Correct SPI income inputs and UK calibration measurements

[juaristi22]

This ports the reviewed SPI income corrections onto the UK publication stack requested in Max's handoff, and corrects England geography and FRS composition measurements that obscured the comparison. National VOA targets previously counted other countries against England facts; generic UC family categories treated some older children as partners.

The current PR head is 1f7d676f2449894bfd794973eaa507aa79bfbfdb, rebased onto main at 289437765a2d5caca97f2974f4d104e9e1e066d5, including merged PRs #874 and #870. The existing PR commits are preserved, with all three durability fixes in one additional commit: immutable report links in the four register annotations, a reachable source-receipt revision in the evidence builder, and explicit historical-register hash documentation. The earlier #874 integration brought 659adb1a fiscal-year target selection, DfT geography and take-up-contract fixes. The retained calibrations were not rerun for this head; the numerical results below describe their original pinned measurement and spine versions.

With the same measurement changes applied to both retained full spines, SPI reduces national loss 39.8%, and targets within 10% rise from 337/371 to 341/371. Both runs remain release-blocked. Interest accounts for 91.5% of the net objective gain; aggregate income-tax and state-pension fit worsen. This conditional comparison does not establish complete administrative measurement correctness or release readiness.

Changes

• Rebase supported SPI monetary inputs from 2022 to the 2024 build period, correct OTHERINV indexing and FRS tax-free-interest accounting, and condition pension draws on receipt. Preserve the under-16 input guard and query stream. Six flows remain explicitly nominal, including the SPI state-pension leaf despite an available engine index; its separate uprating sensitivity remains unmeasured.
• Restrict national VOA A–H and total to country == ENGLAND, matching the selected E92000001 facts within the broader England-and-Wales VOA publication. Local VOA and Scottish bindings remain separate; the household-count proxy for dwelling stock is unchanged.
• Use retained FRS head/parent roles for UC calibration composition, including cohabiting couples and older children. The child-count measures include nonclaimant reported members under 20 or native UC qualifying-child status.
• Restore optional-engine test boundaries: nine SPI-income cases require UK; parsed/path-equivalence coverage uses deterministic uprating stubs without an engine. Add an independent full-registry synthetic UC matrix regression and a reusable authenticated evidence builder.
• Preserve all target values, 44 measure exclusions, reviewed fit-exclusion decisions, take-up, solver settings, gates, CGT algorithms and retained numerical results. The review repair does not change production runtime or rerun calibration.

Matched national comparison

The complete 2024 control and SPI spines originated at d43b4203c6ebe10e062cb3ef3034e66731ea055d and fc49b48200e1e15fe35bb4e15b948992bf03c28d. Measurement code 117860efe96b4fc1e336e0c9bfab10917cd3aa59 calibrates 2025 using UK 2.94.0/Core 3.31.0, Chronicle 6fb700e, 371 identical active targets, 1,500 epochs, family-equal weighting, learning rate 0.02, maximum weight ratio 10 and seed 0.

Metric	Corrected control	Corrected SPI
Final loss	0.021724747	0.013076866
Targets within 10%	337 / 371	341 / 371
England Band A error	+0.049%	−0.027%
England total dwelling-stock error	+0.036%	+0.010%
Income-tax error	−12.067%	−12.709%
Total state-pension error	−7.909%	−11.430%
UC single-with-children error	−31.365%	−27.923%
UC total benefit-unit error	−11.189%	−10.006%
UC childless couples, £27.6–28.8k error	−100%	−100%
UC childless couples, highest included band error	−100%	−100%
Effective sample size	6,127.6	6,110.6

The formerly decisive childless-couple cell depended on two duplicates of one source case, now classified as lone parent in both runs. Awards remain unchanged by remeasurement. Corrected childless-couple support is empty from £26.4k upwards in both spines, and adjacent bands have only one to four unique source cases.

