UK local surface before the Budget build: high-value homes OBR leg, per-vehicle fuel flag, Northern Ireland energy share and housekeeping (#1278)

[juaristi22]

Second PR for #1278: the items that are actionable now and small enough to land before the frozen pre-Budget build.

Base.

• Rebased. This branch is rebased onto main after UK local release: publish a K=15, 150,000-household dense dataset in at most two builds (#1301) #1307's merge (ff3f348) and targets main.
• UK local release: publish a K=15, 150,000-household dense dataset in at most two builds (#1301) #1307 conflicts. Four conflicts with UK local release: publish a K=15, 150,000-household dense dataset in at most two builds (#1301) #1307, resolved as its re-review described:
  • the census family keeps both high-value homes sentences, regenerated;
  • the calibrated node takes {**gate_params, "phase": "terminal", "terminal_code": uk_terminal_only_code_hash()};
  • both sides' terminal tests stay.
• Receipt and membership. They auto-merged to the union (27,249 receipt rows), and a regeneration on the combined tree reproduced them byte for byte.
• Item I. Re-pin the UK Chronicle feed to abf0921 (#1300) #1302 left Northern Ireland's energy share to a change stacked on it (María's ruling of 2026-10-16); item I is that change.
• The reviews.
  • The first review of 17 October is addressed in R1–R7, and the J1 and D messages are reworded.
  • The second is addressed in R8–R10.
• UK local release: publish a K=15, 150,000-household dense dataset in at most two builds (#1301) #1307's remaining fixes. The remaining fixes from the third review of UK local release: publish a K=15, 150,000-household dense dataset in at most two builds (#1301) #1307 ride here too (listed below).

Part of #1278.

What changes (one commit each, then the review fixes and #1307's)

A. The high-value homes benchmarks gate reads the OBR costing facts (bug).

• Problem. The gate's OBR leg looked for an older row shape (observed_measure.source_concept, layout.record_set_spec_id), so it reported awaiting_vendored_facts on every build, and the derived-stock level question (Place high-value homes where the Land Registry sees them (#1240) #1270 review, finding 4) had no OBR comparison.
• Row selection. The leg now selects rows by concept, England, fiscal-year period type, year and basis, and needs every band and the total once each. Each count must be finite and non-negative. Otherwise it reports no_matching_rows, incomplete_rows or malformed_rows, and a missing or non-numeric value, or a field that is not a mapping, is reported rather than raised, since the leg runs after the solve.
• The share. The non-owner-occupied share (40 per cent) comes from the note's own row, and the leg names the year it is stated for.
• What it reports. The owner-occupied sensitivity (99,000 for 2028-29, static), the frame's England total over it, and the derived stock's private-sector count over Table 1.1's total.
  • No owner-occupied ratio. The stock build sets the owner-occupied share from the same OBR share, so that ratio would only repeat this one.
• Disclosure. The shipped band counts withhold the total when it rests on fewer than ten distinct source households. They withhold further bands until the withheld ones rest together on ten, so the total cannot give back their sum. The regional shares and their total follow the same rule.
• Unchanged. The leg stays informational: pass or fail does not depend on it.
• Cache. The calibrated gate node keys on the high-value homes code through its parameters (terminal_code), hashed once per process, so a resumed store recomputes the evidence. The full-gates kernel's hash, which UK local release: publish a K=15, 150,000-household dense dataset in at most two builds (#1301) #1307's preflight kernel inherits, leaves that code out, because the dense solve keys on the preflight.
  • What that buys. An edit to high_value_home_gates.py never re-solves. high_value_homes.py also feeds the target kernels, so an edit there still re-solves: a disclosure fix between two builds that must reuse the solve belongs in high_value_home_gates.py alone.

B. The high-value homes evidence names its scope.

• Scope. Owner-occupiers' main residences only, until the coverage decision (UK: data fixes to model the High Value Council Tax Surcharge (mansion tax) correctly, and the related housing and council-tax gaps #1240 fix 6, tracked here since UK: data fixes to model the High Value Council Tax Surcharge (mansion tax) correctly, and the related housing and council-tax gaps #1240 closed into UK local surface, consumption and high-value homes: remaining items (consolidation) #1278). It is stated in the evidence summary, both gates' details and the census family text.
• Unchanged. Target labels and the gates.json notes.

