Draw SPI incomes at each donor's FRS rank within their cell

[MaxGhenis]

What this changes

PR #529 draws each SPI-synthetic person's incomes from SPI records in their own earnings group, so employees draw pay and the out-of-work draw none. Within the group the draw was still random: microimpute picks a random quantile, conditioned only on age, gender and region. The donor keeps their hours and FT/PT status, so a part-timer drew a full-timer's pay as often as a part-timer's. The SPI has no hours.

This PR draws each income at the person's FRS rank of that income within their cell instead of at a random quantile. This is the uk-data analogue of PolicyEngine/microcosm#840, "rank-preserving replacement". A part-timer on low FRS pay draws low SPI pay. A donor at the top of their cell draws from the top of the forest's distribution for that cell.

• Rank variable: per component. Pay sets the pay draw, profit the profit draw, a private pension the pension draw, and so on.
  • For employees and the self-employed, this is the same as ranking on total earnings.
  • People with both pay and a trade keep each source's own position.
  • People with no earnings (21m weighted) get ranks from their pension, savings and property income. A total-earnings rank would leave their draws random.
  • The FRS does not record Gift Aid or charitable investment gifts, so those are drawn at a uniform random quantile, independently of each other and of every other output.
• Ties and zeros. Within a cell, people are ordered by the income, with equal values in random order. Equal values are common: zeros, and the FRS's heaped pay, where only 16% of positive values are distinct. Each person takes a uniform point in their slice of the cell's weight.
  • The slices tile [0, 1), so the quantiles are exactly uniform in every cell, whatever the ties.
  • Zero is just the lowest tie. An employee with no recorded pay draws from the bottom of their cell and never above one with pay.
  • Mid-ranks are rejected: they put a whole tie, such as everyone without a private pension, at one quantile.
• What uniform quantiles preserve, and what they don't.
  • For each group's first output, the draw keeps the forest's distribution for the cell, in expectation. The forest is queried at the person's exact age, but its training age is uniform noise within the SPI age band. FRS pay rank tracks age within a band, so the rank draw also picks up the forest's chance age pattern. On the full FRS, pay over £100k, summed over the three pay groups, is 0.802m at FRS pay ranks, against 0.845m with random within-cell ranks (SD 0.033m) and a forest expectation of 0.848m. For the EMPLOYEE group alone it is 0.780m against the forest's 0.827m.
  • Later outputs are conditioned on those drawn before, at correlated FRS ranks, so they carry more dependence between incomes than the SPI does. It is still less than microimpute's single random quantile per person gave (table under Evidence).
• Cells: (earnings group, SPI age band, gender, region). These are the finest cells the SPI can tell apart, since it gives age only as a band. The forests are queried at the person's exact age and, after each group's first output, at the outputs drawn before it. So the cells match the forests' conditioning only up to that age noise and the chain.
  • Small cells are not pooled. Ranking a London donor nationally would map their higher national rank onto London's already-higher SPI distribution.
  • A one-donor cell gets a uniform random quantile, as before.
• People with both pay and a trade are split by main source.
  • SPI: MAINSRCE 1 means pay is the main source; 3 (sole trader) or 4 (partnership) means a trade. Any other code (2 occupational pension, 5 other, 6 claims case, −1 not classified) leaves it to the larger income.
  • FRS: the main job's status.
  • Main source of income is one of the variables HMRC stratifies the self-assessment sample by (HMRC, Survey of Personal Incomes 2022-23: public use tape documentation, p. 3; MAINSRCE codes on p. 15). It is not the larger income: 57% of trade-main taxpayers have more pay than profit.
  • The FRS has 294 such people, 1–2 per rank cell, so their main source is most of what links their draw to their own jobs.
• The draw reaches the tails.
  • Before: microimpute 1.8.1, the locked version, picks from ten forest quantiles between 1/11 and 10/11, so no one drew from the top or bottom 9% of the forest's conditional distribution at their predictors.
  • Now: outputs are drawn at the nearest of 1,000 midpoints of that distribution, from 0.05% to 99.95%.
  • The draw reimplements microimpute's chain (each output conditioned on those drawn before) at given quantiles. At microimpute's own random draws and grid, it reproduces predict bit for bit, which is tested. A model type it does not know raises.
• The FRS rows' dividend draw uses the same ranks. Ranks are computed once, on the full FRS at grossing weights, and both draws use them.
  • Before: every FRS adult's dividends were drawn conditioned on a random drawn pay.
  • Now: they are conditioned on pay and savings drawn at the person's own ranks.
  • The FRS dividend column itself is empty until Keep FRS-reported dividends and key them on person_id #498 lands, because frs.py keys it on person.index. Until then the dividend quantile ties and is random.