Unresolved measurement, support and entitlement work

The registry contains 100 included payment-band references: 25 per family, 84 active and 16 excluded. DWP publishes 26 positive-payment categories for the relevant period, including a separate £2,500.01-or-over monthly tail. This register omits that tail. Its next-lower-edge convention gives the first band [£0.12, £1,200.12), and the highest included, bounded-source band [£28,800.12, infinity). A £36,000 annual award therefore enters a row labelled £2,400.01–£2,500 monthly. Catalog and currency-boundary repair belongs with #736 and requires new matched measurements; saved-matrix equality does not establish published-boundary correctness.

DWP family classification uses awarded standard allowance, including single-rate cases with ineligible partners. Reported children can differ from child-element eligibility. FRS relationship counts remain a proxy. The FRS 2024–25 glossary confirms cohabiting benefit units and dependent 16–19-year-olds, who are excluded from its adult definition; its survey-head designation does not establish UC eligibility. Numeric role codes and adult/child file routing remain unverified. The native standard-allowance relationship error and partner eligibility need separate review; calibration relabelling repairs neither entitlement question.

UC reporter redraw still uses legacy is_married/qualifying-young-person predictors, its has_non_child_member screen and transition categories. That can classify a young lone claimant as child_only while calibration uses SINGLE, and can disagree for cohabiting couples. The redraw implementation/settings and existing generated support stay fixed in this comparison. Aligning redraw is unresolved support-generation work.

Both runs fail the unchanged target-fit gate and refuse calibrated H5 export. SPI's unreviewed failures are the two empty high-payment UC bands, self-employment at £50–70k (+34.485%) and state pension at £20–30k (+27.941%). The four UC fit-deferral rationales cite historical #813 composition and all require fresh adjudication; the two-child entry is stale in SPI, and the private-pension-count entry is stale in both runs. Each of the four register reasons now preserves its original text verbatim and appends a dated historical-basis annotation linking the report; the two-child annotation explicitly records its stale treatment result. All signatures, dates and operative decisions remain unchanged, and no approval is renewed. The receipt’s protected_policy_resource_sha256["target_fit_reviewed_exclusions.json"] records the unannotated register used by the historical runs; it is not a checksum of the current annotated register. The other historical receipt hashes also remain the pins of the executed comparisons. No new input-mass parity waiver, known-gap approval or release authorization is supplied. Gate-override observability remains a separate suggestion.

Evidence and validation

The historical pre-solve script asserted nine VOA rows against independently derived England masks and 84 UC rows against runtime-resolved families. It did not serialize individual UC exact-match flags. The new builder authenticates the original receipts, retained preflight/matrix/solve artifacts and spine hashes, then adds 84 explicit saved-matrix row audits per run. The private composition snapshots were first hash-pinned during this repair and remain saved resolver output, not independent historical family evidence. Public output contains only allowlisted aggregates and hashes. The independent synthetic test covers all included boundaries, excluded-edge retention and benefit-unit-to-household sums, with the open-last-band defect marked as known behavior.

Historical validation on 7 September with UK 2.94.0/Core 3.31.0 installed covered 305 passes/two skips for the measurement work, then 458 passes/four skips for a later focused rebase set. Historical Ruff, inventory and pinned-feed regeneration checks passed. These are different scopes, neither a full-repository or certification claim.

Review-repair focused checks on 7 September, before the second rebase: the SPI files pass on engine-free Python 3.13 and 3.14 (39 passes/14 skips each), and with the existing UK engine (52 passes/one skip). The independent UC file passes five tests; the builder passes 15 tests, including artifact-tamper refusal, exact-row and privacy checks. Both retained 84-row audits complete without new calibration. The PR targets main, so hosted CI applies; current hosted results should be read from the checks panel.

Historical validation after the 7 September rebase onto main base 5ab1b056: the same 13 affected test files pass without country engines on Python 3.13 and 3.14 (588 passes/19 skips in each environment). Five affected files pass with the existing UK 2.94.0/Core 3.31.0 environment (81 passes/two skips). These scopes overlap; the counts are not additive. The earlier 113-file fast group was interrupted for the user-requested main update and is not a full-suite pass. The completed checks cover code contracts, not a recalibration or dataset certification.

