 16 October). DIVIDGRO in dividend_income, as the larger of the accounts' dividends and the director's, added about £5.8bn of dividends on FRS-channel people, more than doubling that channel's. The directors moved up the income bands, the bands they left lost support, and the solve failed the weight-ratio gate; with only O2 left out, all seven gates pass. It stays open in UK: track the defects uk-data is fixing that microcosm still has #1095 with this finding, for its own PR.
• UK local file: support is counted on the FRS host, but SPI-drawn incomes carry no tape-record id #1287 (María's ruling).
• The renewal of ons.savings_interest_income (UK spine: savings interest income a third under HMRC SPI and five times under the ONS D.41 row — reconcile the target concept and review FRS interest capture #866, expires 22 October): a separate small PR.
• The local-authority re-measure of C1: María runs it herself.
• policyengine-uk releases after 2.132.6.

Disclosure

Checked against the UK Data Service rules before pushing: the commit messages, every added line, the fragments and this body. No record-level value, extreme, effective sample size or support count appears; the one small cell is given only as "fewer than 10"; obr.esa has no before-and-after figure; the ESA group is given only as support group or not.

Tests

• CI: the checks on this PR's head.
• After the review: the FRS spine, SPI income (engine-free and engine), student-loan, take-up tool and contract, energy pricing and graph tests, 1,073 with 2 skips, all pass.
• Locally, at the head's code: the touched surfaces, 1,479 tests with 1 skip; the SPI engine tests, 88 with 1 skip; CI's engine-uk job list, 213, all pass.
• The feed-gated tests on feed 735820c, at the first push: 103 with 2 skips, all pass.

🤖 Generated with Claude Code

[vahid-ahmadi]

@juaristi22 thanks, this is my review at 2dcb5619b.

Fine once the B3 tests are tightened and CI is green. O2 is out and the receipts are in. The engine bump, the HBAI swap and the loss split are right, and most new tests fail when I revert the code they cover. Three code paths are still not pinned (items 1 and 2), and I couldn't confirm the ESA group coding from the documents I have (item 3).

Points from my review of #1271

• Round 1, items 1 to 6: still fixed.
• Round 2, item 1 (old receipts commit): still public. ed33feb76 is still served by its SHA (I fetched it through the API today). This PR can't fix that. It stays a disclosure call for you.
• Round 2, item 2 (birth months): fixed. test_birth_months_are_spread_at_the_frs_weights_and_kept_by_copies fails if the spine spreads the months with equal weights instead of GROSS4. It also checks that no month is zero and that copies keep their months, which covers my store-zero mutation.
• Round 2, item 3 (engine past 2.131): fixed by B1 and B2.
• Round 2, item 4 (CTC and WTC on both arms): fixed. The receipts (Tax credits) give child and working tax credit as zero from 2025 on both arms and UC at 2025 as £77.7bn on both. These are aggregates rather than shares, but they are weighted totals over the whole frame, so that's fine.

Findings

1. should: the B3 test doesn't pin each ESA group code. The fixture adult has one work-related activity group row (VAR3 1) and one assessment-phase row (VAR3 3). If I change frs_spine.py:1245 to group.isin((3,)), test_esa_support_group_is_read_from_the_esa_rows still passes, so a WRAG-only person going to true would not be caught. One case per code (1 alone, 3 alone, 2 alone) would close this.

2. should: two SPI-stage call sites are not pinned. If I delete spi_income.py:1241 (_refresh_esa_support_component) and spi_income.py:1242 (_clear_trading_loss_beside_profit), nothing fails in test_uk_spi_income.py (engine-free or engine), test_uk_spi_spine.py or test_uk_spi_support.py. The helpers have unit tests, but nothing checks that impute_uk_spi_income_support calls them. By contrast, deleting the _move_losses_brought_forward_to_loss_relief or _owner_manager_flags_from_draws call does fail tests. An assertion on the stage output would cover both: a redrawn row's flag is true, and a redrawn row with a profit has no trading_loss.

3. should: please cite where the ESAGROUP codes come from. _esa_includes_support_component (frs_spine.py:1222) reads VAR3 1 as WRAG, 2 as support group and 3 as assessment phase. The UKDS data dictionary I have labels VAR3 only "Code 0" to "Code 17", and DWP doesn't post the questionnaire. The engine's note that about nine in ten ESA recipients who answer say support group fits code 2 being support, but that is not an independent check. If the codes were swapped, the benefit cap ESA exemption and UC's LCWRA test would flip for most ESA recipients. A reference in the docstring (questionnaire block or DWP variable spec) would settle it.

