[juaristi22]

Part of #1095. This PR collects the #1095 items that pass their A/B measurement against the 28 October national build, and more land here as they pass. It holds three so far:

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

Every figure that rests on the SPI channel is given as a class only (SN 9422 and the disclosure rule).

How the items were measured

• Control: the 28 October national build, attempt 4374e0bf at 7757add (staged on populace-uk-private). Its data path is main's at fddbc70, which adds only wording and the 0.055 UC take-up gate.
• Arms: each item is one national arm (one spine and one calibration) built on fddbc70, with the control's review date (2026-10-28) and feed (272c987a).
• Placebo: main with the SPI stage seed moved from 42 to 142 (local only) sizes the noise of a redraw.
• Cross-machine check: the control's spine was built on another machine. Its 17 stages before the SPI income stage give receipts and records identical to the new spines'. So the arms differ only from the stages the items change.

	Rows within 10%	Final loss	Rows moving more than 5 points against the control
Control	99.0% of 1,300	0.00529	—
Placebo	98.8% of 1,299	0.00560	23
1. SPI age band	98.9% of 1,299	0.00539	11
2. HRAD allocation	99.2% of 1,300	0.00531	0
This branch (1, 2 and 3)	98.8% of 1,299	0.00536	8

Every arm passes every spine gate and all seven calibration gates; this branch's passes with the take-up gate at 0.12 (item 3). Each arm that changes the SPI stage, the placebo included, loses one row to the signed zero-support floor.

This branch is rebased onto main at 945cb23. Since fddbc70, main has moved only in NZ code and in shared ledger compile modules; no UK target or compile-parity file changed.

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

Why

The 2022-23 SPI public-use tape gives age only as a band (AGERANGE). The SPI income stage draws an age for each donor inside its band and trains the stage-1 tape forest on it, so the forest learns no real income pattern within a band. Twins of one band were then drawn from conditionals that differed only by that noise. uk-data#628 reports the same defect in uk-data and proposes training and querying the forests on the band, measured against a draw-seed placebo.

What changes

• The stage-1 tape and reference forests read spi_age_band instead of age. On the tape it is the AGERANGE code, with composite records (-1) as a band of their own. On a frame person it is the band of their whole-year age, through one helper shared by both frames.
• at_or_over_spa still carries the State Pension split inside the 65-74 band.
• The cross-fit fold cells key on the band, since they key on the predictor values.
• The drawn donor ages stay on the donor for their other readers: the band pools' State Pension side, and the property components' tape seed.
• The stage receipt's stage1_age_predictor names the rule. sources.yaml declares the predictor and the fold key.

Measurement

• Movement.
  • Against the control, this item takes four rows outside 10%: dwp.uc.households_children_3, and three HMRC regional cells (income tax at £150,000 to £200,000 in Wales; taxpayers at £50,000 to £70,000 in the East Midlands and London). It takes three UC family rows inside.
  • The placebo takes five rows outside and two inside.
  • The HMRC regional and income-band rows move no more than under the placebo, and the UC rows end closer to DWP's than in either other arm.
• The age gradient (Chronicle: one append-only trace for every build attempt at every rung #628's symptom). This is the weighted rank correlation of age with pay within each SPI age band, on drawn SPI copies against FRS employees.
  • On main, the copies' pay barely rises with age inside the youngest bands, where FRS employees' clearly does.
  • With this change, the copies' pay rises with age inside 16-24 and 25-34 nearly as the FRS rows' does, and falls with age over 55 as theirs does.
  • The placebo changes none of this.

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

Why

The property purchases stage (#1308, porting uk-data#573) fits households' HRAD purchases to Table 6b's 2024-25 higher-rates transactions and value. It allocates them to households by individuals' share of all Table 8a residential transactions (0.881) and value (0.882). #1308's review found that this probably overstates households' purchases (#1095, note of 17 October). María's ruling of 17 October left allocating from Table 8a's band receipts for later.