Current durability-repair validation at 1f7d676f2449894bfd794973eaa507aa79bfbfdb on main base 289437765a2d5caca97f2974f4d104e9e1e066d5: 264 tests passed across the corrected-measurement receipt, signed-deferral register, gate-battery pin and data-contract checks; the saved-evidence builder replay matches the updated receipt byte for byte. Source ancestry and blob hashes, exact receipt-locator and report-link preservation checks, Ruff, formatting, diff and CI inventory checks all pass. This validates the repair and integration; it does not rerun calibration or certify a dataset.

• Corrected comparison and limitations
• 371 paired targets, authenticated matrix evidence and UC support
• UC target inventory and source-boundary caveats
• Reproducible retained-evidence builder
• Historical publication comparison and earlier native validation

Refs #665, #736, #796, #809, #840, #866 and #862. The joint-fit problem remains open; CGT tail work in #878 and timing reconciliation in #875 remain separate.

[juaristi22]

Completed the matched national-only rebuild and calibration comparison requested in the migration handoff. Calibration loss falls 35.8%, but 98.8% of the net reduction comes from the ONS savings-interest target. The joint income-tax/state-pension/UC calibration problem remains unresolved.

The control uses publication commit d43b4203; the treatment uses executed SPI source commit fc49b482. The subsequent evidence commit is 4e2e3c25. Both runs rebuild the full 2024 spine and calibrate to 2025 national targets, using PolicyEngine-UK 2.94.0/Core 3.31.0, identical raw inputs and Chronicle feed (6fb700e), and the same 371 target bindings, exclusions and loss weights. Both use 1,500 epochs, family-equal weighting, seed 0 and a maximum weight ratio of 10. Existing take-up holds and gate policy are unchanged.

Errors below are relative to the target: positive means above target, negative means below.

Metric	Publication control	SPI treatment
Final calibration loss	0.023464462	0.015068903
Targets within 10%	334 / 371 (90.0%)	338 / 371 (91.1%)
OBR income-tax receipts error	−12.151%	−12.562%
OBR total state-pension error	−8.136%	−11.695%
State-pension amount, £20–30k income band	+25.323%	+25.211%
State-pension amount, £50–70k income band	+34.563%	+16.867%
Private-pension recipient count, £100–150k income band	+24.716%	+14.256%
ONS savings-interest error	−18.249%	+0.007%
UC households: single with children	−35.309%	−34.525%
OBR CGT receipts error	+44.180%	+43.744%
Effective sample size	5,868.4	5,758.9

Across all 371 targets, absolute error improves for 181, worsens for 189, and is unchanged for one. Nine targets enter the 10% band and five leave it. The £50–70k state-pension band and private-pension count improve, while aggregate income-tax and state-pension fit deteriorate. These results evaluate the combined SPI treatment; they do not isolate the pension-receipt predictor from income rebasing, the age guard or the OTHERINV correction.

The treatment clears the £50–70k state-pension gate failure but introduces three non-deferred target-fit failures:

• UC couples without children, £27,600–28,800 annual payment: −100%. The treatment estimate is zero before and after calibration, while all calibrated household weights remain positive. For this nonnegative count measure, the aggregate evidence implies an empty realized support cell rather than zero-weight collapse; it does not identify which SPI change removed that support.
• Self-employment income, £50–70k income band: +35.413%.
• Council-tax Band A (VOA): +28.182%.

The £20–30k state-pension failure remains. The private-pension count exclusion is stale in both runs because its error is already within the 25% gate bound; it was preserved for the matched comparison. A22/A23 holds and existing UC/CGT deferrals also remain unchanged.

Both spines complete their build gates, but both national calibrations fail the terminal target-fit gates, so neither exports a calibrated dataset. Household count and total prior weight are equal, but the treatment generates 36 fewer people and 21 fewer benefit units, with different final household support and prior-weight vectors. This is a national-only, single-seed comparison and does not establish local-area accuracy or a release-ready UK dataset.