4. Resolved: O2 is out. Taking it out of this PR is what I would have suggested. Your local branch without it passes all 7 gates, and the uk_weight_ratio fence (gates.json:823) is not something to loosen to pass. The O2 section of the receipts records the finding without any record-level figure. Two points for the follow-up in UK: track the defects uk-data is fixing that microcosm still has #1095:

   • The removed docstring said DWP add DIVIDGRO beside investment income in total income, which means DWP treat the two as separate, while the larger-of rule assumes they overlap. That is worth settling first, because it changes the dividend mass.
   • Two candidates to measure: adding DIVIDGRO only for owner-managers (DIRPERC 1 to 3, as O1 does), or leaving directors' dividends to the SPI channel at the incomes it redraws.

   O1 stays. On 2.132.6, UC unearned income leaves out dividends (income_definitions/unearned.yaml), so its owned_company_income_share is not counted twice in UC.

5. should: the Tax-Free Childcare rate is close to its bound. U2c fits 0.9797, and the fitter's bound is 1.0 (tools/fit_uk_childcare_takeup.py:49). Both TFC rows still miss (spend 0.86, children 1.11), so the rate is mostly absorbing the frame's spend per child. Your U2c ruling, as the receipts record it, is that a bound hit keeps the bound. Please say in the receipt whether 0.98 counts as a hit for that rule. Separately, the universal rate falls to 0.4353. A sentence on why the frame needs so low a standalone rate to meet DfE's universal-only count would help a reader.

6. nit: the fitter docstring says the universal row's allocated term is zero (tools/fit_uk_childcare_takeup.py:291). The code computes that term anyway. Under pe-uk#2177, a child with no standalone take-up who uses 15 hours or less gets base hours and no extended hours. On the allocation side, that child counts as universal-only. At the new usage mean of 31.141 that case is very rare, so the effect is negligible, but the docstring should match the code.

7. nit: UK: data fixes to model the High Value Council Tax Surcharge (mansion tax) correctly, and the related housing and council-tax gaps #1240 doesn't bias the rake, but it overwrites what the donors drew. Starting a declared all-zero cell at its NEED mean doesn't bias the rake: the accommodation margin holds that cell's mean wherever it starts, and the cell's households are a negligible share of the other margins. Their drawn spend (at or below the standing charge) is replaced by the all-dwellings mean, so their energy spend rises above what they drew. That matches the incumbent translation for mobile homes, so it's fine. The changelog fragment could say it in one line. The gate does fail closed: if I disable the undeclared-cell check at stage_health.py:564, test_the_energy_rake_gate_takes_only_the_declared_held_cells fails, and if I drop the start at energy_pricing.py:1125, its test fails too.

8. nit: the Plan 5 rule (you asked for a confirmation). For the 2024/25 stock, the rule at student_loans.py:286 is sound. A repayer whose dates point to a course from 2023/24 can't be repaying Plan 5 in 2025. One caveat: the date is fixed (plan_5_repayments_from: 2026), not derived from the cohort. With the three-year course assumption, a 2023/24 starter leaves in 2026 and first repays in April 2027. So a 2026 build would let repayers take Plan 5 a year before the assumed cohort repays. That is worth revisiting when the build year moves.

9. Disclosure question (a judgement for you). No support count appears anywhere, and "fewer than 10" was accepted on UK: policyengine-uk 2.130.1 and the fourth batch of #1095 ports, with the release's weight year, identifiers and birth months fixed #1271. The rewritten U3 commit (7f862f94e) changes only the measured drop, from 18 bands at 2.8% to 19 bands at 3.1% of the published total, and the margin text. The question is unchanged, and the new receipts widen it: 19 UC payment bands whose matrix rows reach fewer than 10 source households are now named in the PR body, in the receipts (U3 section), in the floor receipt's dropped list and in the exclusion entries. Each named band is in effect a small-cell marker. The receipts give DWP's published households for the dropped bands by family, which is fine on its own. Please confirm the band names are acceptable under the UKDS rules before merge.

   The rest of the receipts are clean. I found no exact count, effective sample size, minimum, maximum, unit-record value or seat-level value in them, nor in the body, the commit messages, the added lines or the fragments. The receipts say weight ratios and the largest weight stay local. The one range given (0.980 to 1.023, C1) is of published households per dwelling across areas. The "at most 0.22 points" is of target-row differences, not microdata.

