[juaristi22]

Part of #1095. This PR collects the #1095 items that pass their measurement, and since 31 October the Phase B items too. Each item is measured in its own national build against the one before it. Every figure that rests on the SPI channel is given as a class only (SN 9422 and the disclosure rule).

A/B items (measured against the 28 October build, then against the 2 November control):

1. The SPI stage-1 forests train and query on the SPI age band (uk-data#628).
2. Households' additional-dwelling (HRAD) purchases at or below £250,000 are allocated from HMRC's Table 8a receipts.
3. Drawn SPI employees' current pay is floored at what their twin's hours earn, and the tape's pay stays their year of pay (uk-data#629).

Phase B items (each measured against the build before it):

• B5: Carer's Allowance claims are drawn after salary sacrifice, keyed on lineage.
• B12: Scottish owners' council tax bands are drawn from their council area's band mix.
• AC3: the two-child-limit rows date births by the engine's cutoff test.
• B7: children in non-parental care are marked for the two-child limit's exception.
• B10: SPI copies take their twin's contributory JSA and ESA within each benefit's own work limits.
• B9: each household's student-loan balance moves to its borrowers.
• B4: HMRC's HICBC liability count and revenue join as measured diagnostics.
• B3, PR 2b (J2 and K): the NTS cars grade the zero-emission share by income, and each road fuel is levelled to its household-paid share.
• B3, vehicle excise duty: each adult's car gets policyengine-uk's VED inputs from the NTS cars.

Phase B items that change nothing in the build:

• B6 (ages from the interview): dropped, by María's ruling of 2 November. The stored ages already sit at the engine's reference date (note).
• B8 (benefit-cap exemptions): every exemption the engine reads is already set. The engine's own gaps are in Lift the Universal Credit benefit cap for a claimant who receives War Widow's Pension policyengine-uk#2792 and Add Guardian's Allowance policyengine-uk#174.

Also since the first review:

• Vahid's five points of 29 October, each in its own commit (reply):
  • the one-hot SPI age band;
  • the pay floor's receipt on the floor of 3;
  • paid_hours;
  • the UC gate's pools;
  • the HRAD docstring.
• The UC take-up gate is back at 0.055 for each family type, by María's ruling of 2 November (follow-up).
• Vahid's second review of 3 November (review, reply):
  • B10's contributory JSA takes JSA's own limits, and both awards an hours test;
  • B10's receipt counts by source household, on the floor of 3;
  • B7's flag is called an upper bound on the exception;
  • the UC take-up gate's note names t1426m.
• The SPI prior shape (UK: every open data item: uk-data parity, engine locks, the local surface, consumption and high-value homes #1095 residual 12) came onto the branch on 3 November and came off the same day, by María's ruling. It left the fit and the over-£100,000 young-child cell's concentration where they were. The concentration tool stays: tools/diagnose_uk_spi_prior_concentration.py.
• Main is merged in at 79a533c72. It brings main's own fixes for the transport bindings' module count (Count the two modules #1342 added to the transport bindings' identity, and merge main's US re-pin (#1496) #1495) and the US spec-engine seed pins (Repin the US spec-engine contracts after the ACS local SSI and immigration stages #1496), whose absence failed this PR's engine-free and engine-us checks.

How the items were measured

• Control: María's national build of 2 November, cut f6e33b93 at SPI donor savings rebased to SPI's 2023-24 bands (savings follow-up to #1418) #1428's 329a99e71 (policyengine-uk 2.165.1, feed 08962f1b), staged on populace-uk-private.

• T1426: this branch with the A/B items and Vahid's fixes, built on this machine with the control's review date (2 November) and feed.

  • Its 17 stages before the SPI income stage give the same evidence and records as the control's. So the two machines' inputs agree.

• Phase B: one build per group of items, in a chain: B-1 against T1426, B-2 against B-1, B-3 against B-2. All use the same review date, feed and lock, and all 1,339 rows are active in every build.

• Main merged in: the branch now merges main at 312dc4d6b, which adds SPI donor savings rebased to SPI's 2023-24 bands (savings follow-up to #1418) #1428's last commit (2dbf7cf, SPI banding on total income net of expenses). A control rebuilt on main and this branch's merged tree measure the whole PR on the current base (below).

• Control (f6e33b93): 98.5% of rows within 10%, final loss 0.00529.

• T1426 (A/B items and Vahid's fixes):

  • 98.8% within 10%, final loss 0.00527.
  • 21 rows move more than 5 points, all HMRC SPI regional and income-band rows. That is reseed-sized: a placebo that only reseeded the SPI stage moved 22.

• B-1 (B5): 98.8% within 10%, final loss 0.00527. No row moves.

• B-2 (AC3, B7): 98.7% within 10%, final loss 0.00532. No row moves more than 5 points.

• B-3 (B10, B9): 98.7% within 10%, final loss 0.00522. No row moves more than 5 points.

• Control on main (ctl-312d):

  • 98.7% within 10%, final loss 0.00539.
  • Its 17 stages before the SPI income stage give the same evidence and records as the control's; from that stage on it differs, as main's SPI banding commit should.
  • Against the control, 10 rows move more than 5 points, all HMRC income-band and regional rows.

