[juaristi22]

The data fixes for the 22 October national and dense build: #1324's calibration-period fix, #1319's Carer's Allowance draw, and the items #1278 lists before the pre-Budget build. It is one PR with ordered commits and one pin pass at the end, so each commit reads on its own. It must be on main by 21 October 23:59.

Closes #1324. Part of #1319: the draw lands here; its stage-health gate, the overlap rule and a bound caseload target follow after the Budget. Part of #1320: ALLPAY2 is counted once here, and the spousal-maintenance split waits on PolicyEngine/policyengine-uk-data#627. Part of #1278 and #1095.

Summary

#1324. Stored value columns measured at their binding's period

• The bug. A plain-column binding read its value column off the frame, which stores it at the survey period (2024), and compared it with a 2025 target. UK national calibration reads stored input columns at the frame period: 76 £ rows compare 2024 values with 2025 targets #1324 lists 76 £ rows that compared 2024 values with 2025 facts: SPI component amounts, statutory maternity pay, workplace pensions and others.
• The fix. A binding's value variable and the terms of its value_expression are now measured at the binding's period: its measurement_period, else the calibration period.
  • A frame-stored value off the frame's period goes through the measure resolver to the engine at 2025, as every engine-computed variable already does.
  • Engine values for stored names go in under a reserved engine_period__ name. Filters, conditions and bands keep their stored reads, and the stored columns are never overwritten.
  • A declared period that is neither the calibration period nor the frame's period refuses.
  • The profit after loss relief reads the frame only at the frame's period.
• Rows that declare the frame's period.
  • ONS road fuel binds its calendar-2024 figure on the frame's own 2024 petrol and diesel columns, under a signed matched_period hold. The row no longer depends on the engine's fuel projection.
  • The seven bus and rail outturns held without an index declare the frame's period. This keeps their measurement and makes it explicit.
  • The England fare-receipt rows and the Northern Ireland bus receipts carry 2025 facts, so they go to the engine at 2025.
• England bus fares. The England fare-receipt row now ends 9.6% above its target (10.1% with the rail row bound), against 2.3% below on the control's stored 2024 read. The cause is in policyengine-uk's pricing, not in the data:
  • person_bus_fare_spending prices each fare-paying boarding at the 2024-25 yield, and moves only single tickets in England outside London with the cap. It gives every single ticket one uncapped fare, £2.748, inferred from the 2024-25 cap grant.
  • So the cap's move from £2 to £3 in January 2025 raises every single fare outside London from about £2.20 (day-weighted over 2024-25) to £2.75: about a quarter, at the default single-ticket share. England's engine total rises about a fifth from 2024 to 2025 at design weights. DfT's local bus fares index for England, which uprates the target, rose 5.7% (FY2024-25 to CY2025).
  • The weights cannot correct a price per journey. Binding the row at the frame's 2024 period would change only the weights: the engine recomputes fares from journeys in every year, so 2025 and reform runs keep the cap assumption either way. María ruled on 2026-10-20 to keep the engine route.
  • The same assumption gives every single journey outside London the full saving under a lower cap. So until the engine changes, a costing of a cap change likely overstates its cost and gains. The engine fix is Bus fares: follow published fares by area and year, and price cap changes from DfT cap savings instead of one uncapped fare policyengine-uk#2542, and its pull request Bus fares: follow published fares by area and year, and price cap changes from a fitted fare table policyengine-uk#2546, approved and not yet released; UK bus fares: keep in step with the engine's per-year fares, take the single-ticket share from cap-regime years, and check cap savings against NTS fares paid #1331 tracks the Microcosm side. The UK national calibration reads stored input columns at the frame period: 76 £ rows compare 2024 values with 2025 targets #1324 changelog fragment carries the same caveat.
• The check. tools/check_uk_measurement_periods.py rebuilds each plain-column row's route from a finished run. It writes a receipt of periods and routes, and names any stored value read at a period the frame does not hold. It warns and does not fail, unless --strict is given. On the 19 October national build it names the rows UK national calibration reads stored input columns at the frame period: 76 £ rows compare 2024 values with 2025 targets #1324 lists.

