UK: port the remaining uk-data fixes tracked in #1095, engine items included

[juaristi22]

#1086 is settled: each country has its own policyengine-core pin (option 2), which landed on its own as #1143. This PR is the policyengine-uk bump that follows it: B0 moves only the UK side of the lock. C1 can still be cherry-picked on its own.

Before merge: a calibration of this head, which is queued. After merge, as for #1100: the signed-difference entries for the net-new columns (C5, C6, B2–B5) and for B1's mechanism change, whose drafts stay outside the repository until signed, then the national build and the certifier.

Summary

This ports every remaining unblocked item on #1095. It covers:

• six fixes that work on the currently locked engine;
• the move to policyengine-uk 2.122.2, with five fixes that need it;
• two signed calibration-register changes.

On the locked engine (C1–C6)

• C1. Target references regenerate from the pinned feed again. UK spine follow-ups: SPI donors funded, WAS coherence, residential split, Child Benefit, release-cut blockers (#1063) #1089 had hand-edited the generated uk/target_references.json. The contract now declares the band-growth index, so regenerating reproduces the committed references. The compiled values are unchanged.
• C2. Access fund. A calendar-month award that annualises above twice the largest annual-coded award is read as annual. This was Vahid's should-fix on UK: port the unblocked uk-data fixes tracked in #1095 #1100.
• C3. LCFS household counts. The counts are named household_adult_count and household_child_count and are counted from age, so they no longer read the engine's deprecated is_adult and is_child. The stage notes are now quoted in full. The imputation is value-identical (see Measurement).
• C4. Great Britain pension-age Housing Benefit spending. dwp.hb.amount_pension_age is bound at the calendar-2025 window: 3/12 of FY2024-25 (£6,851.0m) plus 9/12 of FY2025-26 (£7,114.7m), £7,048.8m.
  • DWP gives each year its own measure, so calendar_year_window gains the selector key source_measure_id_by_opening_year.
• C5. is_blind from the FRS blind registration question (Set is_blind from the FRS blind registration flag (SPCREG1) policyengine-uk-data#523), on the adult and child tabs alike.
• C6. uc_is_in_startup_period from the FRS claim and trade start dates (Set the UC start-up period from the FRS claim and business start dates policyengine-uk-data#527).
  • UC records with no linked claim date draw their recency, keyed on the benefit unit, at the linked share (0.324).

On policyengine-uk 2.122.2 (B0–B5)

• B0. The UK side of the lock moves to policyengine-uk 2.122.2 and policyengine-core 3.32.19. The US side stays on core 3.32.5.
  • Re-derived: the uprating pin stamp, the concept coverage and mapping review, the H2 fixture's dependency, and APPROVED_UV_LOCK_SHA256.
  • The mapping review covers Add the liquid financial assets concept, schema version 2 (NZ v0, G3a) #1120's liquid financial assets reason. It now names 2.122.2's two other person-level money stocks, car_list_price and lifetime_isa_balance, and Universal Credit's split of household capital by claimant and partner count.
  • The UC take-up population reads State Pension age from the new timetable by date of birth.
• B1. The severe disability flag is the tax credit condition on the derived categories (Store the tax credit severe disability flag, not a mix of premium tests policyengine-uk-data#494). It equals the engine's formula, which an engine-lane test checks.
• B2. is_claimant_or_partner and is_hbai_dependent_child come from the adult and child tables (Supply is_claimant_or_partner from the FRS adult table policyengine-uk-data#524, Supply is_hbai_dependent_child from the FRS and move targets and imputations off the deprecated is_child/is_adult policyengine-uk-data#486).
• B3. pension_credit_reported_capital takes the same recorded capital as uc_reported_capital (Use the FRS benefit-unit capital for Pension Credit's capital test policyengine-uk-data#513).
• B4. Shared rent and boarders' and lodgers' rent.
  • liable_for_share_of_household_rent follows Mark the benefit units that share their household's rent policyengine-uk-data#512's rule.
  • CVPAY becomes the payer's rent_paid_as_boarder or rent_paid_as_lodger. UK FRS spine: stop counting lodgers' rent and property losses as property income; sub-letting rent for every tenure #1081, which merged first, already leaves it out of property_income.
• B5. The mixed-age couple Pension Credit saving is stored at survey-year facts, from the engine's own default (Set the mixed-age couple Pension Credit saving at build time, from survey-year ages policyengine-uk-data#519).