• This branch merged with main (t1426m):

  • 98.7% within 10%, final loss 0.00520.
  • Against the control on main, by stage:
    • Its stages before the SPI income stage give the same evidence and records, except frs_relationships (B7).
    • From the SPI income stage on, every stage's evidence differs, as this branch's SPI items should.
    • The stage order differs only where B5 moves the Carer's Allowance stage.
  • Against the control on main, by row:
    • 14 rows move more than 5 points, all HMRC income-band and regional rows. That is fewer than the 22 a reseed-only placebo moved.
    • Before calibration, obr.jobseekers_allowance's shortfall halves (B10).
    • After calibration, dwp.esa_claimants comes inside 10%. The UC households-by-children rows stay near the 10% line, as they do in both controls.

• B-5a (B4, J2 and K) against t1426m:

  • 98.7% within 10%, final loss 0.00525. No row moves more than 5 points.
  • By stage, only lcfs_consumption differs.
  • dwp.uc.households_single_with_children slips outside 10%. It sits at the line in every build.
  • B-5a and the builds after it calibrate at review date 3 November, when B4's exclusions start. The 1,339 rows are the same.

• b5sp (the SPI prior shape) against B-5a: 98.7% within 10%, final loss 0.00524. No row moves more than 5 points. The shape is off the branch (above).

• B-5b (VED) against b5sp: only lcfs_consumption differs, and the calibration is identical: no target reads the new inputs.

• b5j (JSA's own limits, with VED and without the shape) against B-5a:

  • It builds ccc18c08f: 0cddc8464 with the node's declared reads, on the 312dc4d6b base.
  • 98.9% within 10%, final loss 0.00525 (B-5a: 98.7% and 0.00525). No row moves more than 5 points.
  • By stage, only lcfs_consumption (VED), spi_benefit_coherence and uc_take_up differ.
  • Three rows come inside 10%: dwp.uc.households_children_3, dwp.uc.households_single_with_children and hmrc.spi_region.taxpayers_by_region_50000_70000@E12000008. None goes outside.

Every build above passes every spine gate and all seven calibration gates. They measure the items on main at 312dc4d6b. Main is now merged in at 79a533c72, so a control on that main and a build of the merged head are owed before a whole-PR comparison on the new base.

1. The SPI stage-1 forests on the SPI age band (uk-data#628)

Why

The 2022-23 SPI public-use tape gives age only as a band (AGERANGE). The SPI income stage drew an age for each donor inside its band and trained the stage-1 tape forest on it, so the forest learned no real income pattern within a band. Twins of one band were then drawn from conditionals that differed only by that noise. uk-data#628 reports the same defect in uk-data.

What changes

• The stage-1 tape and reference forests read the SPI age band, one-hot, instead of age (Vahid's review). On the tape it is the AGERANGE code, with composite records (-1) as a band of their own. On a frame person it is the band of their whole-year age, through one helper shared by both frames.
• at_or_over_spa still carries the State Pension split inside the 65-74 band.
• The cross-fit fold cells key on the band.
• The drawn donor ages stay on the donor for their other readers: the band pools' State Pension side and the property components' tape seed.

Measurement

• Against the 28 October build, the item moved no more rows than a draw-seed placebo, and the UC rows ended closer to DWP's.
• The copies' pay now rises with age inside the youngest bands nearly as FRS employees' does, and falls with age over 55 as theirs does. The placebo changed none of this.

2. Households' HRAD at or below £250,000 from Table 8a receipts

Why

The property purchases stage (#1308, porting uk-data#573) allocates households' HRAD purchases by individuals' share of all Table 8a residential transactions (0.881) and value (0.882). #1308's review found this probably overstates households' purchases.

The residential nil-rate band was £250,000 for all of 2024-25. A purchase at or below it owes no standard-rate SDLT, so other buyers' receipts in that band are surcharge.

What changes

• The nil-rate band from receipts: households' HRAD at or below £250,000 is Table 6b's band less other buyers' HRAD implied by their Table 8a receipts. The stage refuses an implied figure outside Table 6b's band or above other buyers' own transactions there.
• Above the band, uk-data#573's shares stay. Solving for other buyers' HRAD from receipts goes negative above £1m. The module docstring says the shares probably still overstate households' purchases. Table 8a doesn't split out the 2% non-resident surcharge, so the nil-band figure, if anything, understates them (Vahid's review).
• Targets: households' 2024-25 HRAD target falls from 186,589 transactions and £68.81bn to 151,928 and £63.71bn.

All figures in this section are HMRC's published Annual Stamp Tax Statistics 2024 to 2025, Tables 6b and 8a, as the pinned feed carries them.

Measurement

Households' purchases fit their new target, and no row moved by more than 5 points against the 28 October build. obr.sdlt fits within 0.25% after calibration in every build.

3. Drawn SPI employees' current pay at their twin's hours (uk-data#629)

Why

The SPI income stage keeps each drawn copy's FRS twin's hours and job, while stage 1 redraws the copy's pay from the 2022-23 tape. The tape's pay covers the tax year and holds part-year workers, so part-year pay could land on a current full-time twin, at an hourly rate far below the minimum wage. Such a copy fails the minimum-income tests of the 30 hours' childcare offer, Tax-Free Childcare and the benefit cap more often than the FRS rows it copies.