#1278. Register items

• ONS single years of age 0, 1 and 2 for all persons, on the 15–19 split's precedent (2da556a5a). The references go from 1,319 to 1,322.
• The plan-2 exclusions are renewed on AY2024/25 (expiry 2026-11-26), and two stale trackings are re-pointed.
• The Band D parity exclusion names the per-band council-tax amounts, so the next re-pin does not port them as targets.
• The ORR rail-support row is held out of the national objective. At design weights the household rail subsidy sits about three quarters below ORR table 7270's total, and the solve closed that gap by moving weight onto rail-using households. The 7270-versus-7271 basis ruling stays open (Bind DfT, HMRC, DESNZ, Ofgem and ORR facts for bus fares, road fuel, domestic energy and rail support (#790 ruling: bind facts and calibrate) #890 G, WS-C deferred adjudications and the reconciliation layer #736).

Spine riders

• State Pension at 52 weeks. policyengine-uk reads state_pension_reported back at 52 weeks, so 365.25/7 put every full-rate pensioner 0.34% above the full rate. This is an SP-only exception to UK: track the defects uk-data is fixing that microcosm still has #1095's weekly amounts.
• Paying households' zero council-tax bills. A household that pays council tax, reports no reduction and carries a zero derived bill now takes its cell mean, as a missing bill does. The bill level waits until after the Budget.
• ALLPAY2 is counted once, as miscellaneous income. A new test refuses any FRS field that feeds two person income columns.

#1319. Carer's Allowance claims for carers over the earnings limit

• would_claim_carers_allowance followed reported receipt. A carer over the earnings limit cannot report the allowance, so a reform that widens it reached no new claimant. The April 2025 rise of the limit from £151 to £196 a week brought about 61,000 more people into eligibility (Independent Review of Carer's Allowance Overpayments, July 2025, chapter 7).
• A new ungated spine stage, carers_allowance_take_up, keeps reported receipt as the anchor. It opens the flag at one rate (0.6, sensitivity 0.45 to 0.75) to carers outside Scotland whom only the earnings test keeps from the allowance. The draw is keyed on the person's identity.
• Scotland stays closed, since Carer Support Payment reads the same flag and policyengine-uk does not yet apply its earnings limit. In law Carer Support Payment has an earnings limit, as the allowance does (Carer Support Payment: apply its earnings limit policyengine-uk#2548).
• The receipt is aggregate-only. The stage has no health gate, and the flag stays out of the terminal take-up signal gate.

Hygiene

• The SPI band receipt and 23 exclusion reasons are restated in aggregate terms, with the same meaning, decisions and windows. A disclosure-control item went to the author separately.
• The uk.full.problem and holdout kernels review the local binding adjudications on the build's review date, not the wall clock.

Pin pass

• Two commits at the end. The first regenerates the national target references and membership: 1,319 to 1,322, with measurement_period mirrored onto road fuel and the seven held bus and rail rows. The second regenerates the national compile-parity receipt, the release input coverage manifest and the charter-H2 spine fixture; the local receipt was regenerated and does not move.
• The feed stays abf0921, and gates.json does not move.

Verification

• Touched tests locally. The full suite runs in CI.

• Control national at a90eec3cb: 7/7 terminal gates; 98.4% of 1,215 rows within 10%.

• UK national calibration reads stored input columns at the frame period: 76 £ rows compare 2024 values with 2025 targets #1324 arm (the four UK national calibration reads stored input columns at the frame period: 76 £ rows compare 2024 values with 2025 targets #1324 commits on the control spine): 7/7 terminal gates; 98.8% of 1,215 rows within 10% (control 98.4%). The measurement-period receipt names no stored value read off its period; on the control it names 78 rows.

  • The SPI employment amounts improve most: their mean absolute error falls from 6.1% to 3.5%, and six SPI band rows come back inside 10%.
  • One row moves outside 10%: England's bus fare receipts, at +10.1% (see England bus fares above). The control hid the gap, because it fitted the frame's 2024 values.

• Rail hold-out arm (the UK national calibration reads stored input columns at the frame period: 76 £ rows compare 2024 values with 2025 targets #1324 arm with the ORR row held out): 7/7 terminal gates; 99.1% of the 1,214 rows both arms bind are within 10% (98.8% with the row bound). Three rows come back inside 10%, England's bus fare receipts among them (+9.6%), and none goes out.