Base and overlaps. Branched from main 4cbedbe (uk-data 1.58.0), which includes #529. On 10/14, main daad71c (uk-data 1.58.1, the #576 batch: #555, #540, #518, #541) was merged into this branch at 5621c0d. The merge was clean and main does not touch income.py. On 10/15, main a43e8c5 (uk-data 1.58.2, which adds #597's survey household provenance) was merged in at a017674, again cleanly; the real builds use a43e8c5 as their base. Not part of a released batch yet; it waits for the next batched release (d833).

Interactions with what main gained. create_datasets.py runs impute_income before impute_salary_sacrifice, so #518's limits apply to the pay this PR draws. Stage 1 caps each SPI-synthetic row's salary sacrifice at its drawn pay. Stage 2 moves at most half a donor's drawn pay into salary sacrifice and takes it out of cash pay, so a row drawn at £4,000 or less can gain at most £2,000 of new sacrifice, and final cash pay can reorder after the draw. The capital gains donor selection (cgt_band_donors) also runs before salary sacrifice and reads the drawn incomes. The real builds measure the size of these effects. #540 leaves REGION_MAP unchanged and sets Scottish taxpayer status only in create_spi, which the income imputation does not use. #597 adds source_household_id inside stack_datasets, which impute_income calls only after both draws have run, so it cannot change the ranks, the cells or the drawn values. The copy keeps its donor's FRS household ID as its source. long_geography.py reads that ID only for its area-support diagnostics, which count a donor and its copy once within an area; it does not change local weights.

Open PRs that also edit income.py:

• Keep FRS-reported dividends and key them on person_id #498 keys FRS dividends on person_id and deletes the FRS-half dividend draw, so FRS rows keep their reported dividends. After it, dividend ranks matter only for the SPI copy, where this code ranks the recorded dividends with no change. This PR's FRS-half rewiring is therefore transitional. Whichever of the two merges second drops the FRS-half impute_over_incomes(dataset, model, ["dividend_income"], quantiles) call. Keep FRS-reported dividends and key them on person_id #498's test stub of impute_over_incomes then needs the quantiles argument, as Set UC gainful self-employment from the FRS main-job status #525's did here.
• Rebase SPI income draws to the FRS survey year #532 rebases SPI draws to the FRS year by wrapping output_df = model.predict(input_df) in impute_over_incomes. This PR leaves that line as it is: the rank draw happens inside the model's predict. The two compose. The rebase is a positive scaling per variable, so it keeps ranks.
• Set the UC start-up period from the FRS claim and business start dates #527, Put FRS self-employment losses in trading_loss #528 and Move legacy-benefit reporters onto Universal Credit at Move to UC claim rates #534 also touch the file. Whichever lands second resolves the overlap.

Evidence

Real builds pending. Production builds (base, branch, and a placebo redraw of main) have not run yet. Post-release rebuilds run one at a time through the UK hub's rebuild queue, and each needs more than 100 GB free on the build host's data volume. This PR is row 4 in that queue and has no go yet; the volume had 43 GB free on 10/9, 56 GB on 10/15 and 71 GB on 10/16. Real-build figures for hours and pay, calibration and the policyengine-uk microsimulation will be added here before the PR leaves draft.

Held out of the next batch on CI calibration fit. Under #599's rule, this PR's TESTING build has 2.44pp fewer local targets within 10% than the 1.58.1 release branch's (W3's comment). Net, all of it is constituency self-employment income amounts, almost all in England (decomposition). This PR fits that family better than main's TESTING builds at epochs 10 and 20 of 32, then overshoots at epoch 30, the epoch whose weights the build keeps. Main's 512-epoch release builds dip in the same family at epoch 30 and recover by epoch 50. Whether this PR recovers is not yet measured; the real builds above will show it. Until then the PR stays a draft and out of the batch.

