[juaristi22]

The savings follow-up to #1418, part 1: fix the low savings base at its source instead of making the solve stretch weights. This branch starts from PR C's last commit (0b276e571, merged as #1420).

Draft. The national built from 9f70c71e7 is certified (cut fd0d4bc2): all 50 spine gates, 7 calibration-seam gates and 19 release-cut gates pass. Its release directory assembled once 4dd83e999 moved the contract's pins, and the publisher pre-flight passes. It is staged, not published. The reply to the independent review has its measurements, and an earlier comment has a8e25870c's, the cut (0a4c8e2c) PolicyEngine/policyengine.py#731 still bundles. 91775db4a to 4dd83e999 address the review (below).

Earlier builds:

• The spine from d563f4fbc stopped at the UC take-up gate.
• The spine from 6485af4af passed every gate, but its national stopped at the fit gate on the tax on savings rows.
• The spine from f9676fed9 stopped in the new receipts stage, which hadn't declared its source-household read; 0c2f5314f declares it.
• The national from 0c2f5314f stopped at the release-cut certification on three gates; 3702769f0 to a8e25870c remove their causes (below).
• The national from a8e25870c certified. The review's fixes move the data, so the build from 9f70c71e7 replaces it.
• The spine from 59ef8ecee passed every gate, but its savings rebase hadn't converged (below). 9f70c71e7 bands the tape once, and the national from 59ef8ecee was stopped.
• The national from 9f70c71e7 certified. Its release directory failed the data contract on three pins that 91775db4a hadn't moved; 4dd83e999 moves them, and assembly and the pre-flight pass.

The problem

PR C binds SPI Table 3.7's interest and other income by total-income band. Interest moves to the OBR's savings path from 2023-24.

• The frame's base is low. At design weights it carries about 70 percent of those rows. The solve meets them by stretching the weights of the few households that carry savings, so the totals hit but the column's distribution doesn't.
• The SPI donors are under-aged. They come from the 2022-23 tape and reach the 2024 frame by each column's engine index:
  • other income by GDP per head, ×1.09 from 2022 to 2024, where SPI's own other income grew ×1.65 in the single year to 2023-24;
  • interest by ONS household interest, ×2.27, where SPI's grew ×2.69 in that year.
• The top band's split is flipped. On the tape, the £1m-and-over band's other income outweighs its interest, as in the published table. Ageing the two columns on such different indices flips that split.

The PR C national built from 5aba83cd2 stopped at certification on that band. Its figures stay out of this thread, because the band rests on the SPI tape.

The change

• spi_income.rebase_spi_donor_savings runs before any draw. It scales each donor's interest and other income to HMRC's published 2023-24 Table 3.7 amounts:

  • the factor in each band is the published amount over the FACT-weighted amount of the tape's taxpayers in the band, as Table 3.7 counts taxpayers only;
  • the tape is banded once on its income moved to 2023-24: each income leaf at its own factor, and each savings column at its factor over all bands. The band donors and the Table 3.7 targets band on uprated income, so the migration across band edges is counted once (since 9f70c71e7). A record whose band's factor differs from all bands' can still sit across an edge on its rebased income;
  • a taxpayer is a record the tape flags (TAXPAYER 1) or one whose income crosses the first band by 2023-24. A flagged taxpayer below the first band sits in it (Table 3.7 note 8). Any other record takes the factor of all bands together;
  • a band whose taxpayers hold under a tenth of its published amount takes the factor of all bands, and a band the tape carries none of is left as it is. Both are named in the receipt;
  • a tape whose TAXPAYER codes aren't 1 or 2 is refused.

  The stage-1 forest and the band-donor resample both read the rebased tape. The receipt records factors only, never counts.

• _spi_income_uprating_factors then grows the two columns from 2023-24 to the frame's year on the OBR's savings-income growth (October 2026 EFO, Table 4.5), through the new hmrc_uprating.uk_obr_savings_income_level. This is the path the Table 3.7 band targets already take. A column the rebase leaves as it is grows on its engine index from 2022-23 instead (since 91775db4a). From the frame's year on, the engine projects both at GDP per head, so both columns are now listed in SPI_HISTORY_INDEX_DIFFERS_FROM_PROJECTION.

• uk/hmrc_spi_table_3_7_savings_rows.json vendors the 26 published rows (13 bands × 2 columns) from the feed pin. It is declared in the vendoring register and the country package.

• Docs and changelog: docs/uk-income-anchors-280.md and changelog.d/1418-spi-donor-savings-rebase.changed.md.

The UC take-up gate: 0.055 to 0.06 (second commit)

The first spine built from 667fe532a stopped at uk_stage_uc_take_up. After the take-up redraw, single people's caseload ended 0.0568 of DWP's count further from it than before. PR C's spine was at 0.0523, and the tolerance is 0.055. The cause:

• The rebased SPI savings feed the UC reporter redraw's investment-income predictor, so a few SPI benefit units change report.
• That moves the take-up stage's pool merging, which joins adjacent award bands until each pool rests on enough sources.
• As a result, single people's £1,000–£1,100 band moved from a reachable pool into one whose DWP target exceeds the entitled units.

The reachable pools still realize DWP's counts within 0.005. The tolerance goes to 0.06 (María's ruling of 2026-10-30), as PR A took it from 0.05 to 0.055 for the gilt reshuffle. Pool boundaries that jump with small upstream changes are a follow-up on #1334. The gate's note, the committed-parameter test and the data contract's mirrored digests are updated.

Superseded by 6485af4af, which fixes the pools and returns the gate to 0.055 (below).

The national from f3cb3ddd3: what moved

The spine passed every gate. The national's fit gate blocked on one row. Below £1m, every Table 3.7 interest and other-income band fits within 1% after the solve, and the £1m-and-over other-income row that blocked PR C's national now passes the fit gate.

The share of each group's solved mass sitting on households stretched beyond 3× their design weight, PR C's blocked national against this one:

Group	PR C	This national
Interest carriers below £100k	54%	41%
Other-income carriers from £100k	29%	10%
Other-income carriers below £100k	66%	63%

The last row is the FRS under-reporting part 2 would address.

At design weights, the rebased frame carries more savings than SPI in the £200k–£1m bands, where its top earners are already over-represented, so the solve now cuts those savers' weights instead. The figures for the £1m-and-over band stay out of this thread, because it rests on the SPI tape.

The £1m-and-over interest hold (third commit)

The one row the fit gate blocked on, Table 3.7's £1m-and-over interest amount, is held as a measured distortion, as the ten HMRC £1m-and-over cells are (María's ruling of 2026-10-30), and is still measured on every evaluation. How the rebase sits in the top bands at design weights is a follow-up.

Why other income is short below £100k, and the fixes (fourth to sixth commits)

A read-only investigation of the national from f3cb3ddd3 found that below £100k the frame carries 0.40 of Table 3.7's other income at design weights. It ranked three causes:

1. FRS rows carry a different concept (46% of the gap).
2. SPI copies draw thin, incoherent other income (34%).
3. Taxpayer coverage (18%), which the solve's taxpayer rows already fix.