What changes

• The floor: floor_spi_employment_pay_at_minimum_wage lifts a drawn copy's pay to its twin's usual weekly paid hours (paid_hours, Vahid's review) × 52 × the lower of the minimum wage for the twin's age and the twin's own hourly pay.
• Who it touches: only drawn recipients other than reserved band carriers, whose twin has an employee main job, no trade and usual hours, and whose drawn pay is positive. FRS rows are never modified.
• The tape's year of pay: the floored pay is current pay only. employment_income_annual_share keeps the HMRC income rows on the tape's pay.
• Receipt: an employment_pay_floor receipt and an hourly-pay check. A figure resting on one or two source households is withheld as "<3", on the shared floor of 3 (Vahid's review).
• Gate: uk_stage_hmrc_spi_income_spine_hourly_pay refuses a spine where an eligible copy stays below its floor, or where drawn employees sit below the minimum wage more often than FRS employees by more than two points.
• The UC take-up gate is at 0.055 again for each family type and for the reachable pools (María's ruling of 2 November). With SPI donor savings rebased to SPI's 2023-24 bands (savings follow-up to #1418) #1428's pool joining, every family type ends closer to DWP's count after the redraw.

Measurement

After the floor, drawn employees sit below the minimum wage for their age less often than FRS employees, within the gate's margin, and no eligible row stays below its floor. The HMRC employment-income bands end as without the floor, since their annual measures read the tape's pay.

B5: Carer's Allowance after salary sacrifice

• What changes: the Carer's Allowance take-up stage runs after salary_sacrifice, so its earnings test reads pay after sacrifice. That puts it after the CGT copies, so its draws are keyed on lineage: a copy draws as the member it copies, and a source and its copies in one stratum count as one record of their joint weight. Original rows keep their draws.
• Measurement (B-1 against T1426):
  • Only the Carer's Allowance stage differs, and every calibration row is unchanged.
  • The change turns off the claim flag only on CGT copies that salary sacrifice puts over the earnings limit, and the engine already paid them nothing.
  • So it makes the flag consistent with what the engine pays.

B12: Scottish owners' council tax bands

• What changes: Scottish owners' bands are drawn, at the same uniform, from their council area's chargeable-dwellings mix (Scottish Government, September 2025, A to H), vendored from the pinned feed as uk/scotgov_chargeable_dwellings_by_band.json. Renters and the per-copy bill stay with the bill-level issue.
• Measurement: it runs only in dense builds, so the national builds here can't measure it. It is measured in the next dense build.

AC3: two-child-limit births by the engine's cutoff test

• What changes: the 15 two-child-limit rows read the engine's uc_is_child_born_before_child_limit (a date of birth before 6 April 2017) instead of a birth year of 2017 or later. A child born from January to 5 April 2017 no longer counts as affected.
• Measurement (B-2): the seven rows whose measure reads the cutoff start 1% to 4% lower, and all 15 still fit within 0.5%.

B7: non-parental care and the two-child limit

• What changes: frs_relationships sets child_limit_non_parental_caring_exception for a child whom no adult of their own benefit unit claims as a natural, adopted or step child (household grid codes 3 and 4 from the child, 7 and 8 from the adult), and whom no local authority placed there. This covers kinship and friends care.
  • Multiple births stay at the engine's default: the FRS has no twin field, and its birth dates are blank (María's ruling of 2 November).
  • The flag leaves the signed engine-default register.
• Measurement (B-2): inert in this build. The flag reaches the calibrated dataset, but no flagged child is a third or later child born after the cutoff in a paid UC claim, so no row moves.

B10: SPI copies' contributory JSA and ESA within each benefit's own work limits

• Why: spi_benefit_coherence zeroed contributory JSA on every SPI copy and left contributory ESA with the stage-2 forest's draw, which knows nothing of the unit's claim history. A contributory award follows the person's National Insurance record and incapacity, which the copy keeps; what the copy changes is its work.
• What changes:
  • An SPI copy takes its FRS twin's contributory ESA within permitted work (María's ruling of 29 October; ESA Regs 2008 reg 45(4)): under 16 hours a week, and earnings up to policyengine-uk's esa_exempt_work_higher_earnings_limit at the frame's year, £183.50 a week in 2024-25.
  • It takes the twin's contributory JSA within JSA's own limits (Vahid's second review): under 16 hours a week (JSA Regs 1996 reg 51(1)(a)), and earnings below the age-related amount plus the £5 disregard. Reg 56's prescribed amount is that sum less 1p: £95.49 a week from 25 in 2024-25. The award is the lower of the twin's and the age-related amount less the earnings above the disregard (Jobseekers Act 1995 s. 4(1)(b)).
  • The amounts are policyengine-uk's own parameters at the frame's year. Earnings are employment income plus any self-employment profit, over 52 weeks. Hours are the twin's usual hours, which the copy keeps and the pay floor reads.
  • The operation, with its hours and award, is in the source manifest and the sources schema, and the graph hands the node the hours and the source households.
  • The stage gate checks the rule on every copy. The receipt counts support by distinct source household and withholds a weighted figure that rests on one or two.
  • Contributory ESA leaves the QRF tail-concentration surface, since no forest draw survives in it.
• Measurement (B-3 against B-2, with ESA's limit for both awards):
  • On the copies, contributory ESA falls (the forest had given them more than their twins report) and contributory JSA returns at a small level.
  • Before calibration, obr.jobseekers_allowance's shortfall halves. The ESA rows start further below DWP's and OBR's: without the forest's extra ESA, the FRS base's own shortfall shows.
  • All of them fit after calibration, and dwp.esa_claimants comes inside 10%.
• Measurement (b5j against B-5a, with JSA's own limits):
  • JSA's own limits leave the copies' contributory JSA where B-5a had it, so obr.jobseekers_allowance starts where it did and B-3's halving stands.
  • ESA's hours test takes contributory ESA off some copies, so the ESA rows start slightly lower.
  • All of them fit after calibration.