Offline, measured by review r1. These use this PR's code and forests on the full FRS at FRS weights, with single draws in SPI 2022-23 money. They are not a build: no calibration and no policyengine-uk.

Hourly pay below £5/hour moves from full-timers to part-timers, but the overall share does not fall:

Share of employees below £5/hour	Status quo	This PR
All employees	17.0%	16.9%
Full-time	18.1%	15.1%
Part-time	12.7%	24.2%

• The FRS's own figure is 1.8%.
• The SPI employee group is 33.0m weighted, against 27.6m FRS employees. 10.8% of it has annual pay under £4,160, from part-year and marginal pay.
• Rank draws hand that low-pay mass to the employees with the lowest FRS pay, who are mostly part-timers. The group-definition mismatch predates this PR and is listed below as a follow-up.

Dependence between incomes (weighted Spearman):

Pair	SPI	Forest, independent quantiles	This PR	Status quo
Employees: pay–savings	0.146	0.150	0.204	0.343
No earnings: pension–savings	0.397	0.486	0.551	0.658

Totals on the full FRS:

Output	Status quo	This PR	SPI-cell expectation
Dividends (the FRS-half draw, which carries real weights)	£41.3bn	£60.7bn	£59.9bn
Gift Aid	£7.29bn	£2.54bn	£2.52bn
Charitable investment gifts	£2.80bn	£0.05bn	£0.05bn
Savings interest	£7.16bn	£9.03bn	£7.77bn
Self-employment profit	£87.0bn	£101.2bn	£90.1bn

• Most of this undoes microimpute 1.8's shared quantile and its ten-point grid.
• The profit and savings overshoots come from the forests' tails, not from the ranks: independent quantiles give £103.6bn and £9.18bn.
• Property income for people with no earnings is £23.7bn, against £20.5bn with independent quantiles.
• These move the calibration inputs (dividends, Gift Aid relief); the real builds will show by how much after calibration.

Over £100k the rank draw matches the SPI cells on the build's SPI copy: 0.512m at FRS ranks, against 0.511m from the SPI cells and 0.524m from the forest, counting every adult in the 10,000-household copy at FRS weights. Over £150k it still tracks the forest, but the forest does not track the SPI cells:

Copy adults with pay over £150k	People
SPI-cell expectation	0.205m
microimpute 1.8.1's own draw on this PR's forests (drawn on the full FRS, then subset to the copy)	0.274m
This PR, at FRS ranks	0.306m
Forest expectation	0.325m

Over 1,000 copy subsamples the rank draw's excess over the SPI cells is +45% (SD 9%), so it is not subsample noise. The 0.274m row holds the forests fixed, so it is a controlled comparison, not the base build's own result: base has one forest for people with both pay and a trade, and drawing on the copy alone restarts microimpute's random sequence. Holding the forests fixed, the rank draw adds 32k weighted people over £150k (+11.9%), and the excess over the SPI cells rises from 33% to 49%. The excess belongs to the forests, and the 1,000-point grid exposes more of it than the ten-point grid did. The EMPLOYEE, NO_EARNINGS and SELF_EMPLOYED forests are the same in base and branch. Why the forest's tail exceeds the SPI cells' for these donors is not established. The real builds will report SPI-row pay over £100k and £150k after calibration to HMRC's income bands.

The forest's top tail also depends on its training resample: EMPLOYEE pay over £150k on the full FRS ranges from 0.26m to 0.50m across four resamples, and this PR's resample gives the 0.50m. The EMPLOYEE, NO_EARNINGS and SELF_EMPLOYED forests are identical between base and branch (same group index, so the same resample seed and size), so the base–branch delta isolates the draw rule for those groups.