By María's choice (2026-10-30), three changes address the first two, plus a fix to part 1's rebase:

• 4a6dc728f, the rebase on taxpayers (F). Table 3.7 counts taxpayers only, so each band's factor is now computed over the tape's taxpayers (TAXPAYER 1), and non-taxpayers take the all-bands factor. Over all records, the two lowest bands' taxpayers fell short of the published amounts. Since 91775db4a, a saver who becomes a taxpayer by 2023-24 counts once, as a taxpayer, and a flagged taxpayer below the first band sits in it (the review's finding 10).
• 092de13b9, FRS savings income on SPI's definitions (A1, A2).
  • NS&I Direct Saver and the other NS&I and bank products move to savings_interest_income, since SPI has counted gross-paid interest as bank interest since 2016-17 (note 4). Gilts stay in other income.
  • NS&I interest is taxable savings income in either column, so its tax doesn't move.
  • Income from an annuity not bought with pension funds or from a trust or covenant (FRS PENTYPE 4 and 5) moved from private pension to other income (note 6). The review found that this took it out of policyengine-uk's means tests and household income, and changed its tax (finding 1). Since 91775db4a it stays in private_pension_income, and Table 3.7's other-income rows read it through a calibration-only descriptor (below).
• b82247443, a coherent other-income draw on SPI copies (C). The stage-1 transport locates each copy's draw at its twin's own other income (gilts, trust and annuity income) rather than at an identity uniform that ignored it. The stage spec's analogue map and the coverage manifest follow.

The investigation's frame arithmetic puts other income below £100k at design weights at about 0.66 of its target with these changes, against 0.40 before. The next build measures it.

FRS rows keep their own reported incomes: the investigation's larger lever, an SPI fill of other income on FRS-row taxpayers, is not taken.

The small traders' pay: SPI Table 3.10's employment-income counts (d563f4fbc)

The national from b441bb5a0 costs the Budget's trading allowance measures (Table 6.1 line 15) below PR A's, and both sit below the costing. A read-only investigation found two things:

• The fall from PR A is PR C's Table 3.10 counts correcting an overshoot in PR A's £2,000–£3,000 band, so it isn't a regression.
• The frame's small traders have the wrong mix. In Table 3.10, more than half of the traders below £3,000 also have employment income; the frame's mix leans to traders without pay. An employee's side profit is taxed at their marginal rate, so the mix decides what the measures cost.

By María's choice (2026-10-30), hmrc.spi.employment_income.count_by_self_employment_income_band binds Table 3.10's individuals with employment income over the same eight bands from £1 to £10,000 as the self-employment counts:

• Each person with pay over the year (employment_income_annual, the variable Table 3.6's employment rows read) counts in their own self-employment band, taxpayer or not.
• Table 3.10's column is wider: income assessed under ITEPA 2003 plus other miscellaneous income (its note 4). Like Table 3.6's employment rows, the binding leaves out miscellaneous income, benefits in kind and the contributory benefits.
• The counts grow to 2025 as the self-employment counts do. The zero band and the bands from £10,000 are signed out row by row in target_reference_signed_exclusions.json.
• The target is declared signature-incomplete beside the other Table 3.10 rows. The national references and membership and the national compile-parity receipt are regenerated on the pinned feed (53815b2), with eight new rows.
• The membership's 13 SPI dividend rows and 8 Table 3.10 self-employment amount rows also take the October EFO engine pins that main's policyengine-uk 2.164.2 lock brought (b89d3aab8). The merge had left them on the 2.162.14 pins. The parity receipt doesn't move for them, and builds compile live from the pins.

The tax on savings rows: the OBR path (5c0137fb4, reverted in f9676fed9)

The three ITL Table 2.6 rows bound HMRC's own 2024-25 and 2025-26 projections as a calendar-2025 window: £8.40bn of tax on savings, ×1.52 on the 2023-24 outturn. Table 3.7's savings income grows only ×1.36 on the OBR's path. Bound together, the rows pinned the tax on savings, and so the savings rates' static yield, about 12% above the base the Budget's costing grows.

By María's choice (2026-10-31, option A):

• The rows read the 2023-24 outturn alone and grow it by the OBR's savings income growth, the index Table 3.7's savings rows take: £1.61bn, £2.24bn and £3.66bn, £7.51bn in all.
• This leaves out the leverage that HMRC's projection carries from the frozen Personal Savings Allowance and thresholds.
• hmrc_itl_savings_rate leaves the source-projection families. Only the three rows move in the references, the membership and the parity receipt.

Reverted (f9676fed9, María's ruling of 2026-10-31). The national from 6485af4af stopped at uk_target_fit, as the leverage caveat predicted:

• Under the engine's rules, tax on savings grows faster than savings income, because the allowance and thresholds are frozen.
• So the tax rows grown on the income's path asked for less tax than Table 3.7's savings income gives.
• The solve cut the interest rows in the basic-rate bands to meet them: the £30,000 band ended 30.7% under target and the £20,000 band 12.0% under.

HMRC's own projection carries that leverage, and b441bb5a0's national met it beside the interest rows. So the savings rates' yield stays pinned to the projection. Its distance from the costing is a difference of basis and of behaviour (ISA shifting), not a fault in the base.

UC take-up pools that cannot meet DWP's count (6485af4af)

The spine from d563f4fbc stopped at uk_stage_uc_take_up. Single people's caseload ended 0.0674 of DWP's count further from it than before the redraw (0.0568 on f3cb3ddd3's spine).

• The pools whose counts the draw can reach met them.
• The gap is the pools that cannot meet their count: either DWP's count exceeds every entitled unit, or the reporters alone already exceed it. Those pools kept a clipped rate, and which bands fall there moves with small upstream changes.

By María's ruling (2026-10-31), the stage is fixed instead of the gate widened again:

• The join (replaced). After the source-support pooling, a pool that cannot meet its count joined the adjacent pool with more room for its gap. The review found that a joined pool's single rate moved claims the wrong way between award bands (finding 7).

• Passing gaps (91775db4a). A pool that cannot meet its count is held at its clip, and its gap passes on:

  • a shortfall and an excess elsewhere in the family type offset first;
  • what remains goes to the other pools, nearest first, the one with more room first at equal distance and the lower band on a tie, each taking what its room allows.

  Every band keeps its own rate, so no band's drawn units claim against its own count to meet another's. A family type's caseload still lands on DWP's wherever its entitled units allow. UC results move in later builds.

• The gate returns to main's 0.055.

• What follows: the stage spec, the coverage manifest, the H2 fixture and the data contract's digests. New stage tests cover the gap passing.

The trading allowance's base: the self-employed's gross receipts (f3540cbc6)

policyengine-uk applies the trading allowance against self_employment_gross_receipts, and neither the FRS nor the SPI tape records turnover. Without receipts, the engine relieved profits up to the allowance in full and nothing above. So the Budget's raise to £3,000 missed its main base: traders who already claim the £1,000 allowance against larger receipts and gain the larger deduction. No published source gives turnover beside profit. But under current law a claimant's receipts are their profit plus £1,000 exactly, so only who claims is missing.