B9: student-loan balances to borrowers

• Why: was_wealth splits the household's WAS student-loan balance before any plan exists, so it could sit on a member with no loan while the household's borrower held none. policyengine-uk caps a borrower's repayment at a positive balance.
• What changes: once student_loans sets the plans, each household's balance moves to its members who hold one: by repayments where any of them repays, equally otherwise. A household with no borrower keeps was_wealth's split, and no household's total moves. The stage opens a reader-isolation boundary in the UK graph, since it rewrites a cell an earlier open-surface stage reads.
• Measurement (B-3): the SLC repayment rows' starting estimates move by under 1%, and they all fit after calibration. Their large starting overshoot is not the balance's: the WAS balance is a small part of SLC's loan book, so few repayments are capped.

B4: HICBC as measured diagnostics

• What changes: two national rows from HMRC's Child Benefit statistics (August 2025, Table 16): the Self Assessment taxpayers liable to the charge in 2024-25 and its revenue, both provisional (310,000 and £360m). They are measured at the model's 2025 under a signed hold, since the thresholds are the same in 2024-25 and 2025-26, and held out of the solve by two signed exclusions (María's ruling of 3 November).
• Measurement: the model's charge sits well above HMRC's, and more than half of it on SPI copies: too many high-income parents still receive Child Benefit. Families in payment and obr.child_benefit fit, so it is not the number of claims. The fix belongs in the SPI channel and is UK: HICBC runs well above HMRC's count, mostly on the SPI copies (too many high-income parents in payment) #1508. Neither the SPI prior shape nor PR 2b moved it.

B3, PR 2b (J2 and K): the NTS cars

• What changes: lcfs_consumption reads the National Travel Survey's cars (SN 5340, survey years 2022 to 2024, at the interview weight W3).
  • J2: each car is zero-emission at DfT's licensed-car share (VEH1103) times its household income fifth's relative risk among the cars NTS households keep, normalised so the national share is unchanged.
  • K: household road fuel keeps ONS 07.2.2's level, and each fuel is levelled to its share of household-paid spend: HMRC's litres times DESNZ's cars share times the share of car miles a household pays for.
  • obr.fuel_duties_cars stays held (María's ruling of 29 October), its reason restated with the household-paid shares.
• Measurement (B-5a against t1426m): only lcfs_consumption differs, and no row moves more than 5 points. The top income fifth now holds the most households whose cars are all zero-emission, and household road fuel moves from petrol to diesel at the same total.

B3: vehicle excise duty inputs

• What changes: the same stage writes policyengine-uk's five car inputs for each adult who keeps a household car: fuel type, CO2 emissions, engine size, first registration date and the exemption flag (María's design of 3 November).
  • A household's own cars go to its adults, oldest first. Company cars are left out, and cars beyond the adults are dropped and counted in the receipt.
  • Each car's attributes come from an NTS car of the same income fifth and age band with linked CO2 bands, reweighted so the fleet keeps its age mix, since the newest and oldest cars carry the linked record less often. A zero-emission car has no engine size.
  • DLA higher-rate mobility and PIP enhanced mobility make the car exempt. The list price stays the engine's default. Registration dates are ISO strings.
  • The inputs join the export allow-list, so the battery's digests move.
• Measurement (B-5b against b5sp): only lcfs_consumption differs, and the calibration is identical: no calibration target reads VED. policyengine-uk counts VED in household tax, so household net income falls by each household's duty; HBAI income does not change. An engine test checks each duty regime's rate on the step's own dates.
• Limits:
  • No car pays the expensive-car supplement, since the list price is the default.
  • The step realizes a lower electric share than J2's fuel flag, because a fuel-buying household's first car always burns fuel. Which share to follow is María's call.
  • The registration dates are fixed, so from the year after the build's no car pays first-year duty: Car VED: fixed first-registration dates leave no first-year duty after a dataset's year policyengine-uk#2868.

Tests

CI at 1ac980d1d: all ten checks pass.