Calibration registers (X, Y)

• X. The West Midlands £12,570–15,000 income-tax cell is deferred four weeks in the target-fit register (approved 2026-10-06, expires 2026-11-03). The target stays bound in calibration.
• Y. Two UC payment bands that policyengine-uk 2.122.2 empties join the UC caseload runs ~40% below administrative counts, capping aggregate reform estimates policyengine-uk-data#452 measure exclusions (approved 2026-10-07, expiring with that class on 2026-11-26).

The diagnoses are under Measurement.

From the review (R1, rounds 2 and 3)

• R1. The access-fund repair is recorded. The root stage's checkpoint evidence now carries the paid awards by period code, the threshold and the repaired count. A tab with calendar-month awards but no annual-coded award refuses.
• C6's linked share is a household-weighted average, np.average over the linked rows with the gross4 household weights. It is not a microdf MicroSeries, because microdf reaches an environment only with an engine and the FRS spine runs in the engine-free lane.
• B0's forward-year note cites pe-uk#2173.

Where this departs from the plan

• C5 reads children too. The child tab asks the same question, and Tax-Free Childcare reads is_blind for children.
• C6 puts one identity-keyed draw on the root stage, and the stage notes now say so.
• B1 keeps storing the flag, aligned to the formula, rather than handing it to the formula. The incumbent coverage manifest marks it required.
• B2 leaves the UC take-up population to Max's uk-take-up-per-person-spa branch.
• B4 is rebased on UK FRS spine: stop counting lodgers' rent and property losses as property income; sub-letting rent for every tenure #1081, which merged first. The householder's receipt stays uncredited, as UK FRS spine: stop counting lodgers' rent and property losses as property income; sub-letting rent for every tenure #1081 documents, because policyengine-uk 2.122.2 has no householder-side input for it.

Measurement

Each arm was built and calibrated locally on Chronicle 825406f. The control is A9, main's code. The arms predate the rebase onto #1081, so their spines lack its property-income rules.

• Arm E (C1–C6). Loss 0.00702 (A9 0.00715), 99.0% of targets within 10%, all 7 gates pass.
  • The new Housing Benefit row calibrates to within 0.1%.
  • The LCFS imputation is value-identical wherever its predictors did not move.
• Arm B0 (the engine bump). Loss 0.00769, 98.4% within 10%.
  • uk_target_fit fails on two UC payment bands that hold no benefit unit under policyengine-uk 2.122.2: singles at £26,400–27,600 and childless couples at £24,000–25,200.
  • Each rests on fewer than 10 FRS households. The same spine under policyengine-uk 2.100.0 keeps every supporting award in its band, so the engine alone moves them out. It does so through three corrections:
    • the published LHA rates move awards to neighbouring bands;
    • the 2019 mixed-age couple saving keeps couples who report Pension Credit off UC;
    • Carer Support Payment counts as unearned income, as Carer's Allowance does.
  • Y excludes both bands, beside the 19 of this class already excluded.
  • Align Child Benefit claim exports and opt-out draws with the UK model #1107's checks: the Child Benefit export is registered claims, and the fully charged opt-out pool is exhausted with a 27.4k-family shortfall. obr.child_benefit calibrates to 0.0%.
• Arm head (B5). Loss 0.00765, 98.3% within 10%.
  • Beyond the two UC bands, the West Midlands £12,570–15,000 income-tax cell (hmrc.spi_region.income_tax_by_region_12570_15000@E12000005) fails at +27.4%. It was +8.9% on arm E and +23.6% on B0.
  • The cell's design value is 6.1 times its target, and 6.0 times on main's code. Every other region's cell is 0.6–1.1 times.
  • The cause: the regional cells slice on policyengine-uk's total_income, which leaves out other_investment_income although income_tax charges it.
    • So fewer than 10 SPI records with mostly other investment income sit in the band with their whole tax.
    • Banded on total income including other investment income, the cell is in line with the other regions.
  • X defers the cell until policyengine-uk's total_income includes other investment income. That fix belongs upstream: total_income cites ITA 2007 s. 23, which counts all income charged to income tax.
• With X and Y, replaying the target-fit gate on the head arm's evidence passes, and the other six seam gates passed on that arm.
  • The replay keeps the head arm's weights, which were solved with both UC bands bound.
  • A calibration of this head, the national build and the certifier have not been run.