The full comparison and reproduction instructions and 371-target paired receipt include target/family loss contributions, source/runtime pins and gate verdicts. This publication-baseline comparison is separate from the historical bc11803f / UK 2.92.1 results.

[juaristi22]

Completed the matched national recalibration after correcting the national VOA geography and UC family measurements. This supersedes the interpretation of the earlier comparison.

Both runs use the original full spines, UK 2.94.0/Core 3.31.0, the same 371 targets, 1,500 epochs and unchanged solver settings, exclusions and gates. Before optimization, all nine VOA matrix rows matched England-only masks and all 84 active UC band rows matched the corrected family categories. Every underlying UC award and all unrelated matrix rows are unchanged.

Metric	Corrected control	Corrected SPI
Loss	0.021724747	0.013076866
Targets within 10%	337/371	341/371
England Band A error	+0.049%	−0.027%
England total dwelling-stock error	+0.036%	+0.010%
Income-tax error	−12.067%	−12.709%
Total state-pension error	−7.909%	−11.430%
UC single-with-children error	−31.365%	−27.923%
UC childless couples, £27.6–28.8k	−100%	−100%
UC childless couples, highest band	−100%	−100%

The geography correction resolves the apparent Band A problem: the earlier +22.80% / +28.18% estimates mixed Scotland and Wales into an England benchmark. All nine corrected VOA measures now lie within 0.22%.

The UC correction exposes an underlying support gap. The former two-row high-payment cell was a duplicated lone-parent case. After relationship-aware classification, both corrected spines have no childless-couple support above £26.4k. Several adjacent bands have only one to four unique source cases. The native UK model's standard-allowance relationship error is separate and remains unresolved by this measurement change.

SPI reduces loss 39.8%, with interest accounting for 91.5% of the net gain. Income tax and aggregate state pension still worsen, so this is not uniform improvement. Both runs remain release-blocked: treatment additionally misses self-employment income at £50–70k (+34.485%) and state-pension income at £20–30k (+27.941%), and the unchanged gate flags stale reviewed exclusions. No calibrated dataset was exported.

UC inventory: 100 payment-band references = 25 × four family types; 84 active and 16 excluded. With 11 general UC and 15 two-child-limit targets, there are 110 active UC-related targets. The ordinary bands are £100/month wide (£1,200/year). The highest published band is currently materialized as open-ended, an existing mismatch held fixed here.

Full corrected comparison · all 371 paired targets and 100-band support audit · complete UC target inventory.

Validation: 305 distinct focused tests pass, two existing skips; Ruff, CI inventory and pinned-feed reference regeneration pass. PR remains draft.

[vahid-ahmadi]

Review pass at 5a1581e3 (Claude Code, high effort; worktree checkout, scoped to the in-scope diff 202a66bc..5a1581e3 — the four uk-publication-stack-834 register commits and the five #879 commits; the ten #874 commits underneath are reviewed on #874 and not here. Without the licensed engine: ruff clean on all 18 in-scope Python files; both census tools, the coverage manifest and ci_test_groups --verify current; 541 passed, 5 skipped across the in-scope suites except test_uk_spi_income.py, on which see finding 2).

