[MaxGhenis]

Pilot of option 1 of the UK pensioner calibration review (approved by Max, 2026-10-27). Result: the pilot does not win, so this PR should not merge. It fits the four rows it binds, but on every out-of-sample test it is worse than main and further from the re-index that #1414 merged for #1069. Closed unmerged on that result; the branch is kept, so reopen it if Max rules to adopt option 1 (decision d2111).

A/B result (aggregates only; full tables in docs/evidence/uk-spi-dated-bands/README.md)

Three national builds on one FRS 2024-25 spine, weights solved at 2025 over 1,273 targets: baseline (main 7d45c0c), pilot (this branch plus a "PILOT BUILD ONLY" hold fill, local 71856c6fa), and #1414's candidate.

Test	Baseline	Pilot	#1414 candidate
Taxpaying State Pension recipients, 2025-26 (HMRC implied 8.65m–8.81m)	8.84m	9.26m (+5.1% above)	8.74m (inside)
Non-taxpaying recipients (HMRC implied 3.58m–3.74m)	3.53m	3.13m (−12.6% below)	3.63m (inside)
ITL Table 2.1, UK taxpayers at State Pension age (9.005m)	−0.81%	+3.92%	not run
ITL Table 2.2 residents at State Pension age / 65+ (5% tolerance)	+1.03% / +1.28%	+5.84% / +5.68% (misses)	in sample
ITL misses in 41 calendar-2025 cells	0	3	not run
Dated check: Table 2.5 2023-24 bands missed	6 of 11	6 of 11	not run
SPI pension rows, total error on the 2023-24 definition	−2.0% to −5.0%	−0.02% to +1.7%	not run
SPI pension rows, total error on the 2025 definition	−0.9% to +6.7%	+3.4% to +11.3%	not run
Sole-income SP allowance (#182), 2027-28: affected, cost	1.235m, £84m	1.117m, £75m	1.326m, £93m
Final calibration loss (in sample)	0.00553	0.00535

• The SPI rows and the calibration loss are in sample, and each build is closer on the definition it binds. Neither ranks them.
• The dated check misses the same six bands on both builds, so it doesn't separate them. It does show that the carried-back file misses HMRC's 2023-24 distribution above £50,000.
• Why the pilot loses is not isolated. The candidates are the 2025 weights, the national age-sex population factor, the engine's carry-back indices, the State Pension's carry-back, and HMRC's projection.
• Allowance figures come from policyengine.py 6.11.1 / policyengine-uk 2.151.0, dashboard fa9c42b1.

What changes

The four SPI 2023-24 Table 3.6 band rows (hmrc.spi.{state_pension,private_pension_income}.{count,amount}_by_total_income_band) stop being projected to calendar 2025 by HMRC's all-taxpayer growth per nominal band. Instead they bind the published 2023-24 values at 2023-24, following the CGT dated-measure pattern (cgt_2024_*):

• New dated measures spi_2023_{total_income,income_tax,state_pension,private_pension_income,person_count} (uk_runtime/spi_dated_measures.py, resolver routes in uk_runtime/measure_simulation.py).
  • Carry-back. The engine's source file is carried back to 2023-24 by the engine's own forward rules:
    • savings interest gross of ISA interest;
    • every uprating_indices.yaml input divided by its year-on-year growth;
    • residential stocks by the cumulative house price ratio;
    • year-specific inputs dropped;
    • inputs moved to their _before_lsr/_reported bases, as Simulation.__init__ moves them.
  • Round trip. The engine's own extend_single_year_dataset must return the carried-back file to the stored one for every input column. The resolver refuses otherwise.
  • Measurement. The annual-income simulation runs at 2023. The State Pension pays at 2023-24 rates.
  • Weights. They stay solved at 2025.
• Population. Counts and amounts carry each person's population factor: ONS's population of their sex and age band at 1 October 2023 over mid-2025, using the same bands as the ons.population.* targets.
• Materialization guard. A row may now declare a measurement period that is neither the calibration period nor the frame's period, but only when every variable of the row is dated to that year. This holds whether or not the frame carries period metadata.
  • "Every variable" is one inventory of executable reads (_binding_reads): value terms, each filter's variable or concept on its own entity, and the band.
  • This applies to any measure dated by its name, so a CGT row dated throughout (cgt_<y>_*) may now declare its year too. That is intended.
  • Without frame period metadata, a row whose value terms read an engine-computed variable that no name dates is refused under a declared non-calibration period. "Engine-computed" means not in the frame's own stored_columns; a column the resolver injected does not count as stored.
• Resolver refusals. The resolver refuses three kinds of row:
  • a row that names a spi_<y>_* measure alongside a variable of another year (values, filters, band and value_expression terms);
  • a dated SPI measure bound through a provider kind, named in any field of the binding;
  • any row measured at a year other than the calibration year that also carries an uprating_index, because that would move the fact twice.