Full numbers are in experiments/1095-uk-data-ports-2-receipts.md.

Testing

• Each commit: touched-surface tests.
• B0: the full suite on 2.122.2.
• B5: the full suite, 13,378 passed and 128 skipped. The two failures are environmental:
  • the extras test expects no UK engine in the venv;
  • the live Hugging Face pointer test fails on main too.
• X and Y: the register, gate, target and graph tests, 276 passed.
• Rebased head, on main e3e3d881f:
  • the engine-free lane's 509 test files: 14,929 passed and 133 skipped, with the same two environmental failures;
  • the engine-uk lane's 40 test files, run in full: 170 passed. Its Add the liquid financial assets concept, schema version 2 (NZ v0, G3a) #1120 liquid-asset test had failed on 2.122.2 since the previous rebase.
• Review rounds 2 and 3: the FRS spine and lock-level tests pass (115), and no generated surface moved. The rebase onto Give each country its own policyengine-core pin (#1086) #1143's merge left the tree unchanged.
• US lane (policyengine-uk 2.102.4+ needs policyengine-core 3.32.9+, which the certified US default refuses #1086): on the shared lock, test_batched_scoring_matches_whole_pool_policyengine_us failed on both Pythons. It's a Maryland child-care rate lookup under policyengine-us 2.2.1 on core 3.32.19. With the per-country pins (Give each country its own policyengine-core pin (#1086) #1143), a US install from the new lock resolves core 3.32.5, and test_us_post_export_scoring.py passes there (4 tests). The lock-level test passes on this head.
• Lint: ruff check and ruff format --check pass on all changed Python files.
• CI plan: tools/ci_test_plan.py verify and tools/graph_acceptance_burndown.py --verify pass.

Receipts: experiments/1095-uk-data-ports-2-receipts.md.

🤖 Generated with Claude Code

[vahid-ahmadi]

Automated review pass (Claude Code, high effort) — round 1 at a0f292ae

Verdict: nothing in the code blocks it, but it isn't mergeable yet. That's for reasons the description already lists: the #1086 ruling, the signed-difference entries, and a calibration, national build and certifier run of this head. The ports read correctly against their uk-data counterparts. The engine bump is measured separately from the data fixes (arm E against arm B0), which is the right shape. Two of the calibration-register changes treat a defect as a deferral, and I'd fix those at source.