By María's choice (2026-10-31), a new last spine stage, trading_allowance_receipts, sets each trader's receipts:

• Who claims. Each trader with a positive profit is flagged as a claimant, drawn on their own person_id: 0.64 for profits up to £4,500, 0.32 from £4,500 to £9,000, none above.

• Receipts. A claimant's receipts are their profit plus £1,000. Any other trader's are their profit plus £3,000, expenses at the raised allowance; the costing books traders who switch to the allowance as behaviour. A person with no positive profit keeps zero, which the engine reads as unknown.

• The rates are calibrated to HMRC's policy paper (about 400,000 taxpayers with a lower bill, by about £380 on average) until HMRC publishes claimant counts. An HMRC table of claimants by profit band would replace them.

• Filing. Every trader with a positive profit is treated as filing, as the bound Table 3.10 counts are of filers. So a profit of £1,000 or less is now taxed where the engine used to relieve it in full.

• Two limits, until Add a trading allowance claim status, so allowance reforms reach claimants without turnover data policyengine-uk#2765's claim status is released (UK trading allowance: store the claim status once policyengine-uk#2765 is released #1439; María's ruling of 2026-11-01 on the review's findings 2 and 8). The engine uprates receipts with profit, on the mixed income index, while the allowance is nominal:

  • a claimant's £1,000 grows, so the raise relieves £1,932 more in 2027-28, falling to £1,758 by 2031-32;
  • the other traders' stored £3,000 grows too, so any allowance above £3,000 relieves every other trader from the stored expense. Allowance reforms above £3,000 shouldn't be scored on this national.

  Who gains is approximate too: a drawn non-claimant whose receipts are within the allowance loses the full relief the law gives automatically, and the calibrated rates move that gain onto claimants.

• Wiring:

  • the operation kind in the stage schema;
  • the stage spec, with its mass receipt;
  • the graph, kernel and spine-build wiring;
  • the coverage manifest family and the H2 fixture (51 stages).
  • self_employment_gross_receipts also leaves the register of engine inputs left at their default and joins the export surface's allowed columns, so the data contract's gate digests follow.

Certifying the national from 0c2f5314f (3702769f0 to a8e25870c)

The release-cut certification of the national from 0c2f5314f stopped on three gates. By María's rulings (2026-10-31), all three are fixed at their cause, and no gate is widened.

• uk_weights_audit: a bug on main since Budget 2026 ISA data, PR B of #1418: ISA subscriptions per adult, the Lifetime ISA anchor, additive ISA income, policyengine-uk 2.164.2 #1421 (b9bcd709c).
  • PR B's ISA subscriptions stage fits no model, but it declared an empty fit_weight_records tuple.
  • The stage evidence then carried an empty list for it. The audit reads an empty list as a fitting stage that recorded nothing, so every national certified after Budget 2026 ISA data, PR B of #1418: ISA subscriptions per adult, the Lifetime ISA anchor, additive ISA income, policyengine-uk 2.164.2 #1421 would fail.
  • The stage no longer declares the hook, and a regression test reads its evidence.
  • 3702769f0 also pins the receipts stage in the two stage-order tests that named the ISA subscriptions as the last stage.
• uk_input_mass_parity on charitable_investment_gifts (c5ef12756).
  • The cause. Experiment spines with the other-income analogue (C) undone, then the taxpayer rebase (F) as well, put the cause in F. Gift Aid and charitable investment gifts are sparse leaves that the stage-1 chain draws after, and conditioned on, the income leaves. Rebasing the donor savings on the tape's taxpayers changed which few donors' gifts the copies drew.
  • The fix. F stays. Each gift column's total over the drawn copies, at the stage's household weights, is scaled to the prepared tape's FACT-weighted total. A column the tape or the copies carry none of is left as drawn. The stage evidence records the factors only.
  • The gift figures stay out of this thread, because the columns rest on few SPI records.
• uk_input_mass_parity on jsa_income_reported (a8e25870c).
  • The cause. The national met OBR's total JSA by scaling income-based reporters together with contributory claimants. But income-based JSA is closing with the move to Universal Credit. DWP's Spring Forecast 2026 puts it at £91.0m in 2024-25 and under £1m from 2025-26, against £222.4m and £279.5m of contributory JSA.
  • The fix. dwp.jsa.amount_income_based binds DWP's income-based line at the calendar-2025 window of the two years, about £23.4m, as the pension-age Housing Benefit row is. The solve now meets the rest of OBR's total on contributory claimants.
  • The national references and membership and the national compile-parity receipt are regenerated on the pinned feed (53815b2), with one new row.

The review of a8e25870c (91775db4a to 4dd83e999)

An independent review of a8e25870c raised thirteen findings. By María's rulings (2026-11-01):

• Annuity and trust income (finding 1, 91775db4a).
  • FRS PENTYPE 4 and 5 stay in private_pension_income, where policyengine-uk's means tests (Pension Credit, Universal Credit, Housing Benefit, Council Tax Reduction, Income Support), household income and tax read them.
  • A new descriptor, annuity_and_trust_income, stores their part. It is calibration-only, and the export carries it.
  • Table 3.7's other-income rows read a new measure, other_investment_income_spi_concept: the engine's other income plus that share of the engine's private pension, as the table counts both payments as other income (note 6).
  • A row the SPI stage draws holds the tape's pension, which leaves these payments out, so its descriptor is zero.
  • The FRS spine refuses a pension tab without PENTYPE, and the H2 fixture's pension rows now carry it.
• The receipts encoding (findings 2 and 8). Kept for the Budget's £3,000, with its limits stated in the stage spec, the module and the changelog (above). UK trading allowance: store the claim status once policyengine-uk#2765 is released #1439 tracks storing the claim status once Add a trading allowance claim status, so allowance reforms reach claimants without turnover data policyengine-uk#2765 is released.
• Empty fit-weight records (finding 3, 24e2329f6). A roster test pins the eight stage transforms that declare fit_weight_records, read from the source. A new one has to be reviewed, and either record each fit it makes or declare nothing.
• The receipts test's labels (finding 4, 24e2329f6). They are built with rounded_count, so the test holds whatever small-cell floor main sets.
• The UC take-up gate with UK #1095 A/B fixes: the SPI age band (uk-data#628), households' HRAD from Table 8a receipts, and the SPI pay floor (uk-data#629) #1426 (finding 5). Not changed here. A merge with UK #1095 A/B fixes: the SPI age band (uk-data#628), households' HRAD from Table 8a receipts, and the SPI pay floor (uk-data#629) #1426 takes its 0.12 without a conflict, under this PR's note of 0.055, so the value needs a ruling at that merge.
• The rebase (findings 6 and 10, 91775db4a and 9f70c71e7). Banding on income moved to 2023-24 and the taxpayer basis, as above.
  • 91775db4a moved each record's savings by its own band's factor and repeated the banding until no record changed band. On the production tape that didn't converge, because a record's band sets its factor and the factor sets its band.
  • 9f70c71e7 bands once, moving the savings by each column's factor over all bands. Its receipt records that factor in place of the pass count.
• UC take-up (finding 7, 91775db4a). Gaps pass between pools instead of joining them, as above.
• The gift anchor (finding 9). One factor per column fixes each total, not how concentrated the column is; the changelog says so. A published total for the charity reliefs would bind it (Not in this PR).
• The support floor (finding 11, 24e2329f6). Its counts scope takes Table 3.10's employment count by self-employment band, so a cell without support is dropped and receipted instead of stopping the solve.
• Target tests (finding 12, 24e2329f6). Tests pin the income-based JSA row and the employment count's binding. The claimant rates' calibration, as a committed tool with a post-build check, is part of UK trading allowance: store the claim status once policyengine-uk#2765 is released #1439.
• Texts (finding 13, 24e2329f6).
  • The £1m-and-over interest hold now tracks UK: data fixes for the Budget 2026 household measures #1418.
  • The income anchors doc, the uprating test and the NS&I and gilt comments describe what ships.
  • The employment count's note says SPI rows count benefits in kind.
  • The changelog fragments describe the change, not the branch's history.
• 4dd83e999 moves the data contract's full-manifest pins (policy, gates manifest and spec fingerprint) with the descriptor's export surface, as f3540cbc6 did for the receipts column.
• 59ef8ecee regenerates the H2 fixture and the coverage manifest. The national references, membership and parity receipt don't move. The H2 fixture's synthetic tape doesn't gross to the published amounts, so 9f70c71e7 leaves it as it is.