What checks out

Every number in the PR body reconciles exactly to experiments/receipts/uk-corrected-measurement-comparison.json: loss 0.021724747 → 0.013076866 (−39.8%), within-10% 337 → 341 of 371, England band A +0.049% / −0.027%, income tax −12.067% / −12.709%, state pension −7.909% / −11.430%, UC single-with-children −31.365% / −27.923%, ESS 6,127.6 / 6,110.6; 91.5% of the net gain is the single ons.savings_interest_income@2025 row; all 66 changed matrix rows are VOA or UC. The body's admission that treatment worsens income tax and state pension is honest and matches the receipt. The VOA fix is a declared household_conditions entry country == ENGLAND on the nine national targets at household grain, the Scottish rows keep region == SCOTLAND, and test_uk_council_tax_targets.py:36-102 derives the England mask independently in-tree. The under-16 guard is applied consistently at the stage-1 draw, the stage-2 target, the base redraw and the landing, with the RNG stream preserved and tested. Tax-free interest does not double count (200 rebased + 5 unrebased = 205 in the test). OTHERINV now rides the engine's own obr.per_capita.gdp index like dividends and property rather than being held nominal. The exclusion clock is fail-closed: a required artifact, refused unless a plain date, derived by the certifier with no CLI override, and the inclusive approved_on..expires_on window is shared machinery. The SPI uprating map is a constant that the spec is asserted against, the six nominal flows are listed and pinned by test, and the parity fixture and coverage manifest moved only their manifest shas.

Findings

1. Should-fix — the calibration's UC family typing diverges from the redraw stage's. The new measurement (measure_simulation.py:37-89) defines claimants as head, parent, or any member when no dependent children are recorded; children as non-claimants who are engine qualifying young persons or under 20; couple as two claimants. The #835 redraw stage (uc_reporter_redraw.py:675-684, untouched) still types couple by is_married and children by the engine QYP flag alone, and uses is_married as a stratification predictor. So the stochastic stage draws reporters toward one family mix and the calibration measures another; a lone 16–19 qualifying young person is child_only for the redraw and SINGLE for the measure (the new test fixture's benunit 40 is exactly that case). The body calls the change a "calibration composition proxy" but does not say the redraw keeps the old typing. Either align the redraw's family typing to the shared definition, or state the divergence and why it is tolerable for this comparison.

2. Should-fix — requires_uk markers dropped, so tests fail rather than skip without the engine. The base had @pytest.mark.requires_uk on the typed-weights SPI test; the diff removes it, and none of the five new impute_uk_spi_income_support tests carry it, yet the path reaches _disability_parameters (spi_income.py:1266), which imports policyengine_uk unconditionally. Here test_uk_spi_income.py is 9 failed / 24 passed / 5 skipped. Hosted CI is unaffected because the engine-uk lane installs the extra and the file routes there, but the body's "305 passing, 2 skips" is engine-only. Restore the markers, or stub _disability_parameters the way the rebase test already stubs the uprating factors.

3. Should-fix — four register entries are adjudicated on the measurement this PR replaces. target_fit_reviewed_exclusions.json's four dwp.uc.households_*@2025 reasons cite "honest grain after the #813 re-bind", the family-type and child-count measurement that uc_calibration_* now supersedes; the receipt already shows households_children_2 stale in treatment and single_with_children moved from −35.3% to −27.9%. Holding them fixed for the matched experiment is right; merging them with unchanged reasons ships deferrals whose stated basis is gone. Renew or annotate before the 2026-09-30 clock.

4. Should-fix — "remains draft" is false. The PR is not a draft (isDraft: false); "release-blocked" is accurate (both runs fail uk_target_fit and refuse calibrated export). Convert to draft or drop the word.

5. Question — the UC matrix-row check is asserted, not evidenced. pre_solve_matrix_checks in the receipt carries the nine VOA England-mask checks and all_uc_awards_unchanged: true, but no per-row check for the 84 UC rows the body says were verified against corrected family masks, and the committed experiments/build_879_publication_receipt.py is the older receipt's builder with no England-mask or UC-row code. VOA is covered in-tree by the council-tax test; UC is not.

6. Question — spine-origin pins are indirect. The corrected receipt never names d43b4203 / fc49b482; it records input identity shas that are equal before, after and copied, and those shas appear in the earlier uk-publication-spi-comparison.json, which does pin the commits. The chain holds; the corrected receipt should carry the commits directly.