The residential nil-rate band was £250,000 for all of 2024-25. A purchase at or below it owes no standard-rate SDLT, so other buyers' receipts in that band are surcharge. At Table 6b's gross HRAD receipts per transaction in the band (£625m over 113,500 transactions), other buyers' £265m of receipts there is about 48,100 HRAD transactions and £7.05bn. The all-band share left them about 13,400.

What changes

• The nil-rate band from receipts. Households' HRAD at or below £250,000 is Table 6b's band less other buyers' HRAD implied by their Table 8a receipts, at Table 6b's gross receipts per transaction and per pound. The stage refuses an implied figure outside Table 6b's band or above other buyers' own transactions there.
• Above the band, uk-data#573's shares stay. There the standard rates make up most receipts, and solving for other buyers' HRAD from them goes negative above £1m. So the shares remain, and probably still overstate households' purchases. The module docstring and the stage declaration say so.
• The price mix follows the households' bands. The lowest band's share of households' purchases falls from 54% to 43%.
• Targets. Households' 2024-25 HRAD target falls from 186,589 transactions and £68.81bn to 151,928 and £63.71bn. Both stay above uk-data#573's lower bounds, 87,200 and £33.28bn. The non-residential kind is unchanged.
• Receipt. The stage's derivation adds the nil-band receipts, the implied HRAD and two shares: other buyers' HRAD over their nil-band transactions (0.625), and households' share of the band (0.576). uk_stage_property_purchases_fit needs no change, since every *_share key stays in (0, 1].

All figures in this section are HMRC's published Annual Stamp Tax Statistics 2024 to 2025 (corrected, 11 March 2026), Tables 6b and 8a, as the pinned Chronicle feed (272c987a) carries them.

Measurement

• The purchases stage.
  • Households' HRAD purchases fit at 152,849 against the new target of 151,928 (+0.6%), with the value at £63.71bn exactly and no shortfall. On main they fit 186,520 against 186,589.
  • The same value now spreads over fewer purchases, so their mean price rises. The share of purchasers held at their own stock rises from 37% to 48%.
  • Non-residential purchases are unchanged.
• Movement. No row moves by more than 5 points. Three rows move inside 10% (dwp.esa_claimants and two UC family rows), and none move outside.
• obr.sdlt.
  • Before calibration the model's receipts exceed OBR's £10.16bn by 49.2%, against 50.9% on main. That move is inside the placebo's, which gives 47.2%.
  • After calibration the row is within 0.25% of the OBR figure in every arm.
  • The rise in mean price offsets most of the drop in count, so the surcharge receipts fall less than the purchases.

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

Why

The SPI income stage keeps each drawn copy's FRS twin's hours and job, while stage 1 redraws the copy's pay from the 2022-23 public-use tape. Stage 1 draws on age, sex, region and the bridge predictors, but not on hours. The tape's pay covers the tax year, and the tape holds part-year workers. So part-year tape pay can land on a current full-time twin, at an hourly rate far below the minimum wage.

Such a copy fails the minimum-income tests of the 30 hours' childcare offer, Tax-Free Childcare and the benefit cap more often than the FRS rows it copies. uk-data#629 reports the same defect in uk-data.

On the national spine, drawn employee copies sat below the minimum wage at their twin's hours far more often than FRS employees. Most of those copies had a normally paid twin working 16 hours or more. The tape's low-pay share was above FRS's in every working-age band, which points to part-year pay rather than survey noise. These are private figures, given here as classes only.

What changes

• The floor. A new floor_spi_employment_pay_at_minimum_wage operation of hmrc_spi_income_spine lifts a drawn copy's hmrc_spi_pay leaf to:

  • its twin's usual weekly hours (hours_worked read back at 52 weeks, uk-data#531),
  • times 52 weeks,
  • times the lower of policyengine-uk's gov.hmrc.minimum_wage.non_apprentice rate for the twin's age (read as the engine reads the build year: £11.44 from 21, £8.60 from 18, £6.40 below) and the twin's own hourly pay.