Findings

1. Should-fix: X defers a cell whose cause is a defect in the band variable.
   • The receipts diagnose it: the regional income-tax bands slice on policyengine-uk's total_income, which leaves out other_investment_income, while income_tax charges it. So a few SPI records carry their whole tax into the £12,570–15,000 band: design £339.2m against a £55.1m target (experiments/1095-uk-data-ports-2-receipts.md:100).
   • That's a measurement bug in every regional band, not a West Midlands quirk. It shows up there first because that cell is thinnest.
   • Fix: band on total income including other_investment_income (or on the engine's adjusted_net_income, if that matches the SPI band definition) and drop the deferral. Or keep X, but open an issue now and link it from the register entry, so the four-week expiry has a fix to land against.
2. Should-fix: Y adds two more single-household UC payment bands to an exclusion class that keeps growing.
   • With these, 21 cells of the uk-data#452 class are excluded. Each engine or spine change pushes another one-to-nine-household cell out of its £1,200 band. A census of the class with pooled bands (or a minimum-support rule applied when the registry is compiled) would stop it costing one signed entry per release.
   • Fine to keep Y for this PR; worth an issue.
3. Should-fix: B4 moves CVPAY off the payer, but the householder's receipt is still uncredited, and UK FRS spine: stop counting lodgers' rent and property losses as property income; sub-letting rent for every tenure #1081 removes the same line.
   • rent_paid_as_boarder / rent_paid_as_lodger on the payer is right. But as on UK FRS spine: stop counting lodgers' rent and property losses as property income; sub-letting rent for every tenure #1081, the householder's benefit unit never gets the receipt. The description notes the conflict ("whichever lands second rebases").
   • Suggest: settle the order now. Land UK FRS spine: stop counting lodgers' rent and property losses as property income; sub-letting rent for every tenure #1081 first (it's smaller and reviewed), then rebase B4 on it, so the CVPAY change is reviewed once.
4. Question, for policyengine-uk 2.102.4+ needs policyengine-core 3.32.9+, which the certified US default refuses #1086: the bump pulls in about 600 policyengine-uk commits.
   • The rule changes that move this build include:
     • the published LHA rates (#2022/#2134);
     • the mixed-age couple saving (#1940/#1944);
     • Carer Support Payment as unearned income;
     • the Child Benefit opt-out gate (#2140), which activates Align Child Benefit claim exports and opt-out draws with the UK model #1107's export;
     • the legacy severe disability premium (#1946);
     • is_claimant_or_partner (#1896);
     • the NICs threshold freeze (#1959);
     • BADR's own elasticity (#1980), which matters for forward years only.
   • Not in 2.122.2: the working-age HB abolition (#1910), the published DWP rates (#1925), the HB/CTR non-dependant deductions (#2007), the Pension Credit earnings disregards (#2019), and the property finance-cost relief (#2172).
   • Also still in 2.122.2: salary_sacrifice_returned_to_income flows into employment_income and so into adjusted net income from April 2029. HMRC says the 2029 cap changes NI only. That doesn't touch the 2024-25 build, but it does affect forward-year users of this release, so it's worth a line in the release notes.
5. Question: is the Align Child Benefit claim exports and opt-out draws with the UK model #1107 opt-out shortfall expected? The fully charged pool gives 599.6k opt-out families against a 627.0k target, 27.4k short (receipts:90). obr.child_benefit still fits, so the payment side is fine. But it means the dataset has about 4% fewer registered opt-outs than HMRC's count. Is the rest meant to come from the taper pool now that policyengine-uk pays taper opt-outs, or is the shortfall accepted?
6. Nit: C2's repair threshold depends on the data, and is silent when there's no annual-coded award.
   • The rule reads a calendar-month award as annual when it annualises to more than twice the largest annual-coded award on the tab, so the cut-off moves with each survey year. A tab with no annual-coded award "repairs nothing" without saying so.
   • Suggest: record the repaired count and the threshold in the stage evidence, and raise when the tab has calendar-month awards but no annual-coded ones, instead of passing silently.
   • The arithmetic for the one 2024-25 record is right: £4,500 reported per month, weeklyised to £1,035.62, read as annual gives £1,035.62 × 52/12 ≈ £4,488.
7. Nit: C6 weights the linked-claim share by hand (gross4 sums). That's consistent with the stage's other derived shares and is computed from the data (0.324). The rule follows reg 63 as substituted from 23 September 2020 (SI 2019/1152): no longer limited to new trades, applied at claim start or when a new trade is reported. The limitations listed in the docstring (re-claims, five-year repeats, work-related-requirements group moves) are honest.