Invariants (tested)

policyengine_uk_data/tests/test_spi_income_rank_draw.py (Hypothesis unless noted):

1. Quantiles lie in [0, 1). Within a cell, a lower value never gets a higher quantile, and a weighted row always gets a strictly lower one. The rows' slices tile [0, 1) and each quantile lies in its row's slice, so the quantiles are uniform in every cell.
2. Across seeds, a row's quantile is uniform on its slice, not a fixed point in it (KS test). A lone donor does not always draw its cell's median.
3. With equal weights and m rows per grid point, every grid point is drawn exactly m times.
4. Quantiles do not depend on row order, on strictly increasing transforms of the values, on other cells' values, or on the scale of a cell's weights.
   • Equal seeds give equal results.
   • A row whose value is unique in its cell stays in its slice whatever the seed.
5. Missing values or weights, negative weights and duplicate ids are rejected.
6. Differential test: draw_at_quantiles at microimpute 1.8's random draws and grid equals microimpute's predict exactly.
7. The first output drawn never falls as the quantile rises, and every draw lies within the training range.
   • Each output is drawn at its own quantile: raising one output's quantile never lowers it and never changes an output drawn before it.
   • An unknown model type raises.
8. The group model draws each group at the given quantiles, gives NOT_IMPUTED rows no draw, and without quantiles is deterministic.
9. impute_over_incomes passes quantiles by person ID: a subsample drawn with the full data's quantiles gets exactly those, and more FRS income never draws lower in a cell.
10. draw_quantiles ranks within rank_cells at household weights: in every cell, each person's quantile lies in their slice of the cell's household weight, for every output.
11. Outputs everyone ties on (Gift Aid) get independent quantiles: no two quantile columns share an order (|Spearman| < 0.1 over 2,000 people in one cell).
12. On a built eFRS, SPI-synthetic employees' pay rises with hours at least half as strongly as FRS employees' does. It is skipped without a build; it passed on CI's TESTING build.

test_imputation_source_flags.py: impute_income ranks once, on the full weighted FRS, and passes those same quantiles to both draws.

test_spi_income_earnings_groups.py gains the main-source split: SPI and FRS rows with pay and a trade fall in the employee-main group exactly by the rule above.

Mutation check. These are the eight mutants from review r1, run against the targeted tests:

Mutant (review r1's numbering)	At df3a633	At c134a6b
M1: midpoint of the slice instead of a uniform point	survived	killed
M2: ties ordered by id instead of at random	survived	killed
M3: one seed for every output	survived	killed
M4: equal weights instead of household weights	survived	killed
M6: national ranks instead of cell ranks	survived	killed
M7: the copy's draw without the full-FRS quantiles	survived	killed
M9: every output at the first output's quantile	survived	killed
M10: value order reversed	killed	killed

Script and log: mutate_r1_fixes.py / .log (local evidence folder).

Tests. Six targeted files on the locked microimpute 1.8.1 (test_spi_income_rank_draw, test_spi_income_earnings_groups, test_imputation_source_flags, test_income_imputation_preserves_housing_costs, test_spi_build, test_uc_gainful_self_employment): 168 passed and 8 skipped. Every skip needs a built dataset, and the differential test runs. ruff format --check . reports 243 files formatted. CI at 6c26644: 2,217 passed, 3 skipped and 1 xfailed, with test_built_enhanced_frs_spi_pay_rises_with_hours passing on the TESTING build. CI at c134a6b: green (Test passed in 47m16s). At 5079ea6, the comment-only change below: the six targeted files give 168 passed and 8 skipped, and the differential test runs. CI at 5079ea6 is green: 2,226 passed, 3 skipped and 1 xfailed, and test_built_enhanced_frs_spi_pay_rises_with_hours passes on the TESTING build. At 5621c0d (main merged), seven targeted files (the six above plus test_salary_sacrifice_imputation) give 195 passed and 8 skipped. Every skip needs a built dataset, and the differential test runs. ruff format --check . reports 247 files formatted. At a017674 (main merged again, adding #597), eight targeted files (the seven above plus test_clone_and_assign) give 218 passed and 9 skipped. Every skip needs a built dataset, including #597's new test_built_dataset_sources_are_original_survey_households, and the differential test runs. ruff format --check . reports 247 files formatted, and the changelog fragment check passes. CI at a017674 is green: 2,383 passed, 4 skipped and 1 xfailed. On the TESTING build, test_built_enhanced_frs_spi_pay_rises_with_hours and #597's test_built_dataset_sources_are_original_survey_households both pass.