• Spine twin (the spine riders and UK Carer's Allowance: would_claim follows reported receipt, so earnings-limit reforms reach 4,600 new carers against an official 61,000 #1319) against the control spine:

  • 39/39 spine gates on both.
  • Property purchases: the non-residential value fit is 82.5% (control 81.5%; floor 80%).
  • State Pension: the full-rate amount now sits at the 52-week figure, and the share of recipients above the full rate falls by about a quarter.
  • Council tax: the regional zero-bill outliers in England fall into line with the other regions.
  • Carer's Allowance: the realised draw share is 0.61 against the 0.6 rate, and Scotland stays closed.

• Combined arm (every commit, on the twin spine): 7/7 terminal gates, and the measurement-period receipt names no stored value read off its period.

  • On the 1,214 rows it shares with the rail hold-out arm, 98.8% are within 10% (rail arm 99.1%). Three rows that sat at 9–10% move just past 10%: two low SPI employment bands and one State Pension count band. None moves far.
  • The three ONS age rows fit to within 0.1%.
  • Carer's Allowance: the obr.carers_allowance gap before the solve narrows from −11.6% to −10.1%, and the UC carer element from −15.3% to −14.0%.

• Review round 1 (Vahid, 21 October; no blocker): four follow-up commits and a pin commit.

  • A declared frame period on a value the frame does not hold now refuses.
  • The Carer Support Payment reason is corrected.
  • The council-tax wording quotes the FRS methodology.
  • The UK national calibration reads stored input columns at the frame period: 76 £ rows compare 2024 values with 2025 targets #1324 changelog carries the bus fare caveat.
  • The coverage manifest and the H2 fixture are regenerated.

For review

• Register entries signed juaristi22 / 2026-10-20: the road-fuel matched_period hold (expires 2026-12-11), the plan-2 renewals (2026-11-26) and the rail entry (2026-12-11).
• The council-tax docstring and changelog fragment now quote section 1.3 of the FRS 2024-25 methodology. It does not say whether CTPAID is asked or imputed.

Not in this PR (after the Budget)

• The C1 diagnostics schema field (Publish a JSON Schema for calibration diagnostics schema 8 #1298).
• Carer's Allowance: a stage-health gate, the overlap rule and a bound caseload target.
• HICBC, Marriage Allowance, the Plan 2 stock, car availability, duty quantities, the council-tax level and the tape-record key.
• The engine's bus fare cap pass-through: a spread of uncapped single fares, fitted to the 2024-25 grant and to DfT's fares index (Bus fares: follow published fares by area and year, and price cap changes from DfT cap savings instead of one uncapped fare policyengine-uk#2542, Bus fares: follow published fares by area and year, and price cap changes from a fitted fare table policyengine-uk#2546, UK bus fares: keep in step with the engine's per-year fares, take the single-ticket share from cap-regime years, and check cap savings against NTS fares paid #1331).
• London's fare row starts about a fifth under its target at design weights in every run, and the solve closes it by moving weight onto London bus users. This predates this PR.

🤖 Generated with Claude Code

[vahid-ahmadi]

@juaristi22 thanks, this is my review at 7e5c8b37cd40737b91cc775b0c6b4b60aa6022c7.

Verdict. I found nothing that blocks the 22 October build. The #1324 period change does what the body says, and I traced the engine route end to end: a stored column read at 2025 goes to a pe-uk simulation that extends the 2024 dataset with extend_single_year_dataset, so employment_income and the other stored inputs come back uprated, not carried forward. Other builds are unaffected. The Carer's Allowance draw is sound in its mechanics, and the 61,000 figure checks out against the Independent Review. But the 0.6 rate has no empirical anchor yet, and the eligible pool is wider than the people the 61,000 counts. So any costing of a Carer's Allowance earnings-limit change on this build carries that assumption. I have six "should" items and three nits below. Only finding 1 is a code gap, and it does not bite any row today.