Checked and correct

• C1: the regenerated target_references.json carries the same 5,028 numeric values as before, as a multiset; only labels and metadata move. I couldn't rerun the byte-for-byte regeneration test, because the pinned Chronicle consumer feed isn't on this machine, so it's skipped locally.
• C4: 3/12 × £6,851.0m + 9/12 × £7,114.7m = £7,048.8m. The source_measure_id_by_opening_year selector picks each fiscal year's own DWP measure.
• C3: the LCFS counts are named for what they are and derived from age. The receipts show the imputation is value-identical where predictors didn't move.
• B0: the lock is policyengine-uk 2.122.2 and policyengine-core 3.32.19. The pins, the coverage review, the H2 fixture dependency and APPROVED_UV_LOCK_SHA256 are re-derived.
• Measurement: arm E (C1–C6 on the locked engine), arm B0 (the bump) and the head arm are each calibrated against the A9 control. The gate failures are diagnosed record by record.
• Tests: the 17 test files the PR touches pass locally on 2.122.2 / 3.32.19.

Before merge (from the description, still open)

• The policyengine-uk 2.102.4+ needs policyengine-core 3.32.9+, which the certified US default refuses #1086 ruling, then a rebase.
• The signed-difference entries for C5, C6, B1 and B2–B5.
• A calibration of this head (the target-fit replay kept arm head's weights), the national build and the certifier.

[juaristi22]

Thanks, Vahid. Round 1 is addressed on the rebased head; details by item below.

Rebase. The branch is rebased onto main after #1081, and CI's engine-free failure is fixed:

• The failing test was test_fully_armed_battery_evaluates_gate_for_gate. Its fixed register clock predated X's approval, so X read as not yet in force. X and Y now move the clock forward, as the test's note asks.
• The engine-free job's test files pass locally on the rebased head, apart from the two that need a live Hugging Face pointer or a venv without the UK engine. CI is rerunning.
• The rebase changed nothing else in the branch's own diff.

1. X and the band variable. Agreed, it's a measurement bug, and wider than the regional cells. 47 HMRC targets band on policyengine-uk's total_income:

• the 30 regional SPI cells;
• 14 national SPI band tables, which group by it;
• 3 ITL rows.

I measured the fix through the calibration's own measure path on the head spine, adding other_investment_income to the band. The West Midlands £12,570–15,000 cell goes from 6.15 times its target to 0.92 at design weights. The other 119 regional income-tax cells move by 8% at most.

The fix belongs upstream:

• total_income cites ITA 2007 s. 23, which counts all income charged to income tax. But it leaves out the other investment income that its own income_tax charges as savings interest.
• That also understates post_tax_income, and the incomes that Winter Fuel Payment's test and the Housing Benefit and Council Tax Reduction non-dependant deductions read.

I'm drafting a policyengine-uk issue, and X's register entry will link it once it's filed. X keeps the cell bound in calibration and expires on 2026-11-03.

2. Y and the growing class. Kept for this PR. Agreed the class needs a rule rather than one signed entry per sparse band. The options are:

• pooling sparse bands when the registry is compiled;
• a minimum-support rule;
• support for high awards (UC caseload runs ~40% below administrative counts, capping aggregate reform estimates policyengine-uk-data#452).

3. B4 and #1081. #1081 merged first, and B4 is rebased on it, so the CVPAY line is now #1081's own frs_property_income. B4 keeps only the payer's rent_paid_as_boarder / rent_paid_as_lodger and the shared-rent liability.

The householder's receipt stays uncredited, as #1081's docstring records. policyengine-uk 2.122.2 has no householder-side input for lodger or boarder receipts. It has only the payer's two columns and pays_rent_to_householder. So crediting it would start with an engine variable.

4. The bump. Thanks for the list. B0's changelog fragment now carries the forward-year note:

policyengine-uk 2.122.2 models the April 2029 salary-sacrifice cap by adding sacrifice above £2,000 (salary_sacrifice_returned_to_income) to employment_income, so it also raises adjusted net income from 2029-30. The government's change applies to National Insurance only, so for 2029-30 and later the rules that read adjusted net income (the personal allowance taper, the High Income Child Benefit Charge, Tax-Free Childcare eligibility) see higher incomes for people sacrificing more than £2,000. The 2024-25 build is unaffected.