Not in this PR

• SPI draws are in 2022-23 money and are not uprated to the 2024-25 FRS year before stacking. This is pre-existing; Rebase SPI income draws to the FRS survey year #532 addresses it.
• microcosm#840's other bridge predictors (industry, State Pension age) are not included.
• Per-component ranks plus the chain's conditioning overstate the dependence between incomes (table above). This PR measures it but does not correct it.
• Follow-up: train the forests on the SPI age band rather than a noise age within it. Everyone in a rank cell would then face the same conditional distribution, and the rank draw would reproduce it exactly.
• Follow-up: the SPI employee group carries part-year and marginal pay that FRS employees do not. That puts very low hourly pay on SPI-synthetic part-timers.
• Not changed: draw_quantiles(dataset) and the FRS-half impute_over_incomes(dataset, …) each build the model inputs, so a Microsimulation is built twice. That costs seconds in a build that takes tens of minutes. The FRS-half call is also due to go with Keep FRS-reported dividends and key them on person_id #498.

Review

• r1 (Opus, independent; subfleet job 20261008-171857-ukdata-rank-draw-review): REQUEST CHANGES at df3a633. It confirmed the core: rank_quantiles is unbiased, draw_at_quantiles matches microimpute 1.8.1 exactly, and the person-ID wiring is right. Its findings and the answers:
  1. CI failed at df3a633 (Set UC gainful self-employment from the FRS main-job status #525's test). Fixed in 6c26644; CI green there.
  2. 7 of 8 mutants survived the tests. New tests (invariants 2, 5, 7, 10, 11 and the impute_income wiring test) kill all eight; see the mutation check above.
  3. "SPI distribution preserved" went too far. The claim is now qualified in code, DESIGN and this body, and the dependence is measured above. The band-trained forest is listed as a follow-up.
  4. The offline tail result (b) is the forest and its resample, not a rank bias. Explained in DESIGN.
  5. The hourly-pay shift is disclosed above, with numbers. The group mismatch is a follow-up.
  6. Keep FRS-reported dividends and key them on person_id #498 and Rebase SPI income draws to the FRS survey year #532 were misdescribed. Corrected under "Base and overlaps".
  7. The aggregate changes are disclosed above.
  8. "Percentiles of each cell" now reads "of the forest's conditional distribution".
  9. An unknown model type now raises.
  10. Duplicate ids now raise.
  11. The duplicate Microsimulation build is kept, with the reason under "Not in this PR".
  12. The MAINSRCE claim now cites the SPI documentation. The "impact table" claim now waits for the real builds.
• r2 (delta review of c134a6b; subfleet job 20261008-215049-ukdata-rank-draw-review-r2): REQUEST CHANGES at c134a6b, for written evidence only. Code and tests pass: 168 passed and 8 skipped, the differential test runs, all eight of r1's mutants are killed, and a mixed-type prediction probe matches microimpute exactly. It marked r1's findings 1, 2, 5, 6, 7, 9, 10 and 12 addressed, 11 an acceptable deferral, and 3, 4 and 8 partly addressed. Its findings and the answers:
  1. DESIGN labelled £100k tail figures as £150k. Relabelled. The £150k figures are now shown separately, here and in DESIGN: 0.306m at FRS ranks against 0.205m from the SPI cells.
  2. The 0.802m/0.845m/0.848m figures sum the three pay groups, not EMPLOYEE alone. Relabelled; the EMPLOYEE figures (0.780m against 0.827m) are added.
  3. Nits: the cells are now described as the SPI's resolution of the forests' conditioning, not their exact conditioning. The forests also take exact age and the earlier outputs. This is fixed in the income.py comment (5079ea6, comment only), in DESIGN and in this body. DESIGN's cutoffs now read as percentiles of the forest's conditional distribution.
• r3 (delta review of 5079ea6, GPT-6.1 Sol; subfleet job 20261009-103633-ukdata-rank-draw-review-r3): APPROVE WITH NITS at 5079ea6. All three r2 findings are addressed. It confirmed every new figure against r1's JSON, confirmed the 14 changed code lines are comments (identical Python AST), and checked the new mechanism statements in the code. Its two nits, both on written evidence, are fixed in this body and in DESIGN:
  1. The £150k table's 0.274m is microimpute's own draw on this PR's forests, drawn on the full FRS and then subset to the copy. It is a controlled comparison, not base's own result, and only the EMPLOYEE, NO_EARNINGS and SELF_EMPLOYED forests are shared with base. Relabelled.
  2. DESIGN's 0.521m forest expectation uses the eval's resamples with split forests for people with both pay and a trade, so it compares with the eval's split variant (0.5728m), not the unsplit 0.573m. Relabelled, with the three forests that sum to it.
• r4 (delta review of the merge of main, 5621c0d, GPT-6.1 Sol; subfleet job 20261014-204105-ukdata-rank-draw-review-r4): APPROVE at 5621c0d, with no findings. It confirmed the merge keeps both sides (the auto-merged test_spi_build.py holds both the SCOT_TXP and MAINSRCE tests), that the PR's diff against the new main is unchanged, and that the typed changelog fragment passes Require a typed changelog fragment from each PR, and remove the 32 untyped fragments #555's check. It reran the differential test and the UC suite (106 passed, 2 dataset skips). It asked this body to disclose the salary sacrifice and capital gains donor interactions above. It also corrected one premise: FRS regions can be UNKNOWN (frs.py maps code 14 and missing codes to it). The merge changes neither that mapping nor the rank cells.
• r5 (delta review of the merge of main a43e8c5, a017674, GPT-6.1 Sol; subfleet job 20261015-151205-ukdata-rank-draw-review-r5): APPROVE at a017674, with no findings. Both merge comparisons match: 5621c0db..a0176749 equals main's daad71cd..a43e8c51, and the PR's diff against a43e8c5 equals the approved diff against daad71c. The auto-merged test_imputation_source_flags.py keeps both sides' tests. The rank draw completes before stacking adds provenance. It reran the differential test, test_imputation_source_flags.py and test_clone_and_assign.py (27 passed, 1 dataset skip). It corrected one premise: long_geography.py's source-ID logic produces area-support diagnostics and does not itself suppress outputs or alter local weights. The interactions paragraph above now says so.
• r6 and r6b (reviews of the calibration-fit decomposition, not of code; subfleet jobs 20261016-100120-ukdata-rank-draw-fit-review-r6, GPT-6.1 Sol, and 20261016-101919-ukdata-rank-draw-fit-review-r6b, Opus 5.5 while the Codex lanes were closed): POST WITH EDITS, both applied before posting. r6 reran every script byte for byte and reproduced W3's −2.440328pp with Compare each pull request's calibration fit with its base build's #599's own fit_report. It had the transient reading stated as unproven, and corrected three claims: some same-base rebuilds pass the 2pp allowance (2 of 20), the saved-weight objective is slightly worse on W3's pair, and both workflows use one OA clone. r6b fixed a five-versus-six-build label and wording. r6 recommends keeping the hold until release-settings builds exist.

axiom: n/a: data-construction change (SPI income imputation), no policy rule

🤖 Generated with Claude Code

[MaxGhenis]

Fit measurement for the batch review (W3 no-worsening check, #599, applied by hand):

I compared this PR's TESTING build (run 38661397028, head 5621c0d) with the build of the 1.58.1 release branch (run 38228500493), using each run's final-epoch constituency_calibration_log.csv.

• Share of local targets within 10%: 79.1% → 76.7% (−2.4pp).
• Share of national targets within 10%: +0.6pp.
• National targets whose error rises by more than 0.2: 9.

For scale, five builds of identical data code moved these shares by at most 0.69pp (local) and 0.93pp (national). The basis is in #599. #599 would fail this PR on the local drop. Worth checking which local target families lose fit before this goes into a batch.

[MaxGhenis]

Decomposition of the local-fit drop measured above, from the same public CI logs. Aggregates only: shares of targets within 10%, and sums over constituencies.

Net, all of the −2.44pp comes from one target family: constituency self-employment income amounts.