What I ran

• Worktree at the head, with a fresh uv sync --all-packages --locked --extra uk (policyengine-uk 2.134.4, policyengine-core 3.33.3).
• The touched tests and their neighbours: 16 files, engine-free plus the two touched engine tests (CA take-up, loss relief). 408 passed, 6 skipped.
• tools/ci_test_plan.py verify: ok.
• CI at the head: all 10 checks green.
• The new period tests against the merge base: they import symbols the base does not have (ENGINE_PERIOD_COLUMN_PREFIX), so they cannot run there. This shows the tests are new, not that they would catch a regression.
• Read the whole diff of the runtime code, the contract, the holds and the take-up contract. Read the policyengine-uk code the PR relies on: carers_allowance_pre_overlap, is_gainfully_employed_for_carers_allowance, carers_allowance_earnings, carer_support_payment_pre_overlap, is_carer_for_benefits, the three State Pension variables, extend_single_year_dataset and pe-core's carry-over and uprating branch.
• Sources checked: the Independent Review of Carer's Allowance Overpayments (gov.uk, dated 21 July 2025), DWP benefit statistics February 2026, mygov.scot Carer Support Payment "If you work", and FRS 2024-25 background information and methodology.
• Not run: any national or spine build, tools/check_uk_measurement_periods.py on a real run, and the back-tests. They need the licensed FRS/SPI inputs, which I did not use. So I have not checked the build figures in the body (98.8%, 0.61 realised share, the −11.6% to −10.1% gap).

Findings

1. should target_materialization.py:971 (_refuse_ignored_measurement_period) and ledger_targets.py:1633 (column_at). A binding that declares measurement_period equal to the frame's period passes the guard whether or not its value variable is stored. If the variable is engine-computed (not stored), column_at returns the injected value. The resolver computes that value at the run's calibration period (2025), because it does not read the binding's period. So the row is measured at 2025 while it declares 2024, and nothing refuses. uk_value_read_routes (measurement_period.py:77) then labels it provider:... and does not flag it. This bites none of the nine rows that declare 2024 today: petrol, diesel and the bus and rail columns are stored stage cells (graph.py etb_* cells), and cgt_2024_tax carries its year in its name through _cgt_model_measure. But the next row that declares the frame's period on an engine variable will be wrong silently. Suggest the guard also refuses a declared frame period whose value variable is not in stored_columns (bar the CGT year-named measures), with a test.

2. should carers_allowance_take_up.py, the eligible pool. The pool is persons outside Scotland with care_hours ≥ 35, no reported CA, and over the 2024-25 limit. Two things make it wider than the population the Review's 61,000 counts:

   • The FRS hours field counts care for anyone. Neither the stage nor pe-uk (carers_allowance_pre_overlap) checks that the cared-for person gets a qualifying disability benefit. pe-uk's own is_carer_for_benefits documentation says so.
   • The pool includes every over-limit carer, not just those between £151 and the 2025-26 £196. So the 0.61 "realised share" in the receipt is a share of all over-limit carers, most of whom never become entitled. It is not a take-up rate among the newly entitled (see nit A).

   Neither point changes who is paid for carers who stay over the limit, since the engine still applies the earnings test. But the number paid at baseline 2025 equals 0.6 times the weighted count of pool members who end up under £196, and that count is never reported. The Review gives no take-up estimate (I confirmed: chapter 7 has the 61,000 and nothing on claims). So 0.6, and its 0.45 to 0.75 range, is a judgement call. Ask: add to the receipt the weighted count of pool members entitled at the calibration year, as an aggregate under the existing small-cell rule. That is the earnings-limit back-test the take-up contract already promises. Compare it with the 61,000 before anyone publishes a costing of a CA earnings-limit change on this build.

3. should carers_allowance_take_up.py (module docstring), stage order. The docstring says the stage reads earnings before salary_sacrifice moves pension amounts out of pay. A drawn carer whom that conversion takes under the limit keeps the flag. Because carers_allowance_pre_overlap needs only hours, no gainful employment and would_claim, that person is paid CA in the released file even though they reported none. That changes baseline spending, not just how far reforms reach. The receipt does not count these people. Ask: count them in the receipt now (aggregate), and keep the move after salary_sacrifice on UK Carer's Allowance: would_claim follows reported receipt, so earnings-limit reforms reach 4,600 new carers against an official 61,000 #1319's list. Draw UK QRF stages from uniforms keyed on each record's id, and tie pension contributions to pay #1332 ties pension contributions to pay and also touches frs_spine.py, so this set may move again when it lands.