5. The opt-out shortfall. It's expected under #1107's contract (docs/uk-child-benefit-claim-contract.md):

• With policyengine-uk 2.122.2's default opt_out_charge_share of 1, only fully charged families can opt out.
• The stage records pool_shortfall_families and pool_exhausted instead of drawing taper families, because the installed model would pay them.
• A share below 1 would admit taper families. That's an engine parameter, not a data change.

So the 27.4k shortfall is recorded, not filled, and obr.child_benefit still fits.

6. C2's threshold. Done in d083f8ff5:

• access_fund_annual now refuses a tab that has calendar-month awards but no annual-coded award.
• It records the paid awards by period code, the threshold and the repaired count.
• The root stage carries that record in its checkpoint evidence, beside the capital sentinel count.

7. C6. Thanks; no change.

Still open before merge:

• the policyengine-uk 2.102.4+ needs policyengine-core 3.32.9+, which the certified US default refuses #1086 ruling;
• the signed-difference entries;
• a calibration of this head, the national build and the certifier.

[vahid-ahmadi]

Automated review pass (Claude Code, high effort) — round 2 at 980ffef0

Verdict: nothing blocking in the code; round 1 is answered. Two small follow-ups: link the two upstream issues filed since round 1, and regenerate the concept-coverage files against main. What stands before merge is what the description already lists: the #1086 ruling, the signed-difference entries, and a calibration, national build and certifier run of this head.

Round-1 item	Status	Evidence
1. West Midlands deferral is a band-variable defect	Agreed; link the issue	Confirmed upstream at policyengine-uk 2.122.2. total_income (variables/gov/hmrc/income_tax/total_income.py, cites ITA 2007 s. 23) adds savings_interest_income but not other_investment_income, while taxable_savings_interest_income charges both. It's now filed as policyengine-uk#2174. The X entry in uk/target_fit_reviewed_exclusions.json explains the defect but cites only microcosm#1095. Please add #2174 so the deferral points at the fix that retires it before 2026-11-03. Your 6.15× → 0.92 measurement on the head spine, with the other 119 regional cells moving at most 8%, is a convincing case that the band, not the data, is wrong.
2. Single-household UC band exclusions (Y)	Fine as stated	Kept for this PR. A pooling or minimum-support rule is the right follow-up; worth a tracking item on #1095 so it doesn't stay one signed entry per release.
3. B4 overlap with #1081	Closed	The branch contains the #1081 merge (a50b9cc0). CVPAY handling is only #1081's frs_property_income; B4 keeps the payer's rent_paid_as_boarder/rent_paid_as_lodger and the shared-rent liability. The uncredited householder receipt is documented in #1081's docstring, and the engine has no householder-side input yet.
4. Engine bump and the 2029 salary-sacrifice note	Closed; one link	The B0 changelog fragment's forward-year note is accurate. It would help to cite policyengine-uk#2173, filed today for exactly this: the fix routes the excess to NI-able earnings only.
5. Child Benefit opt-out shortfall (27.4k)	Closed	Recorded rather than filled, as #1107's contract sets out (pool_shortfall_families, pool_exhausted, default opt_out_charge_share = 1), and obr.child_benefit still fits.
6. Access-fund repair threshold	Closed	d083f8ff5: access_fund_annual (frs_spine.py:1390) refuses a tab that has calendar-month awards but no annual-coded award. Its evidence records paid awards by period code, the threshold and the repaired count, and the root stage carries that in its checkpoint evidence. The tests cover repair, refusal and the evidence record (test_uk_frs_spine.py:1062, :2954-3013), and they pass locally.
7. C6 linked-claim share hand-weighted	Still open (nit)	The reply says "no change", but the round-1 nit was about the weighting, not the rule. frs_spine.py:1353-1362 still computes the linked share from a hand-built gross4 array (weight[linked].sum()). Under the microdf-throughout convention, a MicroSeries over the person rows with household weights would do it. Not blocking.