Smaller questions

• State retirement pension is held nominal (hmrc_spi_state_pension_income: None in SPI_INCOME_UPRATING_VARIABLES), so 2022-23 amounts sit in a 2024 build across two triple-lock upratings. It is not the mechanical driver of the state-pension worsening — state_pension_reported is only read for the bridge — but the leaf feeds the auxiliaries the hmrc_spi band targets read. Why nominal when the engine has an explicit state-pension index?
• is_married = famtypb2.isin([5, 7]) in frs_spine.py:430 (pre-existing, out of scope) is why the PR routes around is_married. If that coding is wrong against the FRS dictionary, it is the root of both the "older child as partner" symptom and finding 1.
• _resolve_target_fit_exclusions returns a source that _evaluate_target_fit discards, consistent with the sibling gates, so no gate detail records that an override was in force.
• The body says 305 passing / 2 skipped; the rebase note in 879-corrected-calibration-comparison.md says 458 / 4. Different sets, presumably; worth one line.
• No signed-difference or known-gaps register is touched, yet the under-16 guard and the rebase both depart from the incumbent. Is the UK spine: composition-faithful SPI income draw — bridge predictors and rank-preserving replacement (with-children UC residual) #840 source review the register of record for that?

Nit

uk_population_targets.json re-escapes non-ASCII (the accented characters become \u00ed escapes); the semantic diff is empty, serializer noise only.

Out of scope and unverifiable

#874's content (the engine floor and lock, the Chronicle re-pin, the childcare and bus targets) is reviewed on #874. Reproducing either calibration, the requires_uk uprating-index test, and the engine's family-type and QYP semantics need the licensed engine and data.

---

Nothing blocks on correctness, and the receipts are the most honest part of the PR. 1 is the one to settle before this is used as the basis for anything downstream, since two stages now disagree about what a family is; 2 through 4 are cheap. "Release-blocked" is the right framing; "draft" is not.

[vahid-ahmadi]

Round one is in the comment above (#879 (comment)): nothing blocking on correctness, four should-fixes (UC family typing diverges from the #835 redraw; requires_uk markers dropped, which is also what the two failing spine-uk CI lanes look like; four register reasons cite the superseded measurement; the PR is not a draft). Filing as a comment review rather than request-changes since the receipts are honest and the PR declares itself release-blocked.

[juaristi22]

@vahid-ahmadi Thanks for the review. I've pushed the review fixes and rebased onto main at 5ab1b056. The current PR head is eb7ca8a5. Here's the disposition of each point:

1. UC redraw and calibration classification: I've documented their divergence explicitly. Redraw still uses the legacy marriage/child predictors and eligibility screen, while calibration uses retained FRS roles and the revised child definition. Both retained spines used the same redraw implementation, so the comparison is conditional on that implementation. Aligning redraw and entitlement would change generated support and awards and needs a separate controlled experiment; this PR does not establish that alignment.

2. Missing UK-engine test dependency: This affected hosted CI too: both fast Python lanes had 10 failures. Nine SPI-income cases now carry requires_uk; the SPI-spine path-equivalence test instead stubs the uprating dependency so it retains engine-free coverage. Tests using the actual UK engine remain covered separately.

3. Historical UC deferrals: I've appended review annotations directly to the four affected register entries, preserving their original reasons and identifying the superseded measurement basis. The two-child entry explicitly records that it is inside the bound in the corrected treatment. These require fresh adjudication. Approvers, approval/expiry dates, operative decisions and thresholds are unchanged; the annotations do not renew approval or authorize additional exclusions.

4. PR status: Removed the stale draft-status wording from the description. The PR is ready for review; the dataset remains release-blocked.

5. UC matrix evidence: Added a reusable evidence builder and 84 exact saved-matrix row checks per run to the corrected receipt. The receipt distinguishes historical pre-solve checks from this subsequent audit; composition-file hashes first recorded during review are labelled accordingly. A separate synthetic regression uses independently specified relationship scenarios and covers all 100 included references, 84 active/16 excluded, age and payment boundaries, and benefit-unit-to-household aggregation.