Up to date with main, and CI (30 Oct)

• 9ee6fbde0 merges main into the branch. The main change it brings is PR B (Budget 2026 ISA data, PR B of #1418: ISA subscriptions per adult, the Lifetime ISA anchor, additive ISA income, policyengine-uk 2.164.2 #1421: ISA subscriptions, policyengine-uk 2.164.2). Conflicts resolved:
  • the FRS account routing, keeping main's ISA interest column and this branch's NS&I routing;
  • the FRS spine and SPI engine tests;
  • the vendoring register, keeping both new resources;
  • the data contract's gate digests, recomputed for the merged gates.json;
  • the coverage manifest, regenerated.
• 995969f02 leaves the savings rebase off any tape whose weights don't gross to the population.
  • The band factors assume they do. On the H2 fixture's synthetic tape they ran to millions and broke the band donors.
  • A column whose tape holds under a tenth of the published amount is now left as it is, and the receipt says so.
  • The production tape holds about 0.37 of the published interest and 0.6 of the other income.
• e5ac7a5d8 regenerates the H2 spine fixture for the other-income analogue (C). CI's engine-uk graph test and the integration-uk smoke test refused its stale stage-spec snapshot.
• 53ad29934 starts a missing self-analogue output at zero before the SPI transport. Since C, the transport reads other_investment_income as each twin's own analogue before the stage-1 draw, so CI's engine test of a frame whose FRS rows carry no other income was refused. That column now gets the stage's zero before the transport, as it did after the draw, and the final values are unchanged. Production FRS frames always write it.
• 45c2888fa regenerates the H2 spine fixture on policyengine-uk 2.164.2, the lock's pin since the merge of main. e5ac7a5d8 had built it in a local environment still on 2.162.14, which CI's engine-uk graph test caught.

Not in this PR

• Measured claimant counts for the trading allowance. The claimant rates are calibrated to HMRC's published gainers, not measured. Miscellaneous income's allowance is the engine's (policyengine-uk#2726).
• FRS under-reporting of savings below £100k (part 2) is not addressed here. The build shows how much of that gap remains.
• A published total for charitable giving. No target binds the gift columns, so the final weights move their carriers as the solve needs, and the anchor fixes their totals, not their concentration (the review's finding 9). HMRC's cost of the charity reliefs on gifts of shares and property would bind charitable investment gifts; the feed doesn't carry it yet.
• policyengine-uk's claimant flag. policyengine-uk#2764 proposes an input saying whether a trader claims the allowance in the input data (draft PR policyengine-uk#2765). Once it is released, the receipts stage would store the flag instead of receipts, which removes both limits above (UK trading allowance: store the claim status once policyengine-uk#2765 is released #1439).
• The UC take-up gate after UK #1095 A/B fixes: the SPI age band (uk-data#628), households' HRAD from Table 8a receipts, and the SPI pay floor (uk-data#629) #1426. Its value needs a ruling when the two PRs meet (the review's finding 5).

Checks run locally

All pass. Engine-free tests:

• the new test_uk_spi_savings_rebase.py;
• the SPI income, spine, dependency and acceptance-receipt tests;
• FRS spine, graph, property components, stage health and the HMRC source manifest;
• Carer's Allowance take-up, ledger-fact vendoring, country spec and donor uprating;
• the spine parity instrument and the WAS tests that read the H2 fixture;
• UC take-up and stage health, the population targets and target references, the source stages, the coverage manifest and the data contract;
• the new test_uk_trading_allowance_receipts.py, the stage rosters, the country spec, the graph and the terminal gates;
• for the certification fixes:
  • the ISA subscriptions and private-school stage tests, and stage evidence;
  • the gift anchor's test in test_uk_spi_savings_rebase.py, and the SPI spine, support and rank-transport tests;
  • the population targets, target references, cross-grain declarations and subset relations;
  • the full gates, the battery bindings, the coverage manifest, the compile-parity receipts and the data contract;
  • the pension-age, legacy-closure and benefit-scope targets.
• for the review's fixes:
  • the FRS spine, measure simulation, SPI savings rebase, SPI spine and UC take-up tests, with new cases for findings 1, 6, 7 and 10;
  • the new test_uk_fit_weight_records_roster.py, the receipts and support floor tests, and the pension-age targets;
  • the source stages, the HMRC source manifest, the terminal gates, the national frame, the graph and stage health;
  • the target references, the coverage manifest, the gates, the compile-parity receipts and the data contract.

The engine tests engine/uk/test_uk_spi_income.py and engine/uk/test_uk_spi_state_pension_age.py also pass locally for the gift anchor, and the first again for the review's fixes.

The updated engine tests (engine/uk/test_uk_spi_income.py, engine/uk/test_uk_hmrc_uprating.py, engine/uk/test_uk_itl_savings_tax_rates.py, engine/uk/test_uk_release_input_coverage.py) and the H2 acceptance test run in CI.

🤖 Generated with Claude Code

[juaristi22]

Measurements: the certified national from a8e25870c

Edited 1 Nov after Max's review on PolicyEngine/policyengine.py#731: the savings and property rates are compared with the OBR's income tax line, the trading allowance with income tax plus NICs, and Table 6.1 stays as context.

The national from a8e25870c is certified. All 19 release-cut gates pass. The post ran clean:

• the stage-1 comparison;
• the head-to-heads against 4374e0bf and a6299504;
• assembly and the publisher pre-flight.

It is staged, not published.

Budget measures

The table shows the change in the government balance, in £bn, from PR A's dashboard scoring on policyengine-uk 2.162.14. Years are policyengine-uk's (2028 is 2028-29). The OBR line covers the taxes the model covers, after behaviour (October 2026 EFO); the other measures aren't compared with their own lines yet (—).