4. should Scotland stays closed, and the reason given is only half the story. The body and CARERS_ALLOWANCE_ELIGIBILITY say the engine applies Carer Support Payment no earnings test. That is true of carer_support_payment_pre_overlap, but in law CSP has a take-home pay limit: mygov.scot gives £204 a week now, the same as CA. So keeping Scotland closed is the right workaround, but the cause is a pe-uk gap: the CSP earnings test is missing. That gap also lets the engine pay CSP to any reported Scottish recipient over the limit. Suggest rewording to "pe-uk does not yet apply CSP's earnings limit", plus a pe-uk issue for the test. Until it lands, a Scottish CSP earnings-limit reform reaches no new claimant on this file.

5. should England bus fare receipts end +9.6% (+10.1% with rail bound) after the engine route. The body's diagnosis matches the engine. pe-uk #2546 (bus fares by area and year, with a fitted fare table) looks like the fix it is waiting on. Please link #2546 from the body's "Not in this PR" item. Also note in the release notes that cap-change costings on this build likely overstate their cost, as the body says, so nobody reads them as final.

6. should merge order with your other open PRs:

   • Bump the UK lock to the newest policyengine-uk (2.141.1, then the #2523 fix release) #1333 (pe-uk bump to 2.141.1) shares frs_spine.py, graph.py, source_stages.json, sources.yaml, the coverage manifest, policyengine_uk_concepts.py and the spine fixtures with this PR.
   • Draw UK QRF stages from uniforms keyed on each record's id, and tie pension contributions to pay #1332 shares frs_spine.py, spi_income.py, calibration_measure_exclusions.json and two fixtures.
   • This PR relies on pe-uk's is_gainfully_employed_for_carers_allowance, carers_allowance_earnings and the State Pension variables reading state_pension_reported at 52 weeks. A bump can move any of these.

   Suggest this PR first, as it has tonight's deadline. Then Bump the UK lock to the newest policyengine-uk (2.141.1, then the #2523 fix release) #1333, with the measurement-period receipt, the CA receipt and the 52-week State Pension check re-run on that lock. Then Draw UK QRF stages from uniforms keyed on each record's id, and tie pension contributions to pay #1332.

Nits

A. carers_allowance_take_up.py receipt: rename realized_share to something like drawn_share_of_over_limit_pool, so nobody reads it as a take-up rate (see finding 2).

B. The DWP benefit statistics February 2026 page, which the take-up contract cites as caseload_url, disagrees with itself on CA: the text says entitled cases at August 2025 were 4.7% lower than a year earlier, but Figure 2 shows +62,000. That page also gives 1,376,400 entitled, of which 380,000 are entitled but not paid. Use Stat-Xplore, not that page, when the caseload target binds after the Budget.

C. frs_council_tax.py docstring: the FRS 2024-25 methodology confirms that the band and "some other Council Tax block questions were removed from the field interview from April 2024" and that the other answers "are then imputed during DWP's validation processes". It does not say CTANNUAL is derived from the band. Suggest quoting that wording instead of "DWP derives CTANNUAL from the band", and saying whether CTPAID is asked or imputed.