Invariants (property-tested in tests/engine_free/uk/test_uk_spi_dated_measures.py)

1. No target-side projection on these rows. None of the four rows, or their 52 references, carries an uprating_index. Each is bound at measurement_period 2023 with require_matching_fact_period, and every variable of each row is spi_2023_*.
2. Band partition. Over the published edges, band counts and amounts sum to the unbanded total above £12,570. No one is dropped at an edge or counted twice. This is tested with Hypothesis, including values exactly on an edge.
3. Carry-back round trip. Indexed inputs divided by the factors and then carried forward by the index return the stored values. Held columns are untouched. Inputs are not mutated.
   • A value missing on one side only is a miss; missing on both sides matches.
   • student_loan_repayments is compared on the rows where the engine kept the loan plan, and must come back exactly zero where it reassigned it, as uprate_student_loan_plans sets it.
   • Bereavement Support Payment is held when both separated bereavement inputs are stored, as the engine holds it; council tax benefit is held by declaration (the engine moves it with its household's council tax, and no dated measure reads it).
4. Population factor identity. Where the weights meet ONS's mid-2025 band populations, a band's weighted factor sums to its 1 October 2023 population. The factor is 1 when the population is unchanged, and it is positive.
5. Determinism. The measures and the differential test read only pinned, vendored facts and the engine. There is no randomness.

ITL differential (uk_runtime/itl_differential.py, tools/check_uk_itl_projections.py)

The ITL July 2026 projections become a test of a calibrated file, not targets:

• Cells checked: Table 2.5 taxpayers by total-income band, Table 2.1 UK taxpayers of State Pension age and 65+, and Table 2.2 by area, all at the calendar-2025 window (3/12 of 2024-25, 9/12 of 2025-26).
• Tolerance: 5% national, 10% per area.
• Report: misses by table, band and area, with each cell flagged as bound by a target (in sample) or not.
  • With --diagnostics, "bound" means bound in that run, cell by cell: a diagnostics row (named by band or area) with a missing or positive loss weight.
  • A sum is in sample only when every part is, and nobody the sum counts is outside its parts (outside_parts is zero). The sums are: the all-band count over its bands, the residents' count over its areas, Table 2.1's UK age cell over its twelve areas, and the count below State Pension age over that cell and the all-band count.
  • A diagnostics row binds a cell only at its own period (target_name and period, or the name's trailing @year).
  • Families report in_sample_cells.
  • Only the UK rows of Tables 2.5 and 2.1, and the region-tier rows of Table 2.2, are read.
• Dated check: --dated-year 2023 checks the file's own 2023-24 taxpayers by band, on the spi_2023_* measures, against Table 2.5's 2023-24 outturn.
• Engine test: tests/engine/uk/test_uk_itl_differential_build.py runs this on a built file when MICROCOSM_UK_ITL_DIFFERENTIAL_H5 is set.

Invariants, property-tested in tests/engine_free/uk/test_uk_itl_differential.py:

6. Window weights. The window is a convex combination of the two tax years (weights sum to 1), and it is invariant to row order.
   • Duplicate rows and half windows are refused, among the rows a cell reads.
   • Table 2.5 bands outside ITL_BANDS are refused.
   • The vendored bands sum to the all-band count within 0.5% in every tax year.
7. Partition. The band cells, with the taxpayers below the lowest edge (£12,570), partition all taxpayers. The area cells partition residents. State Pension age and below it sum to all taxpayers. This holds for any person order.
8. Sums in sample. Whenever a sum is in sample, its model count equals its in-sample parts' (the sum, or the difference), on random people. A family is in sample exactly when every cell is.
9. Verdict. A cell misses if and only if |model/expected − 1| exceeds its tolerance.