What looks right

• B1: in uv.lock only policyengine-uk moves (2.130.1 to 2.132.6, PyPI sdist and wheel hashes). The three requires-dist floors match it, core stays 3.33.3 (UK) and 3.32.5 (US), and shasum -a 256 uv.lock equals APPROVED_UV_LOCK_SHA256. uv sync --locked --extra uk resolves 2.132.6.
• B2: every engine read of HBAI net income in the LCFS, NTS, ETB VAT and ETB services stages now uses hbai_household_net_income_before_reset_to_zero. The predictor names on the donor side are unchanged. local_targets.py:383,386 still read the reset equiv_hbai_household_net_income(_ahc), which is right for the published HBAI poverty and decile concepts. Leaving trading_loss subtracted from the survey gross income is right too.
• L1 to L3: the engine (2.132.6) deducts loss_relief from self-employment income only, never below zero. Class 4 starts from self_employment_income before loss_relief (ni_class_4_profits_before_losses), so setting both inputs is not a double deduction. ITA 2007 s.83 and SSCBA 1992 Sch 2 para 3 are the right provisions. On SPI rows, the profit after relief equals the profit the chain drew in both branches of spi_income.py:2741. derive_hmrc_income_auxiliaries runs before the move, while loss_relief is still zero, so nothing is subtracted twice. The HMRC rows, the regional and ITL bands, CGT and the area income all read the profit after relief. The engine's total_income adds the profit before relief, so the swap in total_income_after_loss_relief is correct. band_filter_dimension: total_income is a Ledger filter key, not a variable read. ledger_targets.py mentions self-employment income only in a docstring.
• U2a: the commit keeps your authorship and date, and its changes are identical to UK: source the extended-hours usage draw from DfE additional-hours PTEs (#1126) #1127's head (b91ef34f8). The clipped normal at mean 31.141, sd 4.972, bounds 0 to 30 gives 28.535 hours, which matches 15 + 341,988.09 × 15 / 379,000.
• U2b follows pe-uk#2177's allocation (base hours first from working-parent hours, flags following the allocated hours).
• O1: the DIRPERC codes match the UKDS job dictionary (1 = 100%, 2 = more than 50%, 3 = 50%, 4 = less than 50%).
• The O2 removal is clean: git diff 522d453d5 2dcb5619b touches only the O2 code and test, its fragment, the fixture and generator columns, the regenerated digests and the receipts.

What I checked

• New worktree at 2dcb5619b with a fresh venv (uv sync --all-packages --locked --extra uk).
• git range-diff from 522d453d5: commits 1 to 18 are unchanged, the O2 commit is dropped, and the last six are rebased, with 7f862f94e changed only in its text and the receipts commit added.
• ruff check . passes, and ruff format --check passes on the 92 touched Python files. tools/ci_test_plan.py verify passes.
• At 522d453d5: all 45 touched test files passed, 1,613 tests with 10 skipped.
• At 2dcb5619b: the files covering the rewritten commits pass, 554 tests with 1 skipped. These were the frs_spine, support-floor, student-loan, private-school, take-up, source-stage, coverage-manifest, country-spec, population-target, CGT and spine receipt and parity tests, plus the H parity test.
• Mutations that fail the new tests:
  • L1 losses dropped;
  • L3 copy instead of move;
  • HMRC measure without the relief;
  • birth months at equal weights;
  • the NEED gate's undeclared-cell check removed;
  • the NEED start removed;
  • O1 without the DIRPERC test;
  • the _move_losses… and _owner_manager_flags_from_draws call sites deleted.
• Mutations that pass, so the tests miss them: items 1 and 2, which I rechecked at this head.
• Read in policyengine-uk 2.132.6: the loss relief, Class 4, total_income, esa_includes_support_component, owned_company_income_share and UC unearned income. Also read the changelog from 2.130.1 to 2.132.6 and pe-uk#2177.
• Disclosure scan of the body, the new receipts, all commit messages, the added lines outside the generated fixtures, and the fragments.
• CI at this head was still running when I checked: lint, wheels, select-countries and integration-uk pass; engine-free, engine-uk and engine-us are pending.

I couldn't build a spine or run a calibration (no FRS or feed access), so none of the arm figures, the refitted rates or the receipts' measurements are checked. I also couldn't check the ESAGROUP codes (item 3). I only skimmed U3 (support floor: I checked that its receipt carries no support counts), U4, C1, P1, E1, U1 and D1a beyond the Plan 5 commit, so I have no findings on them.