What looks right

• Period routing. Only an adapter that declares stored_columns and column_at (UKFrameTargetAdapter) changes route. Every other adapter, including the US route, reads columns as before. Engine values for stored names go under engine_period__, so filters, conditions and bands keep their stored reads, and every injection site (national_calibration, full_measure per block and whole frame, the scoring path) uses the shared inject and drop. stored_columns is snapshotted before injection in each of them.
• Uprating on the engine route. I checked this because several routed inputs (employment_income, employee_pension_contributions, loss_relief) have no variable-level uprating. The resolver's simulation loads the H5 and pe-uk's _extend_dataset_with_cgt_provenance calls extend_single_year_dataset, which uprates through uprating_indices.yaml (employment_income is listed there). So 2025 reads are projected values, not the 2024 values relabelled.
• Bands. The SPI band rows band on total_income_after_loss_relief, which the provider computes, so band and value are both read at 2025.
• The refusal. MeasurementPeriodError is a RuntimeError, so it escapes the per-spec except (KeyError, TypeError, ValueError) and refuses the run rather than skipping the row. All nine declared rows pass it.
• The loss-relief guard. The frame read now applies only at the frame's period.
• State Pension at 52 weeks. pe-core's WEEKS_IN_YEAR is 52, and new_state_pension, basic_state_pension and additional_state_pension all divide state_pension_reported by it. So 365.25/7 did put full-rate pensioners 0.34% over the rate.
• ALLPAY2. It is now in miscellaneous_income's field list and gone from private_transfer_income, so it is counted once, not zero times. The new test refusing a field that feeds two income columns is a good guard.
• Council tax. The CTPAID 1 zero-bill rule is narrow (paying, no reduction, zero bill) and reuses the existing cell-mean donors.
• Carer's Allowance mechanics.
  • The draw is deterministic and keyed on person_id.
  • The stage refuses an incoming flag that does not follow reported receipt.
  • Reporters stay true.
  • It refuses any period other than the survey year.
  • The receipt is aggregate-only, with "<10" handling on both sides of the share.
• 61,000. Verified in the Independent Review, chapter 7: "raising the earnings threshold from £151 to £196 per week, bringing an estimated 61,000 more people into eligibility".
• Disclosure. Dropping the SPI band min and max, and returning "<10" for a mean drawn from fewer than ten distinct records, closes a real disclosure route.
• Graph kernels. Reviewing the local binding on the build's review date, not the wall clock, makes the kernels reproducible. Adding spi_income to the SPI stage's helper modules fixes a stale-hash gap.
• ONS ages 0, 1 and 2. These follow the 15 to 19 precedent and stay beside the 0 to 4 sex bands without double-binding a sum.
• Childcare. This PR does not touch Tax-Free Childcare or the funded-hours take-up draws. The is_child_receiving_* names in the diff are only in a new test's list of population variables that keep their stored reads. The new age 0, 1 and 2 rows will move the weight on under-3s, so any count of under-3s will move a little in the 22 October build.

🤖 Generated with Claude Code

[juaristi22]

@vahid-ahmadi thank you, this is a careful review. The changes are on top of 7e5c8b37c in 5 commits, and CI is re-running.

1. A declared frame period on an engine value. Fixed in 9d54588.

• _refuse_ignored_measurement_period now asks the adapter which period each value variable holds without the engine (UKFrameTargetAdapter.value_period):
  • a stored column holds the frame's period;
  • a CGT measure dated by its name (cgt_2024_tax) holds the year in its name. That rule moves to cgt_calibration.UK_CGT_DATED_MEASURE, so the measure provider and the adapter share one copy;
  • any other variable holds none.
• When a binding declares the frame's period and the run calibrates at another period, the guard now refuses unless every value_variable or value_expression term holds that period.
• New tests cover an engine value, an engine term in an expression, the resolver path, a dated CGT measure and a misdated one, and a run calibrated at the frame's own period.
• All 24 bindings that declare 2024 pass. The plain-column rows read stored columns (petrol, diesel, bus and rail) or cgt_2024_tax, and the rest use provider kinds.

2. The eligible pool and the 0.6. Agreed on both points. The pool counts care for anyone, and it holds every carer over the 2024-25 limit, not only those who come under £196.

• The stage receipt cannot carry the count you ask for. The stage reads the earnings test on the frame's own values at the survey year, and it refuses any other period, so it cannot see entitlement at 2025.
• So the back-test runs on the built national file instead. It reports two weighted counts: the drawn carers entitled at 2025, and the whole pool under the 2025-26 limit. Both are compared with the Review's 61,000, with banded support, and no figure where either side rests on one to nine.
• It goes on UK Carer's Allowance: would_claim follows reported receipt, so earnings-limit reforms reach 4,600 new carers against an official 61,000 #1319. No costing of a Carer's Allowance earnings-limit change on this build should be published before it.
• And yes: realized_share is the drawn share of the over-limit pool, not a take-up rate.

3. Salary sacrifice. Agreed: such a carer is paid in the baseline, so this is not only about how far reforms reach. The same back-test counts them, as drawn carers whom the released file holds under the survey year's limit. The move after salary_sacrifice stays on #1319's list, with #1332 in mind.

4. Carer Support Payment. Reworded in 3c3b473. Scotland now stays closed "since Carer Support Payment reads the same flag and policyengine-uk does not yet apply its earnings limit". The new wording is in the module, the stage notes and eligibility, and the changelog. The coverage manifest and the H2 fixture are regenerated in 2279fe8. The engine gap is PolicyEngine/policyengine-uk#2548.