• Across the 11,228 shared local targets, this family contributes −2.49pp and the remaining targets contribute +0.05pp. The remaining targets' own within-10% share rises by +0.06pp, within the largest same-code change seen (0.70pp).
• Its share within 10% is 0.809–0.825 on the five same-code builds of main (4cbedbe data code, Compare each pull request's calibration fit with its base build's #599's basis) and 0.766 on the 1.58.1 release branch's build, against 0.335–0.366 on all six of this PR's builds.
• The miss is in England. Across the 543 English constituencies, the share is 0.762–0.829 on those six builds of main and 0.252–0.282 on this PR's six. Of the 280 targets lost in W3's pair, 277 are English. Scotland and Wales are about unchanged.
• Across the 20 pairs of this PR's four builds on 4cbedbe against those five, the local change is −1.74 to −2.81pp, of which this family accounts for −2.57 to −2.80pp.

The epoch-30 drop may be an early-training transient; the branch's converged fit is unknown. A TESTING build calibrates for 32 epochs and keeps the weights logged at epoch 30. In utils/calibrate.py, Adam runs at lr 0.1 on log weights and saves them every 10th epoch. The table shows the share of constituency self-employment income amounts within 10%, with sum(estimate)/sum(target) over all constituencies in brackets:

Build	Epoch 10	Epoch 20	Epoch 30	Later
main 4cbedbe, TESTING (5 builds)	0.188–0.203 (0.984–0.986)	0.474–0.491 (0.996–0.998)	0.809–0.825 (0.997–1.000)
this PR on 4cbedbe, TESTING (4 builds)	0.309–0.320 (0.968–0.969)	0.523–0.540 (0.972–0.973)	0.342–0.366 (1.029–1.031)
main 1.58.1 release build, 512 epochs	0.295 (0.972)	0.618 (0.964)	0.475 (1.006)	0.838 at 40, 0.975 at 50, 0.971–0.998 from 60 to 510
main 1.58.2 release build, 512 epochs	0.294 (0.973)	0.645 (0.961)	0.489 (1.007)	0.846 at 40, 0.982 at 50, 0.975–0.998 from 60 to 510

• This PR against main's TESTING builds. This PR's builds fit the family better at epochs 10 and 20, then overshoot between 20 and 30.
• National bands. Over the same ten epochs, the share of the 13 national self-employment income amount bands within 10% rises from 0.308 to 0.615–0.692 on this PR's 4cbedbe builds. On main's builds it goes from 0.615–0.769 to 0.692.
• Main's release builds. They make a similar trade between epochs 20 and 30: the national amount bands go from 0.538 and 0.692 to 0.846, while local amounts fall to 0.475 and 0.489. Both recover by epoch 50 while keeping 0.846 of the amount bands. That is consistent with a temporary trade-off between the national band targets and the local amounts during training. It does not show that this PR would recover.
• Local authority calibration. Main's TESTING builds overshoot this family at epoch 20 (sum 1.037–1.039) and recover by 30 (1.008–1.013). This PR's 4cbedbe builds overshoot at 30 (1.040–1.043).
• Loss. The logged stochastic training losses are similar, but they use different weights from the saved fit. On W3's pair, the saved-weight objective is slightly worse (0.078571 to 0.078940), while the national share within 10% improves by 0.62pp. Neither shows that this family is fitted as well.

What this doesn't show: whether this PR's fit recovers in a 512-epoch build. Main's release builds recover, but at epoch 30 their sum ratio is 1.006–1.007, against 1.029–1.039 across this PR's six builds. They also fit the income model on a weighted resample of about 100,000 SPI records, where TESTING uses about 10,000. No build of this PR at release settings exists yet. The seeded production builds queued for it will measure this: base, branch and two placebos, one at a time on the build host. Until they run, the PR stays a draft and out of the batch.

For #599: the reported W3 pair fails its matched TESTING check. Across the 20 same-base comparisons above, local-share changes range from −2.81 to −1.74pp, so some pass and some fail its 2pp allowance. A matched TESTING rebuild tests reproducibility under that protocol. Separate matched release-settings builds test whether the regression persists after further calibration. These logs alone don't justify changing the gate.