Local runs on the commits since 22703ee0d:

• B10: test_uk_spi_benefit_coherence.py (34 tests: JSA's prescribed amount, award cut and hours, ESA's hours, the receipt's floor), the engine test of the limits at 2024 and 2025, and the node's declared reads in test_uk_graph.py.
• On the merged tree: the gate battery pins, microcosm-data's contract, the transport bindings' identity, the country spec and the spec-engine inventory and field usage (629 tests), and B10's other readers (1,004 tests). None fail.
• Generated surfaces:
  • The H2 parity fixture regenerates byte-identically.
  • The release input coverage manifest, the CI test plan and the spec-engine coverage are current.
  • The US seed protocol and seed map match main's pins.
• Earlier: VED's tests (test_uk_nts_car_attributes.py, the LCFS and source-stage tests, and the engine test of the duty regimes), and the full engine-free suite at 2b49a00ac (351 files, 13,226 tests; two fail only on this machine: one expects the UK extra to be absent, and the Orrery reconstruction test meets a local Node error).

New and changed test files, against main:

• New:
  • tests/engine/uk/test_uk_spi_permitted_work_limit.py reads ESA's and JSA's limits from policyengine-uk at 2024 and 2025;
  • tests/engine/uk/test_uk_vehicle_excise_duty_inputs.py charges each VED regime on the step's own dates;
  • test_uk_nts_car_attributes.py and test_uk_spi_prior_concentration.py.
• Engine tier: test_uk_spi_income.py.
• Engine-free, UK:
  • test_uk_carers_allowance_take_up.py, test_uk_frs_relationships.py, test_uk_frs_spine.py and test_uk_graph.py;
  • test_uk_lcfs_consumption.py, test_uk_measure_simulation.py, test_uk_measurement_period.py and test_uk_population_targets.py;
  • test_uk_property_localisation.py, test_uk_property_purchases.py, test_uk_receipt_counts.py and test_uk_source_stages.py;
  • test_uk_spi_benefit_coherence.py, test_uk_spi_income.py, test_uk_spi_pension_age.py, test_uk_spi_rank_transport.py and test_uk_spi_spine.py;
  • test_uk_spine_acceptance_receipt.py, test_uk_stage_health.py, test_uk_student_loans.py and test_uk_target_references.py;
  • test_uk_uc_capital_coherence.py, test_uk_uc_target_measurements.py and test_uk_weighted_integrity.py.
• Engine-free, shared: test_country_spec.py, and microcosm-data's test_contract.py.
• Test support: uk_atomic_support_fixtures.py and uk_student_loans.py.

Follow-up

• HICBC: the fix is UK: HICBC runs well above HMRC's count, mostly on the SPI copies (too many high-income parents in payment) #1508.
• VED: the electric share (above) is María's call, and later years' first-year duty is Car VED: fixed first-registration dates leave no first-year duty after a dataset's year policyengine-uk#2868.
• The merged base: a control on main at 79a533c72 and a build of the merged head (above).
• Phase B builds still to run: B-4 (the gambling stake level, held for Calibrate the QRF sign gates out of fold, for every gated UK imputation #1391) and B-6 (copy ids re-keyed). The target-side items B11, B14 and B15 follow them.
• NS4 (benefit-cap support): still on a thin base. The repair options are an SPI-channel rule for copies' cap state (UK SPI channel after #840: residuals, post-merge measurements and stale pointers #1288) or a support stage.
• uk-data#573 still allocates every band by the shares. Raising this there is María's call.

Generated surfaces: source_stages.json (in step with sources.yaml), the sources schema, the release input coverage manifest, the H2 parity fixture, B4's target references and membership, and the gate digests mirrored in microcosm-data's contract.

🤖 Generated with Claude Code

[vahid-ahmadi]

@juaristi22 thanks, this is my review of 457af0274.

Verdict: the three items look sound, with four should items and a nit. The age band is the right fix for #628. The nil-band arithmetic reproduces the body's numbers: £265m over £625m/113,500 is about 48,100 transactions, and the target falls by about 34,700. The pay floor runs where it should, after stage 1 and the uprating and before employment_income, stage 2 and pension coherence, with the twin cap and the kept year of pay. Nothing here blocks the A/B arms.

Findings

1. should: composites (-1) sit next to 16-24 on an ordinal predictor. spi_age_band is numeric, not one of SPI_STAGE1_CATEGORICAL_PREDICTORS (spi_income.py:2252, 2623). So any split above -1 that doesn't separate -1 from 1 puts composite records in the leaves 16-24 queries land in. Before, composites drew an age across 16-70, so they spread over every band. Now they can only mix with the youngest one, and composites are the tape's high-income amalgams. Please either keep composites out of the stage-1 training set, one-hot the band, or check on the arm that 16-24 draws' top tail is not higher than on main. A line in the predictor rule saying which would do.
2. should: the new floor receipt hardcodes ten. spi_income.py:936, 940, 951-952 and 961 use < 10 and "<10", and the gate docstring says "fewer than ten" (stage_health.py:1425). UK: take up #1095's actionable items on policyengine-uk 2.165.1 #1422 moves every UK floor onto receipt_counts.UK_DISCLOSURE_FLOOR. Whichever lands second should move these onto the constant and the helpers (small_cell_share, rounded_count), or this receipt will be the only one still on ten. The gate's "a withheld share is not judged" branch then follows too.
3. should: one tolerance loosens two checks. uk_stage_uc_take_up goes from 0.055 to 0.12 (gates.json:633). That widens the "no further from DWP than before the redraw" leg, which the floor needs. But the same number bounds the reachable pools, which the notes say realize DWP's count within 0.0015. A second parameter for the pools, left at 0.055 or tighter, would keep that check meaningful while the other is wide. Please also name the issue that tracks the return to 0.055, so the revert isn't left to memory.
4. should (a question): does Table 8a's nil-band receipts figure for other buyers include the 2% non-resident surcharge? If it does, part of those receipts is not HRAD surcharge. Then other buyers' implied HRAD is too high and households' too low (property_purchases.py:473). The direction is the safe one, and it may be small. One sentence in the module docstring saying which way it goes would do.
5. nit: paid_weeks (spi_income.py:893) is annual paid hours (hours / 52 * 52), not weeks. paid_hours would read better.