Measure	Years	Part 1 (b441bb5a0)	This PR (a8e25870c)	OBR line	Table 6.1
Trading allowance to £3,000	2028 to 2031	−0.097 / −0.097 / −0.093 / −0.097	−0.148 / −0.149 / −0.148 / −0.147	−0.221 / −0.165 / −0.171 / −0.178	−0.225 / −0.165 / −0.175 / −0.180
Savings income rates +2pp	2029 to 2031	+0.610 / +0.646 / +0.678	+0.582 / +0.611 / +0.636	+0.443 / +0.475 / +0.510	+0.520 / +0.555 / +0.590
Property income rates +2pp	2029 to 2031	+0.422 / +0.438 / +0.451	+0.402 / +0.417 / +0.429	+0.394 / +0.403 / +0.412	+0.435 / +0.445 / +0.455
Carer's Allowance earnings taper	2030, 2031	−0.038 / −0.075	−0.037 / −0.073	—	−0.040 / −0.080
Domestic electricity at zero VAT	2026, 2027	−0.789 / −0.332	−0.788 / −0.332	—	−0.755 / −0.305
Fuel duty rise deferred	2026, 2027	−0.290 / −0.501	−0.290 / −0.501	—	−0.575 / −0.780

What moved

• Trading allowance. The receipts stage lifts the cost by about half, to about HMRC's static base: about 400,000 taxpayers with a lower bill, by about £380 on average. The rest of the costing is behaviour, traders who switch to the allowance, plus miscellaneous income, which this scoring environment doesn't relieve (policyengine-uk#2726).
• Trading allowance, against the OBR line. From 2029-30 the cost trails income tax plus NICs by 10–17%. 2028-29 is the first Self Assessment cash year, with about 1.5 years of liability.
• Savings rates. The yield is 25–31% above the OBR's income tax line (part 1: 33–38%). That compares a static figure with a line after behaviour.
  • This PR pins the 2025 tax on savings to HMRC's Table 2.6 projection, by rate (0.98, 1.00, 1.00). On that base, +2 points gives £0.48bn in 2025 and £0.58bn in 2029-30.
  • EFO paragraph 3.17 gives the measure about £0.6bn a year after a savers' response of around £0.2bn, mainly ISAs. That implies a static yield of roughly £0.75bn across all taxes, against this PR's static income tax of about £0.61bn.
  • The split of that response by tax isn't published. So the gap reads as at least half behaviour, probably most of it, with Max asked to verify.
  • The earlier reading here, a difference of base level with ISA shifting small, is withdrawn.
  • Behaviour belongs in policyengine-uk, not in this dataset, so no data change is proposed for savings.
• Property income rates. The yield is about 5% below part 1 and 2–4% above the OBR's income tax line.
  • Against part 1 the 2029-30 yield falls £20m:
    • £8m is extra finance-cost relief. Nothing binds landlords' finance costs, and this branch moved them toward HMRC's published level.
    • £10m is a lower tax rate on property income outside Scotland.
    • £2m is more property income in Scotland, where the rate doesn't rise.
  • Against the income tax line there is no shortfall to explain. The partnership rescale in Certify the Microcosm UK national from microcosm#1428 on policyengine-uk 2.164.2 (0a4c8e2c; local area 6528c865 or c9df4b86) policyengine.py#731 is withdrawn, and the PRIS bounds stay a lead (comment).
• The other measures barely move.
• Certification fixes. They move none of these measures by more than £0.001bn against the uncertified build from 0c2f5314f.
• The pension tax allowance measure fails in the scoring template on policyengine-uk 2.162.14, not on the data.

🤖 Generated with Claude Code

[juaristi22]

Independent review of #1428 at a8e25870c

This is an independent adversarial review. Five reviewers each took one area:

• the SPI savings rebase and the gift anchor;
• the FRS definitions, the UC pool join and the ISA fix;
• the trading allowance receipts stage;
• targets and the ledger;
• a cross-cutting pass, with trial merges against main, UK: take up #1095's actionable items on policyengine-uk 2.165.1 #1422, UK #1095 A/B fixes: the SPI age band (uk-data#628), households' HRAD from Table 8a receipts, and the SPI pay floor (uk-data#629) #1426 and UK: company cars from the FRS job table (dashboard #316) #1430.

Two verifiers then reproduced each finding, on synthetic data or on the pinned engine (policyengine-uk 2.164.2), and a last pass looked for gaps. Nothing here uses licensed data. Paths are relative to packages/microcosm-build/src/microcosm/build/ unless they start with packages/, tools/, docs/ or changelog.d/.

Verdict:

• Fix before the national is used: 1 and 2. The certified national from a8e25870c carries both.
• Merge hazards to plan for: 3 to 5.
• The rest: modelling, test and text issues.

Fix before the national is used

1. Annuity and trust income leaves the means tests and household income (uk_runtime/frs_spine.py:615-621, :725).

• What moved: 092de13b9 moves FRS PENTYPE 4 (a non-pension annuity) and 5 (a trust or covenant) from private_pension_income to other_investment_income.

• The engine's means tests: in policyengine-uk 2.164.2, private_pension_income is the only input that carries this income into:

  • Pension Credit income;
  • UC unearned income;
  • the Housing Benefit, Council Tax Reduction and Income Support income lists.

  None of them reads other_investment_income. At 2.164.2 it is also missing from hbai_household_net_income, market_income and capital_income.

• Microcosm's own sums skip it too:

  • the SPI no-record rule (uk_runtime/spi_income.py:2466-2471, :2506-2530);
  • the CGT taxable-income proxy (uk_runtime/cgt_imputation.py:198-207);
  • the ISA stage's income bands (uk_runtime/isa_subscriptions.py:776).

• Benefit examples (synthetic):

  • a pensioner with State Pension of £200 a week and a purchased annuity of £50 a week sees Pension Credit income fall from £250 to £200. That gives about £27 a week of Guarantee Credit they aren't entitled to.
  • a UC claimant with £100 a month of trust income gets £1,200 a year more.
  • In law both count: UC Regulations 2013 reg. 66(1)(i) and (j), and SPCA 2002 s. 15(1)(d).

• The tax moves too. The body says it doesn't, but that holds only for NS&I:

  • Why: these rows go from non-savings income to savings income. So they now get the Personal Savings Allowance and the starting rate for savings, Scottish taxpayers pay UK rates on them, and they enter the base of the +2pp savings measure.
  • Example: on the engine, State Pension of £11,500 plus £3,000 of this income pays tax of £386 as pension (Scotland £366.70) and £0 as other income.
  • The law: trust and covenant payments are generally non-savings income (ITA 2007 s. 18). Only the income element of a purchased annuity is savings income, and the capital element is exempt (ITTOIA s. 717). Yet the whole payment is booked.

• Suggestion: keep these payments in private_pension_income for the engine, and feed them to the Table 3.7 other-income measure through a calibration-only descriptor, as frs_annual_income's descriptors work.

2. Non-claimants' receipts give spurious relief at any allowance above £3,000 (uk_runtime/trading_allowance_receipts.py:84).