5. Bus fares. The body now links PolicyEngine/policyengine-uk#2542, the draft fix PolicyEngine/policyengine-uk#2546 and #1331. The #1324 changelog fragment now says that a costing of a bus fare cap change on this release likely overstates its cost and gains (3d765aa).

6. Merge order. Agreed: this PR tonight, then #1333 with the measurement-period receipt, the Carer's Allowance receipt and the 52-week State Pension check re-run on its lock, then #1332.

Nits

• A. Deferred to UK Carer's Allowance: would_claim follows reported receipt, so earnings-limit reforms reach 4,600 new carers against an official 61,000 #1319's receipt follow-up, together with the back-test fields, so the receipt changes once.
• B. Noted. When the caseload target binds after the Budget, it will come from Stat-Xplore, not that page.
• C. Fixed in d47b48c. The docstring and the changelog fragment now quote section 1.3 of the FRS 2024-25 methodology: the band and "some other Council Tax block questions" left the field interview from April 2024, and the other answers "are then imputed during DWP's validation processes". They also say that the methodology does not state whether CTPAID is asked or imputed.

🤖 Generated with Claude Code

[vahid-ahmadi]

@juaristi22 thanks, this is my second review, at 2279fe8211ee48d26672d0783a581f400e116e21. It covers the five commits on top of 7e5c8b37c and your reply.

Verdict. Nothing in the delta blocks the 22 October build. The new period guard closes finding 1 properly, and the Carer Support Payment, bus fare and council-tax wording are fixed. Findings 2 and 3 and nits A and B are deferred to #1319, and I accept your reason why the stage receipt cannot carry the count. One thing I got wrong in round 1: the merge order. #1333 is stacked on #1332, so the order has to be #1329, then #1332, then #1333 (see item 6). I have one new ask and one nit.

What I ran

• Fresh worktree at the head, uv sync --all-packages --locked --extra uk (policyengine-uk 2.134.4, policyengine-core 3.33.3).
• The engine-free tests of the touched areas, 18 files (measurement period and its receipt and routes, Carer's Allowance take-up, council tax, the CGT calibration, observation period and targets, measure simulation, coverage, source stages, ledger and population targets, the FRS spine, national calibration, full measure, and the shared target materialization): 536 passed, 1 skipped.
• The engine tests the delta touches: Carer's Allowance take-up, CGT observation period and the release input coverage manifest: 9 passed.
• Checked that the new guard tests bite: with target_materialization.py put back to its 7e5c8b37c version, four of the new tests fail (both engine-value cases, the resolver case and the misdated CGT case). I restored the file afterwards.
• tools/ci_test_plan.py verify: ok.
• CI at the head: all 10 checks green.
• Listed every binding that declares measurement_period in uk_population_targets.json: 24, all 2024. Five are plain-column rows (road fuel, the bus and rail rows, cgt_2024_tax); the other 19 use parameter_gated_threshold. That matches your count.
• Re-read the FRS 2024-25 background information and methodology: the council-tax wording you quote is in section 1.3, "Recent improvements to the FRS".
• Not run: any national or spine build, or the back-tests. They need the licensed FRS/SPI inputs.

Round-1 items

1. Fixed (9d54588). The guard now refuses the frame's period on any value term the frame does not hold at that period, and value_period is fail-closed: anything that is not a stored column or a name-dated CGT measure returns None. I checked the two things that could break it. frame_period is always a string on the UK adapter (ledger_targets.py:1629, from the national frame's string time_period), so the string comparison does not refuse stored rows by type. And UKFrameTargetAdapter is the only adapter with a frame_period, so no other route skips the new check. Sharing UK_CGT_DATED_MEASURE between the provider and the adapter is the right way to keep them in step.

2. Deferred, agreed. Your reason holds: the stage reads the earnings test at the survey year and refuses any other period, so the 2025 entitled count has to come from the built file. Running both counts against the 61,000 there is better than what I asked for.

3. Deferred, agreed. Counting the salary-sacrifice cases in the same back-test is fine.

4. Fixed (3c3b473, 2279fe8). The wording is right in the module, the stage notes, the eligibility text and the changelog, and policyengine-uk#2548 is open for the engine gap.