What I ran. I read the full diff at 457af0274 and checked the HRAD figures by hand. I also checked that the SPI leaf reconciliation runs on the donor tape, before the floor, so the floor can't break it. CI is green at this head. I didn't run tests locally or build anything, and I didn't check the private A/B figures.

🤖 Generated with Claude Code

[vahid-ahmadi]

@juaristi22 Round 2 at 2b49a00ac. I re-checked my five points from 457af0274, read your two replies, and read the Phase B diff against main (312dc4d6b): B5, B7, AC3, B9, B10 and B12. I checked B10 and B7 against the legislation. I ran the PR's engine-free tests for the stages it touches (spi_benefit_coherence, student_loans, frs_relationships, carers_allowance_take_up, stage_health, property_localisation, uc_target_measurements): 281 passed. I didn't build anything or look at the private figures.

Earlier points

1. Composites next to 16-24: fixed. spi_age_band is the first of SPI_STAGE1_CATEGORICAL_PREDICTORS (spi_income.py:185-190), so _encode_predictor_frames one-hots it. Tape and query codes are both float, so their dummy columns match (spi_income.py:2663, 3028-3030). The query side never produces -1, so the composite column is always 0 for frame people, as it should be.
2. The floor receipt's ten: fixed. It now uses UK_DISCLOSURE_FLOOR/UK_WITHHELD throughout (spi_income.py:984-1042).
3. One tolerance, two checks: fixed. maximum_reachable_pool_deviation is separate (stage_health.py:1542-1546, 1586) and both values are 0.055 (gates.json:655-656). See below on your follow-up.
4. Non-resident surcharge: fixed. The module docstring says which way it goes (property_purchases.py:39-44).
5. paid_weeks: fixed, now paid_hours (spi_income.py:939).

On your point-3 follow-up. I agree the caseload check can go back to 0.055 and nothing is left to track. One caveat on where the evidence comes from. The gate note cites the T1426 spine at 97fd4d0e4 (gates.json:658). That merges #1428 at 329a99e71, before #1428's SPI banding commit 2dbf7cf56, so the "every family type ends closer after the redraw" reading was taken on the old SPI stage. The body says B-1 to B-3 and t1426m all pass every spine gate, so the 0.055 gate has also passed on the current base. Please name t1426m's commit in the gate note next to 97fd4d0e4, since that is the spine the gate now runs on. If the receipt shows how far inside 0.055 the closest family type sits on t1426m, one figure there would also help. The body already says the T1426 and B-chain A/B measurements were taken against the pre-2dbf7cf56 control, and that ctl-312d/t1426m measure the PR on the current base, so I have nothing to add there.

New findings

1. Should: B10 gives contributory JSA ESA's permitted-work limit, which is about twice JSA's own earnings limit (spi_benefit_coherence.py:517-530). Contribution-based JSA needs the claimant's earnings not to exceed the prescribed amount (Jobseekers Act 1995 s.2(1)(c); JSA Regs 1996 reg 56: the age-related amount plus the earnings disregard, less 1p). For a claimant aged 25 or over in 2024-25 that is £90.50 + £5 − £0.01 = £95.49 a week. Below that, s.4(1)(b) reduces the award by earnings above the disregard. Failure case: an SPI copy aged 30 earning £7,800 a year (£150 a week), whose twin reports £4,706 of contributory JSA (£90.50 × 52), keeps all £4,706 at this head. Under the law it would get nothing, because £150 is above £95.49. At £50 a week it keeps £4,706 where the award would be about £2,366 a year (£45.50 a week). The docstring says the ESA limit is only a ceiling for JSA (spi_benefit_coherence.py:66-68). But the body puts part of the halving of obr.jobseekers_allowance's pre-calibration shortfall down to B10, so the gap matters here. Fix: test jsa_contrib_reported against reg 56's amount at the copy's age (the engine's JSA age-related amount + £5), and either reduce the award by earnings above £5 or zero it above the prescribed amount. Or, if María keeps the ESA limit for both, have the receipt report the copies' weighted JSA between £95.49 and £183.50 a week, so the JSA row's movement can be read with that in mind.
2. Nit: B10's "the test reads no hours" doesn't quite hold. spi_benefit_coherence.py:66 says the SPI tape has no hours. But the copy carries its twin's usual hours, which the pay floor reads (SPI_EMPLOYMENT_PAY_FLOOR_HOURS_COLUMN). ESA reg 45(4) also needs fewer than 16 hours a week. Either test the hours too, or reword the docstring to say why the twin's hours aren't used.
3. Nit: B7's flag is an upper bound on the legal exception (frs_relationships.py:204-218). UC Regs 2013 Sch 12 para 4(2) (before its revocation from 6 April 2026) needs a friend or family carer to meet one of conditions (a) to (h): a child arrangements order, special guardianship, guardian's allowance, a kinship care order, guardianship, a permanence order, or taking the child on where they would otherwise likely be looked after. The FRS grid shows none of these. So every child with no parent or step-parent in their benefit unit is flagged, including informal arrangements that wouldn't qualify. It is inert in this build, so one line in the rule's comment calling it an upper bound would do.