• Mechanism:
  • every trader not drawn as a claimant is stored with receipts of profit + £3,000;
  • the engine relieves min(profit, allowance − (receipts − profit)) and grows both columns by per-capita mixed income;
  • so any allowance above £3,000 × that index gives every non-claimant relief. That includes every trader above £9,000, and it comes from the stored expense, not from data.
• Engine runs for a non-claimant with 2024 profit of £50,000 or £2,000 (2027-28 / 2028-29):

Allowance	Deduction 2027-28	Deduction 2028-29
£3,000	£0	£0
£3,500	£296	£183
£4,000	£796	£683
£5,000	£1,796	£1,683

• Before this PR they got nothing at any of these levels. A £50,000 trader's 2028-29 income tax at £5,000 falls from £9,547 to £8,874. So any raise beyond the Budget's £3,000 is mis-scored; a rough estimate from public counts is hundreds of millions to about £1.5bn a year at £4,000 to £5,000.
• Later conflict: once pe-uk#2765's claim status lands, any non-claimant full relief it computes would also read the stored expense.

Merge hazards

3. The empty fit-weight-records failure is a class, and #1430 brings it back.

• The chain: b9bcd709c fixes the ISA stage, but any stage that declares the hook and returns () fails the same way:
  • it is snapshotted as empty (uk_runtime/stage_evidence.py:69-79);
  • it is rehydrated to () (uk_runtime/release_certification.py:216-219);
  • it then fails the release-blocking uk_weights_audit (uk_runtime/battery_bindings.py:1469-1471), with no exemption.
• UK: company cars from the FRS job table (dashboard #316) #1430: its company_cars stage (company_cars.py:1373-1375 at 40968ff53) returns (). A probe with its real class fails the audit, and the two PRs merge without conflict in either order.
• The same class elsewhere:
  • nts_bus_travel's published-shares fallback stores ();
  • uc_reporter_redraw, salary_sacrifice and spi_housing_shell make weighted fits with no hook, so the audit never sees them.
• Suggestion: a roster-wide test that each stage either fits and records or declares nothing.

4. With #1422, the new receipts test fails (packages/microcosm-build/tests/engine_free/uk/test_uk_trading_allowance_receipts.py:121-125).

• It asserts the literal "<10" for five traders. The file is new, so it merges without conflict, and then UK: take up #1095's actionable items on policyengine-uk 2.165.1 #1422's floor of three makes it fail ('3-9' == '<10').
• test_uk_uc_take_up.py:254 sits in a conflict hunk that has to be rewritten by hand.
• Fix: build the expected labels with rounded_count.

5. With #1426, the UC take-up gate silently becomes 0.12 (uk/gates.json:638).

• The merge: maximum_caseload_deviation merges to UK #1095 A/B fixes: the SPI age band (uk-data#628), households' HRAD from Table 8a receipts, and the SPI pay floor (uk-data#629) #1426's 0.12 outside the conflict markers. Only the notes conflict, and this PR's note says 0.055. So keeping this PR's note ships 0.12 under a text that says 0.055.
• What it needs: UK #1095 A/B fixes: the SPI age band (uk-data#628), households' HRAD from Table 8a receipts, and the SPI pay floor (uk-data#629) #1426 scopes 0.12 to the UC award fix of UK: every open data item: uk-data parity, engine locks, the local surface, consumption and high-value homes #1095, and the pool join (7) changes what the gate measures. The value needs a ruling and a re-measured spine.

Modelling

6. The rebase counts band-edge migration twice (uk_runtime/spi_income.py:1925-1926).

• Mechanism:
  • Each factor is the published 2023-24 amount, banded on 2023-24 income, over the tape banded on its 2022-23 nominal income. So it absorbs net migration across band edges into the open top bands.
  • Every consumer then bands again on uprated income, and counts that migration again: the band-donor pools (:1471) and the Table 3.7 targets.
• Synthetic tapes: the rebased frame's £1m-and-over interest comes to between 1.02 and 1.35 times published, depending on the tail. The no-growth control gives exactly 1. The £200k–£1m bands also over-carry, and narrow bottom bands under-carry.
• Suggestion: band the tape on income moved to 2023-24, or iterate the factors until the rebased tape, banded on its own uprated income, reproduces the published amounts.

7. The UC pool join moves claims the wrong way between award bands (uk_runtime/uc_take_up.py:547).

• Mechanism: a joined pool applies one realized rate to every drawable unit in both bands. So an overshooting band's drawn units start claiming, and a short band's stop.
• Real stage function, join off and on:

Case	Join off (claims per band)	Join on (claims per band)	UC spending
A band overshooting on reporters	200 / 200	280 / 70	−39%
A band short of its count	60 / 100	38 / 162	+45%
Reporters above the family total	150 / 80	150 / 0

• Why it matters: the gate checks family totals only, but 92 of the 100 family-by-band payment rows bind in the national solve.
• Suggestion: holding the unmet band at its clip and passing its gap to the neighbour gives the same family total, with less per-band error.

8. The receipts encoding also misplaces who gains (uk_runtime/trading_allowance_receipts.py:185-189).

• Full relief: drawn non-claimants whose real receipts are within the allowance lose full relief, which ITTOIA s. 783AE gives automatically. Only partial relief needs the s. 783AM election. With the rates calibrated to HMRC's 400,000 gainers, their gain moves onto drawn claimants.
• Uprating drift: a claimant's £1,000 grows with the index, so the raise's extra relief is £1,932 in 2027-28, falling to £1,758 in 2031-32. That is 3.4% to 12.1% below £2,000, not "slightly", the wording at :35-37, in the stage spec and in the body.
• Capital allowances: the profit + £1,000 identity also fails for SPI copies with capital allowances, which can't have claimed partial relief.

9. The gift anchor fixes the total but not the concentration (uk_runtime/spi_income.py:2851-2852).

• Mechanism: one scalar per column scales every SPI row. So if a few extreme draws dominate a column, they still dominate it, and every other carrier shrinks.
• Example (synthetic): with one large draw holding 98% of the mass, ordinary carriers fall from £1,000 to about £16.
• Coverage: band carriers at full Table 2.5 counts and support copies at half coverage are scaled alike.
• Why it matters: Gift Aid works per person, through rate-band extension and adjusted net income, and no target binds gifts. The stage evidence records factors only, so concentration in the real build can't be read from it.

10. The taxpayer basis double-counts, and the guard falls back silently (uk_runtime/spi_income.py:1933, :1954-1962, :2066-2069).

• Two double counts:
  • The factor divides 2023-24 published taxpayers by 2022-23 tape taxpayers, and also scales 2022-23 non-taxpayers by the all-bands factor. So savers who became taxpayers count twice: two donors, one a non-taxpayer, who are both taxpayers next year give 800 where the truth is 400.
  • Table 3.7's lowest band includes taxpayers below the Personal Allowance (note 8), and they are double-counted the same way.
• Latent fallbacks:
  • a column the grossing guard skips stays at 2022-23 values, yet is still grown from 2023-24 under the published-bands label;
  • the guard is per column, so a thin top band can get a factor near a million;
  • is_taxpayer is always written, so the receipt always says "taxpayers";
  • TAXPAYER is neither a required column nor validated.