6. Spine provenance: The receipt now explicitly links the control origin (d43b4203) and treatment origin (fc49b482) to their source receipts and matching spine hashes. All original receipt fields and numerical results are preserved. Replaying the builder after the rebase produced a byte-identical receipt.

The updated report also incorporates Max's measurement-review points. All nine national VOA band/total bindings match England. FRS roles support the composition proxy, but do not establish UC partner eligibility or fix the pinned model's separate standard-allowance relationship error.

The UC boundary issue remains open under #736: DWP publishes 26 positive-payment bands per family, including £2,500.01-or-over monthly, whereas the register includes 25. Consequently, the highest included band is bounded in the source but currently absorbs the omitted open tail. Fixing endpoint semantics and handling that separate published tail requires a new matched calibration. The regression explicitly records the current mismatch; it does not certify it as correct.

For the smaller points, the report acknowledges that a state-pension uprating index exists but was not selected for this treatment, with its sensitivity unmeasured. Exact FRS numeric-code validation and gate-override provenance remain follow-ups. The #840 evidence is not a new parity waiver or release approval. I left the unrelated Unicode serialization cleanup untouched.

Post-rebase validation: 588 passed / 19 skipped on each of Python 3.13 and 3.14 across the affected engine-free tests, and 81 passed / 2 skipped with the pinned UK engine. These are overlapping test scopes. Ruff, formatting, diff checks and CI test inventory checks passed. Hosted CI is still running at posting time, with no completed failures. No calibration was rerun for these review fixes; the reported comparison retains its original code and artifact pins, and the fit gates remain unchanged.

[vahid-ahmadi]

Second pass at eb7ca8a5 (Claude Code, high effort; worktree checkout against 478fa88a, the byte-identical rebased replay of the reviewed head per git range-diff; the interdiff is ten files, all within the six dispositions; 577 passed, 19 skipped, 0 failed across 14 in-scope suites without the engine; ruff check clean and ruff format clean on the changed files; both census tools, the coverage manifest and ci_test_groups --verify current).

1. Documented, not aligned — accurate and adequately placed. uc_reporter_redraw.py is untouched; the divergence is stated in the report's "Redraw and nominal-income limitations" section, in the PR body, and in each of the four register annotations, and the statement matches the code. Since #883 is where the redraw and capital-donor predictors move to the shared claimant helper, I am content with this being a declared limitation of the comparison here.

2. Verified. Nine requires_uk markers on the SPI-income tests; the SPI-spine path-equivalence test stubs _spi_income_uprating_factors with deterministic factors and still exercises the full stage transform — on the old head it fails engine-free with ModuleNotFoundError, so the stub is load-bearing. Engine-free run: 39 passed, 14 skipped, 0 failed.

3. Verified. Only the four reason strings changed, each preserving the original text as a prefix; approvers, adjudication ids, approval and expiry dates unchanged. The two-child entry records −21.5% in treatment, inside the bound, and that the annotation renews nothing.

4. Verified. Not a draft; no draft wording in the body.

5. Verified as far as it can be here. The receipt carries 84 exact saved-matrix row checks per run, all exact_saved_matrix_match: true, with the pre-solve and audit scopes distinguished; build_879_corrected_measurement_receipt.py is committed and fails closed on missing artifacts; the tamper and exact-row tests pass; the new full-registry synthetic test covers 100 references, 84 active and 16 excluded, with the open-last-band defect pinned as known. The byte-identical replay needs the restricted artifacts.

6. Verified. Both runs carry spine_origin_code_pin (d43b4203, fc49b482) with the source receipt path, revision, sha and artifact sha.

Three durability nits. The register annotation links point at a branch blob rather than a commit sha; the builder's SOURCE_COMMIT is the pre-rebase head, which is not in main's history; and the receipt's protected-resource sha for the register predates the annotations (consistent with "historical pins preserved", but one sentence in the report would stop a reader treating it as drift).

---

Approving. Release-blocked stays the right description of the dataset; the code and receipts are honest about it.