  The cap at the twin's own rate means no copy is held to a higher rate than the FRS person it copies. The first two commits carry over María's floor from the childcare fixes work.

• Who it touches. Only drawn recipients other than reserved band carriers, whose twin has an employee main job, no trade and usual hours, and whose drawn pay is positive. FRS rows are never modified.

• When it runs. After the stage-1 draw, the band-donor resample and the State Pension leaf guard, and before employment_income is derived. So stage 2, pension coherence (the twin's employer rate on the floored pay) and every later stage read the floored current pay. The transport receipt keeps the unfloored draws' uniforms.

• The tape's year of pay. The floored pay is current pay only. The row's employment_income_annual_share becomes its unfloored over its floored employment income, so the annual measures that the HMRC income rows read still read the tape's pay.

• Receipt. The stage evidence gains an employment_pay_floor receipt. It also gains an hourly-pay check: the drawn employees' weighted shares below the minimum wage for their age before and after the floor, beside the FRS employees' share, and the eligible rows still below their floor. Any figure on one to nine source households, or beside a side on one to nine, is "<10".

• Gate. A new release-blocking stage gate, uk_stage_hmrc_spi_income_spine_hourly_pay, refuses a spine in either case:

  • an eligible drawn employee stays below its floor;
  • after the floor, drawn employees sit below the minimum wage more often than FRS employees by more than two points.

  Its details give verdicts only. It joins the spine gate scope. The gate battery's three digests and the spine certification part's two move with it, in microcosm-data's contract.py and its mirror test.

• The UC take-up gate. uk_stage_uc_take_up allows each family type 0.12 of DWP's count, up from 0.055, until UK: every open data item: uk-data parity, engine locks, the local surface, consumption and high-value homes #1095's UC award fix lands, and then returns to 0.055 (María's ruling of 29 October).

  • The floor gives fewer low-paid copies UC, so single people's caseload before the take-up redraw falls close to DWP's.
  • The redraw's usual downward move then reads as about twice the old allowance. After the redraw, single people end a little below the spine without the floor, all in the middle award range.
  • The shortfall of middle and higher awards that the redraw cannot fill is the award fix's.
  • The one tolerance also bounds the reachable pools, which realize DWP's count far more closely.
  • The gate's digests are re-pinned with it.

Measurement

The floor was measured two ways:

• On this branch: one arm with all three items. It passes every spine gate, the take-up gate at 0.12 included, and all seven calibration gates. Against the SPI age band alone, it moves no row by more than 5 points. One HMRC regional cell (total income at £50,000 to £70,000 in the East Midlands) goes outside 10%.
• On its own: with the HRAD allocation and the take-up gate widened on a local branch, against the HRAD allocation alone.
  • No row moves by more than 5 points, and the starting loss is lower.
  • Four rows already near 10% move outside it: dwp.esa_claimants, two UC family rows, and the HMRC regional taxpayers at £50,000 to £70,000 in London.
  • The HMRC employment-income bands end as without the floor, since their annual measures read the tape's pay.
  • Employee National Insurance ends further from OBR's figure, still within 10%: floored current pay raises the engine's NICs.
• The floor itself:
  • It lifts a large share of eligible drawn employees.
  • After it, drawn employees sit below the minimum wage for their age less often than FRS employees do, within the gate's two-point margin.
  • No eligible row stays below its floor.
• Across the award ranges: a diagnosis of UC awards against DWP's award bands finds the floor's single claimants leave only the middle range. Single people and lone parents were already short of DWP's middle and higher awards there before the floor; that is UK: every open data item: uk-data parity, engine locks, the local surface, consumption and high-value homes #1095's UC award item.

Tests

Touched surface, JUnit counts on this branch:

• Engine-free (the SPI, purchases, gate, contract, manifest and spec files): 898, 0 failures, 1 skipped.
• Engine tier (H2 parity and the SPI engine tests): 26, 0 failures.