11. The new employment count sits outside the support floor (uk/calibration_measure_exclusions.json:34-45).

• The gap: the floor's counts scope lists Table 3.10's self-employment rows, but not the new employment count. Yet its reason (:76) says Table 3.10's bands are dropped and receipted, not refused.
• Effect: a zero-support employment cell would abort the solve's initialisation, or fail uk_target_fit, instead of being dropped. It's latent.

Tests and texts

12. The new targets and claimant rates have no tests or evidence.

• Targets: in-memory mutations pass all fifteen engine-free target test files. These are changing the JSA row's value_variable, dropping the employment filter, and removing band_upper_bound. Compare test_uk_pension_age_benefit_targets.py:88-111 for the Housing Benefit sibling.
• Claimant rates: 0.64 and 0.32, and the £4,500 and £9,000 edges, are calibrated off-repo. No script or evidence is committed, and no gate re-measures gainers or the average gain.

13. Texts:

• the £1m-and-over interest hold tracks this PR (uk/calibration_measure_exclusions.json:177), so its expiry message will point to a merged PR;
• docs/uk-income-anchors-280.md:
  • :65-69 contradicts :87;
  • :88-90 leaves out the taxpayer basis and the guard;
  • :401-402 describes the old other-income analogue;
• tests/engine/uk/test_uk_hmrc_uprating.py:125-149 describes the removed donor path;
• the "NS&I and gilt" texts are now stale:
  • uk_runtime/spi_support.py:139-143, which is also the stated reason other income stays out of the stage-2 predictors;
  • uk_runtime/spi_spine.py:218-219;
  • uk/uk_population_targets.json:14875 and :22071;
• the employment count's note (:3901) says benefits in kind are excluded, but SPI rows count PAY + EPB + TAXTERM;
• the changelog fragments describe branch history, including a 0.06 tolerance that never shipped, and fixes to code that never shipped;
• uk_runtime/frs_spine.py:750 reads PENTYPE with a silent zero, where the file refuses other missing columns. The H2 fixture has no PENTYPE column, so it never exercises the new routing.

Checked and found true

• The option A revert is exact: git diff 5c0137fb4^ f9676fed9 is empty over all seven files option A touched.
• The vendored Table 3.7 rows match the pinned feed: 26 rows, and the sha matches.
• The eight employment-count rows check out: their uprating, signed exclusions, membership and parity rows. So does the JSA window: income-based plus contributory comes within 0.1% of OBR's total.
• Contract and ledger: the contract digests, the coverage manifest and the ledger accounting are consistent.
• The receipts stage's wiring is consistent, and the claimant draw is deterministic and independent of other draws.
• Merging with current main is clean, and UK: take up #1095's actionable items on policyengine-uk 2.165.1 #1422's floor-of-ten guards find nothing in this PR's added lines.
• CI is green at a8e25870c.

Disclosure-control items went to the author separately.

🤖 Generated with Claude Code

[juaristi22]

The review of a8e25870c: what changed

Thanks for the review. The fixes follow María's rulings of 1 Nov, in five commits:

• 91775db4a: the modelling fixes (findings 1, 6, 7 and 10);
• 24e2329f6: tests, limits and texts (3, 4, 8, 11, 12 and 13);
• 59ef8ecee: the regenerated H2 fixture and coverage manifest;
• 9f70c71e7: the rebase banded once (finding 6). The spine from 59ef8ecee showed that the iterated banding doesn't converge;
• 4dd83e999: the data contract's full-manifest pins. 91775db4a had moved the release-cut digests for the new export column but not these pins, so the release directory failed the contract until this commit.

The rebuild: the national from 9f70c71e7

The national from 9f70c71e7 is certified:

• all 50 spine gates pass, and so do the 7 calibration-seam gates and the 19 release-cut gates;
• the post ran clean: the stage-1 comparison, the head-to-heads against 4374e0bf and a6299504, and, after 4dd83e999, assembly and the publisher pre-flight.

It is staged, not published. PolicyEngine/policyengine.py#731 still bundles a8e25870c's cut, which carries findings 1 and 2.

The table shows the change in the government balance, in £bn, from PR A's dashboard scoring on policyengine-uk 2.162.14, as in the measurements comment. Years are policyengine-uk's (2028 is 2028-29). The OBR line covers the taxes the model covers, after behaviour.

Measure	Years	a8e25870c	9f70c71e7	OBR line	Table 6.1
Trading allowance to £3,000	2028 to 2031	−0.148 / −0.149 / −0.148 / −0.147	−0.173 / −0.175 / −0.173 / −0.170	−0.221 / −0.165 / −0.171 / −0.178	−0.225 / −0.165 / −0.175 / −0.180
Savings income rates +2pp	2029 to 2031	+0.582 / +0.611 / +0.636	+0.587 / +0.615 / +0.640	+0.443 / +0.475 / +0.510	+0.520 / +0.555 / +0.590
Property income rates +2pp	2029 to 2031	+0.402 / +0.417 / +0.429	+0.414 / +0.429 / +0.441	+0.394 / +0.403 / +0.412	+0.435 / +0.445 / +0.455
Carer's Allowance earnings taper	2030, 2031	−0.037 / −0.073	−0.035 / −0.070	—	−0.040 / −0.080
Domestic electricity at zero VAT	2026, 2027	−0.788 / −0.332	−0.789 / −0.331	—	−0.755 / −0.305
Fuel duty rise deferred	2026, 2027	−0.290 / −0.501	−0.290 / −0.500	—	−0.575 / −0.780

What moved:

• Trading allowance: the cost rises about a sixth, all of it income tax.
  • NICs don't move. Every drawn claimant's profit stays under Class 4's lower profits limit, so only income tax changes.
  • The rise sits in the SPI copies. Their income draws read the rebased tape, and more of them gain, by more. FRS rows don't move.
  • Against HMRC's policy paper: on this national the raise gives about 436,000 taxpayers a lower bill in 2028-29, by about £407 on average. HMRC gives about 400,000 and £380, and a8e25870c gave 409,000 and £370.
  • So the claimant rates now overshoot. They were calibrated on b441bb5a0, and on this national they overshoot HMRC's static base by about a sixth. That, not a closed behaviour gap, is why the cost now meets the OBR's income tax plus NICs line from 2029-30.
  • UK trading allowance: store the claim status once policyengine-uk#2765 is released #1439 makes re-measuring gainers after each build a committed check.
• Savings rates: up under 1%, still 25–33% above the OBR's income tax line. The tax on savings in 2025 sits at 0.99, 0.99 and 1.00 of HMRC's Table 2.6 projection, by rate. Max's check on how much of that gap is behaviour still stands.
• Property rates: up about 3%, to 5–7% above the OBR's income tax line. The rise also sits in the SPI copies: their property draws read the rebased tape, and they carry more property income. FRS rows barely move.
• The other measures barely move.
• The pension tax allowance measure still fails in the scoring template on policyengine-uk 2.162.14, not on the data.