🤖 Generated with Claude Code

[juaristi22]

@vahid-ahmadi thanks. Pushed at 20c52989c. The commits up to f57c50091 are unchanged; five review commits follow it, and the receipts commit moved to the end, so git range-diff 2dcb5619b...20c52989c shows five new commits and the receipts rewritten.

From #1271

• Round 2, item 1 (old receipts commit): María's disclosure call, outside this PR.

Findings

1. ESA group codes: fixed in 968a9f945. test_each_esa_group_code_sets_the_support_component_on_its_own keeps one ESA row per case: codes 1 and 3 give false, 2 gives true, and 0 records no group. Your group.isin((3,)) mutation now fails the work-related activity case.
2. SPI-stage call sites: fixed in 968a9f945. The engine test test_the_stage_resets_the_esa_group_and_clears_a_loss_beside_a_profit runs impute_uk_spi_income_support with the fake forests, which draw a positive value for every target, and asserts on the stage's output: each redrawn row's esa_includes_support_component is true, FRS rows keep theirs, and a redrawn row with a profit keeps no trading_loss. Deleting either call now fails it; I checked both.
3. ESAGROUP codes: cited in d918388e2. The FRS 2024-25 variable listing (UK Data Service SN 9563, 9563_frs2425_variable_listing_eul.xlsx, sheet BENEFITS_table) maps VAR3 on ESA rows (BENEFIT 16) to the questionnaire's ESAGroup and gives its format: 1 Work-related activity group, 2 Support group, 3 Assessment phase. The "Code 0" to "Code 17" labels in the data dictionary are VAR3's generic range across benefits; the per-benefit formats are in that sheet. The docstring and the changelog fragment cite it.
4. O2: both follow-up points are now in UK: track the defects uk-data is fixing that microcosm still has #1095's directors' dividends item, under "Next". Thanks for checking that UC unearned income leaves out dividends, so O1's income share is not counted twice.
5. The Tax-Free Childcare rate: in the receipts' new U2c section. 0.9797 is an interior optimum (L-BFGS-B stopped inside 0 to 1), so the bound-hit rule does not apply. Even at a rate of 1 the frame's Tax-Free Childcare spend is 0.88 of HMRC's while its children with used accounts are 1.13 of HMRC's, so the fit balances the two misses at 0.86 and 1.11: the residual is the frame's spend per child, about 0.78 of HMRC's, not the rate. On the universal rate: with every scheme at a rate of 1 the frame already gives 1.77 times DfE's universal-only children, and at the fitted extended-hours rate about a third of the children eligible for extended hours take none and join the universal-only pool, so 0.44 meets DfE's count.
6. The fitter's docstring: fixed in 84e1ffece. It now says the universal row's allocated term is computed and near zero, for the case you describe. No fitted value changes.
7. The UK: data fixes to model the High Value Council Tax Surcharge (mansion tax) correctly, and the related housing and council-tax gaps #1240 fragment: 09a2c8f1b adds the line. A started cell's households take its NEED mean in place of the spend they drew, as the incumbent translation for mobile homes has it, and the rake is not biased.
8. The Plan 5 year: María confirmed the rule for the 2025 build. 18be8d205 puts your caveat beside PLAN_5_REPAYMENTS_FROM: the year is fixed, not derived from the cohort, and is to be revisited when the build year moves.
9. The band names: María's call under the UK Data Service rules; she will confirm before merge.

At the new head the files the review touches pass: 1,073 tests with 2 skipped (FRS spine, SPI income engine-free and engine, student loans, the take-up tool and contract, energy pricing, and the graph tests). CI is running.

🤖 Generated with Claude Code

[vahid-ahmadi]

@juaristi22 thanks, this is my round-2 pass at 20c52989c.

Good to merge once CI is green and you have made the band-names call (item 9). Items 1 and 2 are now pinned, and my mutations fail the new tests. I checked the ESA group codes against the SN 9563 variable listing myself and they match. The two new findings below are nits.

Status of my round-1 points

1. ESA group codes pinned one by one: fixed (968a9f945).
2. SPI-stage call sites pinned: fixed (968a9f945).
3. ESAGROUP codes cited: fixed and verified (d918388e2).
4. O2 follow-ups: done (now in UK: track the defects uk-data is fixing that microcosm still has #1095).
5. Tax-Free Childcare rate near its bound: answered in the receipts' new U2c section. I agree with the answer.
6. Fitter docstring on the universal row: fixed (84e1ffece).
7. UK: data fixes to model the High Value Council Tax Surcharge (mansion tax) correctly, and the related housing and council-tax gaps #1240 fragment: fixed (09a2c8f1b).
8. Plan 5 year: fixed (18be8d205). The comment at student_loans.py:66 records the caveat and that the rule is confirmed.
9. Small-cell band names: still open. It is your call, and the body says you will make it before merge. Nothing has changed here: the 19 dropped bands and the 12 bands under renewal are still named in the body and the receipts.