Checked and fine

• B5: lineage_roots refuses an unmapped or duplicated lineage. Originals keep their person_id draws. lineage_balanced_positions treats a source and its copies in one stratum as one record, and the within-one-record assert is widened to match.
• B9: allocate_balance_to_borrowers keeps each household's total (with a hard check), leaves households with no borrower or no balance alone, and splits by repayments, or equally where nobody repays.
• AC3: the measure now reads uc_is_child_born_before_child_limit, so a January to 5 April 2017 birth no longer counts as affected.
• B10 limit: 16 × £11.44 = £183.04, rounded up to £183.50 (ESA Regs 2008 reg 45(4) and (9A)), which matches the test constant.
• B12: the loader refuses a council area missing a band or repeating one, Northern Ireland and renters keep their bands, and the Scottish receipt is separate.

CI at 2b49a00ac: all ten checks pass (engine-free and engine-uk on 3.13 and 3.14, engine-us, integration-uk, lint, wheels).

Verdict: my earlier points are all fixed and I agree with the 0.055 return. B10's JSA limit is the one item to settle before the JSA row's movement is quoted.

🤖 Generated with Claude Code

[juaristi22]

@vahid-ahmadi thanks for round 2. Your finding, both nits and the gate-note caveat are addressed, and pushed at 1ac980d1d. Main is merged in at 79a533c72 for its own CI fixes.

Your point-3 follow-up (the gate note). The UC take-up gate's note now names t1426m (#1426 at 2b49a00ac) next to the T1426 spine, since that is the spine on the current base. On t1426m, single people, the family type that moves least, end 0.016 of DWP's count closer than before the redraw. The gate allows 0.055 further.

1. Contributory JSA now has JSA's own limits (spi_benefit_coherence.py, _contributory_work_limits).

• A copy keeps its twin's contributory JSA only if it works under 16 hours a week (JSA Regs 1996 reg 51(1)(a)) and earns below the age-related amount plus the £5 disregard. Reg 56's prescribed amount is that sum less 1p.
• Within those limits, the award is the lower of the twin's and the age-related amount less the earnings above £5 (s. 4(1)(b)). Above them it is nil.
• The amounts are policyengine-uk's own parameters at the frame's year (gov.dwp.JSA.contrib, gov.dwp.JSA.remunerative_work.claimant_hours): £71.70 under 25 and £90.50 from 25 in 2024-25, so the limit from 25 is £95.49 a week. An engine test reads them at 2024 and 2025.
• Your two cases are tests. At £150 a week the copy gets nothing. At £50 a week it gets £45.50 a week, £2,366 a year.
• ESA keeps its permitted-work limit, now with its hours test (below).
• Measured in b5j against B-5a, the build before it on the same base: JSA's own limits leave the copies' contributory JSA where B-5a had it, so obr.jobseekers_allowance starts where it did and the halving the body puts down to B10 stands. ESA's hours test takes contributory ESA off some copies, so the ESA rows start slightly lower. All of them fit after calibration. The build passes all seven gates with 98.9% of the 1,339 rows within 10% (98.7% in B-5a), and no row moves more than 5 points.

2. Hours. Both awards now read hours. The copy keeps its twin's usual hours (hours_worked over 52 weeks), which the pay floor already reads as the copy's. JSA needs under 16 hours a week (reg 51(1)(a)) and ESA under 16 (ESA Regs 2008 reg 45(4)). The docstring says so, and adds that neither award is reduced for the copy's pension payments.

3. B7's flag is now called an upper bound on the exception in the rule's comment (frs_relationships.py). It names para 4(2)'s conditions, which the FRS grid doesn't show.

Also in B10: the within-limit receipt now counts support by distinct source household and withholds a weighted figure that rests on one or two of them, as the other receipts do (receipt_counts). It gave row counts before. The graph now hands the node the hours and the source households: the first b5j spine stopped at the node without them, and a test now holds the node's declared reads to the work tests' columns.

Since your review (details in the body): the SPI prior shape came onto the branch and came off again (María's ruling of 3 November), and B4's HICBC diagnostics, PR 2b's J2 and K, and the VED inputs are on the branch. The HICBC gap is #1508. The builds measure each item on 312dc4d6b, so a control on 79a533c72 and a build of the merged head are still owed before a whole-PR comparison on the new base.

CI at 1ac980d1d: all ten checks pass.

[vahid-ahmadi]

@juaristi22 Round 3 at 1ac980d1d. I read 0cddc8464, 860731004 and 1ac980d1d and your reply, and checked the JSA limits against the Jobseekers Act 1995, the JSA Regs 1996 and policyengine-uk's parameters. I ran the engine-free tests for spi_benefit_coherence, the graph and frs_relationships (94 passed). I also re-ran my two cases through _contributory_work_limits on a synthetic frame. I didn't build anything or look at the private figures.