5. Fixed (3d765aa). The body links policyengine-uk#2542 and #2546 and UK bus fares: keep in step with the engine's per-year fares, take the single-ticket share from cap-regime years, and check cap savings against NTS fares paid #1331, and the changelog carries the caveat. One update: #2546 was approved today but is still a draft and not released, so it is not in Bump the UK lock to the newest policyengine-uk (2.141.1, then the #2523 fix release) #1333's 2.141.1 lock. The caveat stays true after the bump, until a later pe-uk bump picks #2546 up. The body still calls it "the draft fix"; "approved, not yet released" would be more accurate.

6. Withdrawn, my order was wrong. Bump the UK lock to the newest policyengine-uk (2.141.1, then the #2523 fix release) #1333's base is uk-keyed-draws-pension-coherence, which is Draw UK QRF stages from uniforms keyed on each record's id, and tie pension contributions to pay #1332's branch. So Bump the UK lock to the newest policyengine-uk (2.141.1, then the #2523 fix release) #1333 cannot reach main before Draw UK QRF stages from uniforms keyed on each record's id, and tie pension contributions to pay #1332 without retargeting, and retargeting would carry Draw UK QRF stages from uniforms keyed on each record's id, and tie pension contributions to pay #1332's commits with it. My Draw UK QRF stages from uniforms keyed on each record's id, and tie pension contributions to pay #1332 review already said Draw UK QRF stages from uniforms keyed on each record's id, and tie pension contributions to pay #1332 then Bump the UK lock to the newest policyengine-uk (2.141.1, then the #2523 fix release) #1333. The one order is:

   1. UK pre-Budget data: measure stored values at their period, draw Carer's Allowance claims, and the #1278 items (#1324, #1319, #1278) #1329 tonight.
   2. Draw UK QRF stages from uniforms keyed on each record's id, and tie pension contributions to pay #1332, rebased on main after UK pre-Budget data: measure stored values at their period, draw Carer's Allowance claims, and the #1278 items (#1324, #1319, #1278) #1329. It shares frs_spine.py, spi_income.py, calibration_measure_exclusions.json and the parity fixtures with this PR, so it regenerates the H2 fixture and fixes the register count.
   3. Bump the UK lock to the newest policyengine-uk (2.141.1, then the #2523 fix release) #1333, rebased on that, with the measurement-period receipt, the Carer's Allowance receipt and the 52-week State Pension check re-run on its lock.

   Your reply agreed to my round-1 order, so please treat this as the correction.

• A. Deferred, agreed. One receipt change with the back-test fields is better than two.
• B. Noted.
• C. Fixed (d47b48c). The quoted wording and the section number match the methodology, and saying it does not state whether CTPAID is asked is the honest answer.

New

7. should UK Carer's Allowance: would_claim follows reported receipt, so earnings-limit reforms reach 4,600 new carers against an official 61,000 #1319 does not yet carry what this thread settled. Its body still has only the general back-test item, and it has no comments. Please add one comment there listing: the two weighted counts on the built national file (drawn carers entitled at 2025, and the whole pool under the 2025-26 limit) against the 61,000; the salary-sacrifice count; the realized_share rename; the stage move after salary_sacrifice; and "no costing of a Carer's Allowance earnings-limit change on this build before the back-test". Otherwise they live only in this PR's thread, which closes when it merges tonight.

8. nit target_materialization.py:1002-1012. The guard checks the value terms only. A filter, condition or band on an engine-computed variable would still be read at the calibration period under a declared frame period. None bites today: the five plain-column rows filter only on country or region, which are stored. A one-line note in the docstring would stop the next person assuming the guard covers filters too.

What looks right

• The guard splits a value_expression on + exactly as _expression does, and takes the entity the same way as _prepared_column_values (from_entity, then the spec's entity), so it checks the columns the read actually uses.
• The early return when the declared period equals the calibration period is right: the engine then computes every value at that period. The new test covers it.
• The misdated CGT case (cgt_2025_tax declared at 2024) refuses, and the resolver path refuses too, so both routes are covered.
• The regenerated coverage manifest and H2 fixture change only the Carer Support Payment wording and their hashes.

🤖 Generated with Claude Code