Re-measure the hmrc SPI band exclusions: three 'sparse cell' premises fail on spine-i (hundreds of in-band carriers) #807

Summary

Measured support inside every excluded hmrc/* SPI band cell on the current candidate-of-record spine (spine-i, sha 0cce0399…), prompted by the wealth-top-tail investigation (chronicle#205 context): the spine's wealth tail is full and WAS-faithful, the historic "£1m cap" was a capital-gains artifact fixed by the HMRC Table 3 stage — so the top-tail exclusion premises deserved re-measurement rather than inheritance.

Method: for each excluded cell, unweighted records and weighted mass of carriers (value_variable > 0) with total income in-band, under two income definitions — hmrc_spi_assessable_income (the SPI stage's own concept) and a component-sum proxy matching the model's total_income (ITA 2007 s.23, which excludes capital gains). Script: uk-candidate-eval/scripts/measure_excluded_cell_support.py; receipt: out/excluded_cell_support_spine_i.json.

Findings, by class

A. The 11 1_000_000_to_inf cells — premise HOLDS, wording should be tightened

Total-income support above £1m is 0–2 records on spine-i, even though the spine now carries capital gains to £20.4m and 6,925 records above £1m wealth. The CGT band donors carry large gains but not £1m+ incomes, and total_income excludes gains, so these exclusions stand.

Request: tighten each reason to say "total income (ITA 2007 s.23, excl. capital gains)" explicitly and cite the measured support. The previous drift of "no gains over ~£1m" into "no wealth over £1m" shows how a loose phrase becomes a false premise downstream.

B. Three mid-band "sparse cell" exclusions — premise appears FALSE, propose lifting

Hundreds of in-band carriers is not "too few records for the solver". A plausible history: these were adjudicated pre-#803, when excluding a sibling band silently widened its neighbour (the #792 movers, up to +858%) — i.e. the adjudication may have been reacting to band-edge corruption rather than real sparsity. Their v9 fits were overshoots (+27% to +34%), not support gaps, which also contradicts the sparse-support reason.

Request: re-measure post-#803 and lift if the fits land; the calibration should get these three back.

C. Three thin 500k–1m cells — defensible, revisit with the next spine

dividend_income_income_band_500_000_to_1_000_000 (6–21 records), property_income_count_500_000_to_1_000_000 (2–12), self_employment_income_count_500_000_to_1_000_000 (4–19). Genuinely thin, but not zero; worth re-measuring on the next spine build (the was_wealth pension split landed on main this morning; a fresh spine is building) before renewing them at expiry.

Not touched

The 16 dwp/uc_payment_dist exclusions (uk-data#452) and the rest of the register are out of scope here — though note the v10 calibration measured UC caseload regressing to −25.6% (from −14.5%) with the 16 band exclusions in place and unchanged UC support, i.e. those exclusions removed the upward gradient on UC weights. Worth weighing when uk-data#452's M2 (the #755 pension split) gets its first calibrated measurement.

Approver context: the register entries carry approved_by: juaristi22, expiries 2026-11-25/26.