Rebase and CI

• The engine-free fix is legitimate. In test_uk_battery_bindings.py::test_fully_armed_battery_evaluates_gate_for_gate, the fixed register clock predated X's approval, so X read as not yet in force. Moving the clock forward is what the test's note asks. It passes locally.
• New: conflicts with main. Main has moved by Add Axiom graph adapter surfaces and simulate.rules_by_ref@1 (NZ v0, G1) #1118 and Add the liquid financial assets concept, schema version 2 (NZ v0, G3a) #1120 (the NZ hub: Axiom adapter and the liquid financial assets concept). git merge-tree shows two conflicts, both in generated files: docs/concept-coverage/policyengine-uk.md and packages/microcosm-frame/tests/golden/concept-coverage/policyengine-uk.json. Regenerate them after merging main; nothing hand-written conflicts.
• CI at this head: wheels passed. engine-free (3.13, 3.14), lint and select-countries were still pending when I checked.

[vahid-ahmadi]

Automated review pass (Claude Code, high effort) — round 3 at 692e8b7b

Verdict: nothing blocking in the code; #1086 is settled and CI is green on every lane. What stands before merge is what the description lists: the signed-difference entries, and a calibration, national build and certifier run of this head. One question on how the core-pin commit lands.

What changed since round 2 (git range-diff of a50b9cc0..980ffef0 against e3e3d881..692e8b7b)

• New first commit, P (261bb8cd): per-country policyengine-core pins, per the policyengine-uk 2.102.4+ needs policyengine-core 3.32.9+, which the certified US default refuses #1086 ruling (option 2).
• B0 (eba82ce4): now moves only the UK side of the lock, and reviews the liquid financial assets concept's unmapped reason against 2.122.2. That reason now names car_list_price, lifetime_isa_balance and UC's claimant-and-partner capital split.
• Every other port commit is either identical or differs only in the regenerated gate-battery digests from the two rebases.
• The receipts file records the two extra rebases and the US-lane evidence.

Findings

1. Should (process): land P as its own PR, or confirm with the policyengine-uk 2.102.4+ needs policyengine-core 3.32.9+, which the certified US default refuses #1086 ruling's author that it can ride here. The ruling's last instruction was "Keep it to the lock. The PR should change no engine version. The policyengine-uk bump follows as its own PR." P itself matches that exactly:

   • In uv.lock, engine versions are identical at e3e3d881 and 261bb8cd: core 3.32.5, policyengine-uk 2.100.0, policyengine-us 2.2.1, spm-calculator 1.0.0.
   • The fork appears only at B0: core 3.32.5 and 3.32.19, policyengine-uk 2.122.2.

   But here P ships in the same PR as B0's bump and fourteen ports. The description already says P can be cherry-picked alone. I'd do that: P goes in fast as a small lock-only PR, and this PR rebases onto it.

2. Nit: B0's forward-year note still doesn't cite pe-uk#2173 (changelog.d/1095-policyengine-uk-2-122.changed.md). This was round 2's item 4. #2173 is the issue for the 2029 salary-sacrifice excess reaching ANI, so the note should point at it the same way X points at #2174.

3. Nit, carried from round 2: C6's linked share is still hand-weighted. At frs_spine.py:1355-1359, weight[linked].sum() is taken over a raw gross4 array. A household-weighted MicroSeries would follow the microdf-throughout convention. Not blocking.

Round-2 items

Item	Status
Link pe-uk#2174 from X	Done. The register reason, the receipts and test_uk_battery_bindings.py (assert "pe-uk#2174" in deferral.reason) all cite it. X expires 2026-11-03.
Regenerate concept coverage against main	Done. Regenerated on the new concept schema (d824f5e97d58), with 285 inputs and 40 covered. The branch merges cleanly into main.
Cite pe-uk#2173 in B0's note	Open (nit 2).
Y's sparse-band class needs a rule	No tracking item on #1095 yet. Worth one line so it doesn't stay one signed entry per release.
C6 hand-weighted share	Open (nit 3).

#1086 / P checked

• [tool.uv] conflicts pairs every US-installing extra with every UK-installing extra. The US-installing extras are microcosm-build[us], microcosm-data[us] and microcosm-frame[policyengine]; the UK-installing ones are each package's [uk].
• No other pyproject.toml in the workspace installs an engine.
• CI installs one extra per lane (--extra us, --extra uk), which the conflicts allow.
• APPROVED_UV_LOCK_SHA256 is re-pinned at P and again at B0.