Hold

uk/uprating_holds.json carries the four families as a proposed matched_period hold: unsigned, status proposed. The compile refuses a target held at an older period without a signed hold (microcosm#1123), so test_every_committed_hold_is_signed fails on exactly these four until María or Max signs. That is the intended blocker. A new test, test_every_reference_measured_at_an_earlier_year_is_declared, requires a declaration for every national reference measured at an earlier year with no index; it catches the failure that stopped the first pilot national.

The A/B pilot national was built from a local branch, never pushed, that fills the hold's approval fields with an explicit "PILOT BUILD ONLY, not a register signature" label.

Review

• Round 1 (49660f1, Opus 5.5 via Subfleet): request changes. Answered in 87a4729, 1471ebf and 2aa2607.
• Round 2 (2aa2607, GPT-6.1 Sol via Subfleet): request changes.
  • P1: Bereavement Support Payment and council tax benefit failed the round trip on the real spine, which the engine probe also found. Fixed in 058690b.
  • P2s (concept filters, provider fields, metadata-less frames, per-cell ITL samples), the student-loan limitation and the nits: fixed in 073e9b5.
  • 073e9b5 also lists uk_runtime.spi_dated_measures among the UK target compile modules, since ledger_targets imports it.
• Round 3 (073e9b5, GPT-6.1 Sol via Subfleet): request changes, with four P2s, three P3s and one plausible risk. All are answered in 58411eb:
  • provider rows have their value reads checked against the declared period (_provider_value_reads, _read_honours);
  • ITL sums carry their parts and outside_parts;
  • diagnostics bind only at their own period;
  • annotation fields (metric_name, notes, source_lines) are skipped by the dated-measure scan;
  • families report in_sample_cells;
  • the hold preflight reads executable bindings and treats a null index as none;
  • entity counts hold no period on either frame;
  • round-trip labels compare by value (same_labels).
• Round 4: not dispatched. The pilot lost the A/B, so this PR is not heading to merge. Any piece split out to land on its own gets its own review.

Status

• ITL July 2026 differential, with a stated tolerance.
• Write-up on extending option 1 to every SPI band family and to the US SOI band rows (docs/evidence/uk-spi-dated-bands/README.md).
• Engine round trip on the real 2024-25 spine.
• Baseline vs pilot A/B on one spine, run with real policyengine.py runs for the allowance (above). The pilot does not win.
• A signed hold: moot unless Max adopts option 1 (decision d2111).

axiom: n/a: microsim calibration-target measurement change, no policy encoded

🤖 Generated with Claude Code

[MaxGhenis]

Closing unmerged: the option 1 pilot lost the A/B.

What happened. The pilot fits the four SPI pension band rows at 2023-24, but on the out-of-sample tests it is worse than main:

• It puts taxpaying State Pension recipients at 9.26m, against HMRC's implied 8.65m–8.81m.
• It puts non-taxpaying recipients at 3.13m, against 3.58m–3.74m.
• It misses three ITL Table 2.2 cells that main passes.

#1414's re-index (merged) is inside both ranges, so it stands.

Where things are.

• Head 58411eb answers the round-3 review (GPT-6.1 Sol). It also records the A/B, aggregates only, in docs/evidence/uk-spi-dated-bands/README.md.
• Locally, the CI engine-free plan ran 21,364 passed and 5 failed. Four failures are the orrery tests, which need the npm @axiom-foundation/orrery package. One is test_every_committed_hold_is_signed, which is the intended blocker on the unsigned proposed hold.
• Round 4 was not dispatched.

Methodology call. Whether to bind any SPI or SOI band family at its own year is queued for Max as d2111, with a recommendation of no. If he rules to adopt option 1, reopen this PR. The branch is kept.

Reusable pieces. Two parts could land on their own, each proposed separately:

• the ITL differential check (uk_runtime/itl_differential.py, tools/check_uk_itl_projections.py);
• a period check on provider reads. On main, provider bindings skip the period check.