Earlier points

1. B10's JSA limit: fixed. The new limits match the law. Reg 56's prescribed amount is A + D − 1p, where A is the s.4(2) age-related amount and D is the reg 99(2) disregard. Reg 99(3) switches off Sch 6 paras 5 to 8 and 11 for contribution-based JSA, so the general £5 of para 12 applies. The age-related amounts are £71.70 under 25 and £90.50 from 25 in 2024-25, and £72.90 and £92.05 in 2025-26, matching gov.dwp.JSA.contrib. The 16-hour test follows reg 51(1)(a). The award is cut pound for pound above £5 (s.4(1)(b), reg 80) and converted to weekly and annual amounts at 52. My cases at this head: £150 a week gives 0, £50 a week gives £2,366 a year, £95.49 a week gives £0.52, and £50 a week under 25 gives £1,388.40.
2. Hours in the docstring: fixed. Both awards now test the twin's hours_worked / 52 against 16, and the graph declares hours_worked and household_source_id, with a test that holds the node to them.
3. B7 upper bound: fixed (frs_relationships.py:204-220).
4. Gate note: fixed. It names t1426m at 2b49a00ac and gives the 0.016 margin for single people (gates.json:658).

New findings

1. Nit: neither award is cut for the copy's pension payments, and the copies carry SPI pension income (spi_benefit_coherence.py:77; private_pension_income comes from the tape's PENSION, spi_income.py:416). The docstring now says so, but not why. JSA Regs 1996 reg 81 takes off every pound of pension payments above £50 a week. ESA Regs 2008 reg 74 takes off half of every pound above £85 (subject to reg 75, which leaves out survivors' pensions, war pensions and some health insurance). Failure case: a copy aged 58 with no earnings and a £15,600 a year private pension (£300 a week), whose twin reports £4,706 of contributory ESA (£90.50 × 52), keeps all £4,706. Under reg 74 the cut is half of (£300 − £85) = £107.50 a week, which is more than the award, so it gets nothing. On JSA the cut is £250 a week, so again nothing. Fix: cap each award at 52 × the weekly rate less the reg 81 or reg 74 deduction on private_pension_income / 52, and add that column to the node's declared reads in graph.py. Or, if you leave it out, have the receipt report the copies' weighted ESA and JSA where pension income is above £85 or £50 a week, so the reader can see how much rests on it.
2. Nit: the weekly tests read annual SPI pay over 52 weeks, which is strict for anyone who worked part of the year (_weekly_earnings, spi_benefit_coherence.py:863-869). The body already says the tape's pay covers the tax year and holds part-year workers, for the pay floor. The same averaging bears on the work limits. Contribution-based JSA lasts at most 182 days (Jobseekers Act 1995 s.5(1)), and the usual claimant has just left a job. A copy whose tape pay is £10,000 earned in 26 weeks, followed by 26 weeks on JSA, averages £192 a week and gets nothing, though its earnings in the weeks it claimed were nil. You measured the JSA row unchanged from B-5a, so a docstring line saying the averaged test understates part-year claimants is enough.

Checked and fine

• The parameters are read at the frame's year with pe-uk's fiscal-year values: the engine test asserts £71.70 and £90.50 at 2024 and £72.90 and £92.05 at 2025, and it passes in CI's engine-uk jobs (I couldn't install the engine here).
• Strict < against £95.50 vs reg 56's "not more than £95.49": the two differ only between £95.49 and £95.50, where the cap already gives under 1p a week.
• min(twin, cap) never lifts a copy above its twin's award (tested).
• The higher £20 disregard of Sch 6 para 9 (part-time firefighters, lifeboat crew, auxiliary coastguards, reserves) can't be seen on the tape. Leaving it out is right.
• The receipt now counts distinct source households, and it withholds a weighted figure, and the before/after change, when one or two households carry it.

CI at 1ac980d1d: all ten checks pass.

Verdict: all four round-2 items are fixed. The two new points are Nits: a docstring line or a receipt figure would cover each, and neither blocks.

🤖 Generated with Claude Code

[juaristi22]

@vahid-ahmadi thanks for round 3. #1426 merged before the two nits went in, so they are in #1391, in its last commit (70bd560):

1. Pension payments. The awards still aren't reduced, since B10 tests work alone. The receipt now shows how much rests on that. For each award, it gives the support and weighted award of the SPI copies whose private_pension_income over 52 weeks is above the award's threshold. For JSA that is reg 81's £50, read from policyengine-uk's gov.dwp.JSA.contrib.pension_disregard with the other JSA limits. For ESA it is reg 74's £85, a constant because policyengine-uk has no parameter for it. That figure is the most the reductions could take off. A figure is withheld where one or two source households carry it, or carry the rest of the award, which it would give back beside the whole. The node now declares private_pension_income, the graph test holds it to that, and the H2 fixture follows. Applying the reductions would change the awards beyond B10, so the receipt sizes that move first.
2. Part-year workers. The module docstring now says that the 52-week average understates part-year claimants, above all of contribution-based JSA, which lasts at most 182 days (Jobseekers Act 1995 s. 5(1)). _weekly_earnings points to it.

🤖 Generated with Claude Code