Evidence for 1 to 3 and 5

• Item 1. test_each_esa_group_code_sets_the_support_component_on_its_own (test_uk_frs_spine.py:1430) keeps one ESA row per case. I ran five mutations of frs_spine.py:1249 and each one fails it:

  • isin((3,)) fails the WRAG case;
  • isin((1,)) fails the assessment case;
  • isin((2, 3)) fails the WRAG case;
  • support read as group == 3 fails it, along with the older test;
  • treating every code other than 2 as outside fails the no-group case.

  isin((1, 2, 3)) survives, but it is equivalent: a code-2 row also sets support, so the result can't change.

• Item 2. test_the_stage_resets_the_esa_group_and_clears_a_loss_beside_a_profit (tests/engine/uk/test_uk_spi_income.py:941) fails if I delete spi_income.py:1241 or spi_income.py:1242. I ran it with the four SPI test files, as in round 1.

• Item 3. I fetched 9563_frs2425_variable_listing_eul.xlsx from the UKDS documentation site (it is open documentation, not data).

  • On sheet BENEFITS_table, the ESA row (BENEFIT 16, questionnaire key 54, WageBen[7]) puts Busamtdk, ESAType and ESAGroup in the Var1, Var2 and Var3 columns.
  • The decode block gives ESAGroup as 1 Work-related activity group, 2 Support group, 3 Assessment phase.
  • The FRS 2024-25 data dictionary agrees: it labels the derived ESAGROUPHD and ESAGROUPSP "Multi-purpose variable No3".

• Item 5. The fitter minimises the sum of squared relative misses (fit_uk_childcare_takeup.py:179). Both TFC rows scale linearly with the rate, so the optimum is (0.88 + 1.13) / (0.88² + 1.13²) = 0.980. That matches 0.9797, so it is an interior optimum, and the bound-hit rule doesn't apply. The residual being spend per child (0.88 / 1.13 = 0.78) follows from the same figures. I can't check the universal-rate explanation (1.77 times DfE, about a third joining the pool) without a spine, but it is consistent with 0.4353.

New findings

1. nit: the body still opens with "Status: draft." (line 3 of the body). The PR is now out of draft.
2. nit: the docstring rewrite left a ragged line at tools/fit_uk_childcare_takeup.py:296. "The expectation is the quantity a take-up rate" now breaks early before "controls". This is cosmetic, and ruff doesn't reflow docstrings.

No new disclosure problems. The new U2c section gives only ratios to HMRC and DfE totals, plus the fitted rates. The new body section and the receipt bullets carry no counts. The only other new numbers are test totals and target-row counts.

What I checked

• Range-diff. git range-diff 2dcb5619b...20c52989c: the merge base is f57c50091, which is 2dcb5619b^, so commits 1 to 24 are identical. The five review commits are new and the receipts commit is rewritten. git diff 2dcb5619b 20c52989c touches 8 files and changes no build code apart from the two docstrings and the PLAN_5_REPAYMENTS_FROM comment.
• Setup. A new worktree with a fresh venv (uv sync --all-packages --locked, then --extra uk), on policyengine-uk 2.132.6.
• Tests. The FRS spine, SPI income (engine-free and engine), student-loan, energy pricing and stochastic-tools (childcare fitter) test files: 299 passed, 1 skipped.
• Lint. ruff check . and ruff format --check on the touched Python files both pass. tools/ci_test_plan.py verify is ok.
• Mutations. The eight listed above.
• Citation. The SN 9563 variable listing, sheet BENEFITS_table, and the FRS 2024-25 data dictionary.
• Disclosure. A scan of the new receipts diff, the body diff and the commit messages.
• CI at 20c52989c. lint, wheels, select-countries, integration-uk and engine-uk (3.13 and 3.14) pass. engine-free (3.13, 3.14) and engine-us (3.13, 3.14) were still pending when I checked.

I couldn't build a spine or run the fitter, so the U2c ratios and the universal-pool share are not checked against data.

🤖 Generated with Claude Code