Fix before the national is used

1. Annuity and trust income: back in pension income (91775db4a). Done as you suggested.

• FRS PENTYPE 4 and 5 stay in private_pension_income, so the means tests, household income and tax read them as they did before this PR.
• A calibration-only descriptor, annuity_and_trust_income, stores their part. The export carries it.
• Table 3.7's other-income rows read a new measure, other_investment_income_spi_concept: the engine's other income plus a share of the engine's private pension. The share is each person's descriptor over their stored private pension, so a reform that changes private pension moves the measure with it.
• Rows the SPI stage draws carry a zero descriptor. Their private pension is the tape's, which books these payments as other income.
• The FRS spine now refuses a pension tab without PENTYPE. The H2 fixture's pension rows carry PENTYPE, so the fixture exercises the routing.
• The body's "the tax doesn't move" now covers NS&I only, which stays in interest.

2. Non-claimants' receipts above £3,000: kept and documented, with a follow-up.

• María kept the encoding for now. It scores the Budget's £3,000 measure, and storing the claim status from Add a trading allowance claim status, so allowance reforms reach claimants without turnover data policyengine-uk#2765 will replace it once that is released.
• The stage spec, the module and the changelog state the limit. Any allowance above £3,000 × the index relieves every non-claimant from the stored expense, so such reforms shouldn't be scored on this national.
• UK trading allowance: store the claim status once policyengine-uk#2765 is released #1439 tracks storing the claim status once #2765 is released, which removes the limit.

Merge hazards

3. Empty fit-weight records: a roster test (24e2329f6).

• test_uk_fit_weight_records_roster.py reads each uk_runtime module's source. It pins the eight stage transforms that declare fit_weight_records.

• UK: company cars from the FRS job table (dashboard #316) #1430's company_cars stage fails it on merge. Its author then has to record each fit or drop the hook.

• Not fixed here:

  • the NTS stage's published-shares fallback still stores (). It's latent, because the production donor takes the interview path;
  • uc_reporter_redraw, salary_sacrifice and spi_housing_shell still fit without the hook, so the audit doesn't see them.

  Both are follow-ups for after this PR. A run that made no fit should say so in a way the audit accepts, and the three stages should record their fits.

4. The receipts test's labels (24e2329f6).

• They're built with rounded_count, so the test holds under either floor.
• test_uk_uc_take_up.py changed again for finding 7, so its conflict with UK: take up #1095's actionable items on policyengine-uk 2.165.1 #1422 still needs resolving by hand at that merge.

5. The UC take-up gate with #1426: not changed here.

• The value needs a ruling when the two PRs meet. The body's "Not in this PR" says so.
• On the spine from 9f70c71e7, the UC take-up gate passes at 0.055.

Modelling

6. Band-edge migration: counted once (91775db4a, 9f70c71e7). This is your first suggestion.

• The tape is banded on its income moved to 2023-24. That is each income leaf times its 2022-23 to 2023-24 factor, leaving out finance costs and the loss draws, plus any part of total income the leaves don't itemise.
• Iterating, your second suggestion, didn't converge on the production tape. 91775db4a moved each record's savings by its own band's factor and repeated the banding. Records kept changing band from pass to pass, because a record's band sets its factor and the factor sets its band. The spine from 59ef8ecee took the twentieth pass, and its receipt said it hadn't converged.
• 9f70c71e7 bands once. Each savings column moves by its factor over all bands, on the tape's 2022-23 bands. Banded so, the rebased tape reproduces the published amounts band by band. A record whose band's factor differs from all bands' can still sit across an edge on its rebased income.
• I checked this locally against your measure: the tape's rebased savings, re-banded on each record's own rebased 2023-24 income, against the published bands. The single pass came closer than both the 2022-23 banding and the twentieth pass, with the top bands near their published amounts. The figures rest on the tape, so they stay out of this thread.
• The receipt records each column's banding factor in place of the pass count.
• Tests:
  • one builds a tape whose income grows across an edge, and checks the record is banded where the consumers will find it;
  • another checks that a saver whose savings are a large part of their income is banded on them.

7. The UC pool join: replaced by passing gaps (91775db4a). This is your suggestion.

• A pool that can't meet its count is held at its clip, and its gap passes on.
• Opposite gaps in the family type offset first. What remains goes to the other pools, nearest first, each taking what its room allows. At equal distance the pool with more room goes first, and on a tie the lower band.
• Every band keeps its own rate.
• The tests now cover an overshooting pool, a short pool, and gaps that offset before passing.

8. Who gains, and the drift (24e2329f6, #1439).

• The module, the spec and the changelog give the drift exactly: £1,932 more in 2027-28, falling to £1,758 by 2031-32 (3.4% to 12.1% below £2,000). That replaces "slightly".
• Full relief for drawn non-claimants whose receipts are within the allowance moves with the claim status in UK trading allowance: store the claim status once policyengine-uk#2765 is released #1439. So do SPI copies with capital allowances, which can't have claimed partial relief. The frame doesn't carry capital allowances today, so that needs a flag from the SPI draw. Both are now noted on UK trading allowance: store the claim status once policyengine-uk#2765 is released #1439.

9. The gift anchor's concentration: acknowledged.

• The changelog says the anchor fixes totals, not concentration.
• The stage evidence keeps factors only. A concentration statistic for columns resting on few SPI records would be disclosive.
• A published total for the charity reliefs on gifts of shares and property would let a target bind them. That is in "Not in this PR".

10. The taxpayer basis and the guard (91775db4a).

• A taxpayer is a record with TAXPAYER 1, or one whose income crosses the first band by 2023-24. So a saver who becomes a taxpayer counts once.
• A flagged taxpayer below the first band sits in it (note 8).
• A band under a tenth of its published amount takes the all-bands factor, and the receipt names it.
• A column the grossing guard skips grows on its engine index from 2022-23, not on the published path.
• is_taxpayer is written only when the tape has TAXPAYER, and the receipt's label follows it.
• TAXPAYER codes other than 1 and 2, or blanks, are refused.

11. The support floor (24e2329f6). Its counts scope now takes Table 3.10's employment count by self-employment band. A cell without support is dropped and receipted.

Tests and texts

12. Target tests (24e2329f6).

• New tests pin the income-based JSA row's selector, window and binding, and the employment count's variable, grouping, filter and bands.
• The claimant rates' tool and a post-build gainers check are in UK trading allowance: store the claim status once policyengine-uk#2765 is released #1439's "Done when".

13. Texts (24e2329f6).

• The £1m-and-over interest hold now tracks UK: data fixes for the Budget 2026 household measures #1418.
• The income anchors doc no longer contradicts itself. It covers the taxpayer basis and the guard, and its other-income analogue is gilt interest.
• The uprating test describes the two-leg savings factor.
• The NS&I and gilt comments, and the two target notes, describe what ships.
• The employment count's note says SPI rows count benefits in kind (PAY + EPB + TAXTERM).
• The changelog fragments describe the change, not the branch's history.
• The FRS spine refuses a missing PENTYPE (see 1).

🤖 Generated with Claude Code