Test files:

• test_uk_spi_pension_age.py: the query predictors carry the band; the band and the State Pension flag still split State Pension inside 65-74.
• test_uk_spi_rank_transport.py: the fold cells on the band.
• tests/engine/uk/test_uk_spi_income.py:
  • the stage-1 queries carry the band and no drawn age;
  • the reference forest is counted once per cross-fit fold in use;
  • the floor through the whole stage: each twin and its copy carry one age and pay, and the test checks the twin cap, the kept year of pay, and the employer contribution that follows the floored pay.
• test_uk_property_purchases.py: the band allocation on synthetic rows, the refusal when receipts imply more HRAD than other buyers' transactions, and the packaged targets reproduced from the published table figures.
• test_uk_spi_income.py (engine-free): the floor on synthetic rows, the twin cap, the hourly-pay check and its small-cell withholding, the comparison with FRS rows, and the year-of-pay helper.
• test_uk_stage_health.py:
  • the hourly-pay gate holds drawn employees to their floor and to the FRS share plus two points;
  • the take-up gate pins 0.12 and passes the floor's pattern.

Follow-up

• uk-data#573 still allocates every band by the shares; raising this there is María's call.

Generated surfaces: source_stages.json (in step with sources.yaml), the sources schema, the release input coverage manifest, the H2 parity fixture, and the gate digests mirrored in microcosm-data's contract. The age band moves the synthetic CGT, purchase and rental facts that the fixture sizes after the SPI stage.

🤖 Generated with Claude Code

[vahid-ahmadi]

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

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

Findings

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

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

🤖 Generated with Claude Code

[juaristi22]

@vahid-ahmadi thanks for the review. All five points are addressed. #1422 has landed, and main is merged in at bc7f59aac.

1. Composites next to 16-24: the age band is now one of the stage-1 categorical predictors, one-hot encoded with gender, region and main source, so composite records are a category of their own (dae46ae94). The predictor rule, the spec's categorical list and the changelog say so. The next build re-measures the three items with it.
2. The floor receipt's ten: the receipt withholds through UK_DISCLOSURE_FLOOR and UK_WITHHELD ("<3"). Its descriptions, the hourly-pay gate's note and docstring, and its tests follow (f1092f14c). Two test cases shrink their small side to two households, so they still test the withholding, and the engine test's floored households read "3-9" (8f0ec8fd7). UK: take up #1095's actionable items on policyengine-uk 2.165.1 #1422's floor guards pass on the merged tree.
3. One tolerance, two checks: the reachable pools keep their own maximum_reachable_pool_deviation at 0.055, beside maximum_caseload_deviation at 0.12 (d68eb5cbb). The caseload check returns to 0.055 when UK: every open data item: uk-data parity, engine locks, the local surface, consumption and high-value homes #1095's UC award item ("UC awards below the top band are too low") lands, per María's ruling of 29 October.
4. Table 8a and the non-resident surcharge: HMRC's commentary does not say whether Table 8a's receipts include it. HRAD makes up more than half of the residential receipts it reports, so they read as whole SDLT receipts. The module docstring now says which way the surcharge goes: read as HRAD surcharge, it raises other buyers' implied HRAD and lowers households', so households' nil-band purchases are, if anything, understated (f3d9c0af0).
5. paid_weeks: renamed paid_hours (bc5adc97c).

The H2 fixture is regenerated on the merged tree (bc7f59aac). The full UK engine tier passed locally before main's merge, and CI runs it on the merge. #1095's Phase B joins this PR next (María's call of 31 October). Each item is measured against the national control built from #1428 at 329a99e71 (cut f6e33b93), on policyengine-uk 2.165.1, the lock main and #1428 share.

[juaristi22]

@vahid-ahmadi a follow-up on point 3: #1428 is now merged in at the control's commit, and on the T1426 spine every family type ends closer to DWP's count after the take-up redraw than before it. So the caseload check is back at 0.055 (0ddc23673, María's ruling of 2 November), and nothing is left to track. The pools keep their own maximum_reachable_pool_deviation, also at 0.055.