Tests run locally at 692e8b7b (fresh worktree, uv sync --all-packages --locked, the engine-free environment)

• test_engine_core_pins.py: 14 passed.
• All 18 engine-free test files this PR touches: about 1,220 passed, 7 skipped, 0 failed.
• ruff check .: clean.
• ruff format --check: none of the PR's files would be reformatted.
• CI at this head is green: lint, wheels, select-countries, engine-free, engine-us and engine-uk (each on 3.13 and 3.14), and integration-uk.

Still before merge (from the description, and unchanged)

• The signed-difference entries for the net-new columns (C5, C6, B2–B5) and for B1's mechanism change. The drafts are still outside the repo.
• A calibration of this head, the national build and the certifier. The receipts' head arm was built as 65a08ddb6 with both UC bands bound; the Y replay reuses those weights.

[juaristi22]

Thanks, Vahid. Rounds 2 and 3 are addressed: #1143 has merged, and this PR is rebased onto it with the fixes (05b8f073d).

Round 3

1. P as its own PR. It landed as Give each country its own policyengine-core pin (#1086) #1143, so it's no longer part of this PR.
2. Salary sacrifice above the April 2029 NI cap raises adjusted net income, but HMRC says the cap leaves ANI unchanged policyengine-uk#2173. B0's forward-year note now cites it, the same way X cites total_income leaves out other_investment_income, which income_tax charges policyengine-uk#2174.
3. C6's linked share. It's now a household-weighted average over the linked rows: np.average with the gross4 household weights. That is the idiom microcosm's stages already use for weighted shares (provenance_flags.py, frs_take_up.py).
   • I didn't use a MicroSeries. microdf only reaches an environment with an engine, because policyengine-core, -uk and -us depend on it. The FRS spine runs in the engine-free lane, so microdf would become a new runtime dependency of microcosm-build, and the lock would move for one share.
   • The FRS spine tests pass (101), and no generated surface moved.

Round 2

• Y's sparse-band class. Agreed: it needs a rule rather than one signed entry per release. The rule belongs in the follow-up PR, with a tracking item on UK: track the defects uk-data is fixing that microcosm still has #1095.
• The total_income leaves out other_investment_income, which income_tax charges policyengine-uk#2174 link and the concept-coverage regeneration were already in the head you reviewed in round 3.

Before merge

• A calibration of this head is running. Its results go in the receipts.
• The signed-difference entries, the national build and the certifier follow the merge, as they did for UK: port the unblocked uk-data fixes tracked in #1095 #1100.

[juaristi22]

Calibration of this head (arm head-r3). Built locally from c7fe76e8a, the same tree as the pushed head 05b8f073d, on Chronicle feed 825406f, with local staging only.

• Fit. Loss 0.00712, against 0.00715 for arm A9 (main's code without this PR). 98.3% of 1,167 targets are within 10%, and the ESS is 4,874.
• Gates. With X and Y in force, all 7 calibrated-seam gates pass, and so do the exported file's terminal gates. Unlike the earlier replay of Y, this solve runs without the two UC payment bands.
• Input mass. The certifier's input-mass gate passes, with no stale, expired or dormant exclusion.
• Rows to watch:
  • The West Midlands £12,570–15,000 income-tax cell is at +28.9%. X defers it until 2026-11-03, and Include other investment income in total income policyengine-uk#2184 is the open upstream fix.
  • obr.vat is at +24.1%, under its 25% bound.
  • The new dwp.hb.amount_pension_age row calibrates to 0.0%.
• C6. The weighted-share rewrite leaves uc_is_in_startup_period unchanged: no value differs on the 142,129 person rows this spine shares with the earlier head arm.

The receipts in experiments/1095-uk-data-ports-2-receipts.md still list this calibration as owed. The follow-up PR records it there, so this PR needs no further CI run. The signed-difference entries, the national build and the certifier follow the merge, as for #1100.