C. Registers track #1278.

• What moves. Three consumption-tax measure exclusions named UK consumption-tax levels: open items after #1167 #1207, and 19 uprating holds named UK local build (K=25, 2026-10-06): data-quality defects on the target surface and the spine #1123 (16) or UK consumption-tax levels: open items after #1167 #1207 (3). Both issues closed into UK local surface, consumption and high-value homes: remaining items (consolidation) #1278. Two hold reasons that still named UK local build (K=25, 2026-10-06): data-quality defects on the target surface and the spine #1123 now point here too.
• Unchanged. Approvals, windows and adjudications, so nothing is re-signed (María's ruling of 2026-10-16).
• The SLC exclusions. The two slc.borrowers.plan_2_* exclusions move from closed UK spine: SLC plan-2 borrower stocks at a third of SLC's — the FRS flag captures repayers, not the liable stock #868 to UK local surface, consumption and high-value homes: remaining items (consolidation) #1278 too (María's ruling of 2026-10-17), and their adjudications record the move. The plan-2 and plan-5 uprating hold keeps UK spine: SLC plan-2 borrower stocks at a third of SLC's — the FRS flag captures repayers, not the liable stock #868.

G. The disclosure floor's study conditions are recorded. This answers #796's open #609 question.

• Conditions. No study the build reads sets a minimum above CD171's ten:
  • the FRS 2024-25 (SN 9563) adds nothing to the End User Licence;
  • the SPI 2022-23 tape (SN 9422) adds two conditions: no record attribution, and analyses go to HMRC ten working days before distribution;
  • LCFS, ETB, WAS and NTS add nothing.
• Where. The note lives in uk_runtime.receipt_counts.
• The baselines tool. The weighted-integrity baselines tool no longer asks about a threshold of 30, anywhere: its docstring, its default's comment, the --sdc-minimum-count help and the caller obligations it writes into every receipt. Those obligations now carry the SPI tape's two conditions.

J1. Zero-emission cars are drawn per vehicle.

• The draw. Each of a household's vehicles is zero-emission at the VEH1103 battery-electric and fuel-cell share. A household with n vehicles is therefore a fuel buyer at 1 - (1 - rate) ** n, against the same identity-keyed uniform. One-vehicle households keep their flag.
• Rule name. rate_rule becomes per_car_zero_emission_share in the stage spec, the rendered roster and the sources schema.
• Receipt. The fuel-flag receipt adds two shares: the flagged share among vehicle households, and the share of vehicles in fuel-free households.
  • The flag is per household, so a zero-emission car beside a petrol or diesel one keeps its household's fuel, and the second share sits below the per-car share.
  • The shares on a side difference into the households with and without a vehicle, crossed with fuel or none. So every share on that side is "<10" when any non-empty group rests on one to nine distinct positive-weight sources: the frame's source households, the donor's diaries.
• Regenerated. The coverage manifest and the H2 spine fixture.
• Not yet. Who holds the cars by income stays random; a later PR conditions it.

E. The rowwise solve takes its review date.

• Change. prepare_uk_full_solve, the doctrine solve and the rotated holdout accept now and pass it to the binding review. Builds keep today's date.
• Why. Seventeen engine-free tests reviewed the committed register on today's date, so they would have failed from 2027-01-15. They now review on its latest approval (COMMITTED_BINDING_NOW), as a build would on that day. Nothing is asserted at import.
• New tests.
  • A solve and every holdout fold refuse a date past the windows, and all five folds receive the date.
  • test_no_committed_local_adjudication_is_about_to_lapse fails fourteen days before the earliest local expiry, as the uprating holds' clock does.
• Follow-up. The uk.full.problem kernel and the holdout kernel should take the build's review date.

H. Division tails. The LCFS stage's report-only division capture adds each division's top-1% and top-10% spend shares, on the frame and the donor.

• The floor. It counts the distinct source households that spend anything, since the frame stacks copies of one. A share is "<10" when one to nine of them sit:
  • in its top group;
  • outside it, since the division's total ships beside the shares;
  • or between it and a shipped smaller fraction's group.
• Copies. A frame that marks copies without naming their sources is refused.
• What it answers. The next build says whether the over-draw sits in the tail.

D. The parity register is re-pinned to uk-data a43e8c5 (from 8629dbb).

• New concerns:
  • Pension-age Housing Benefit, ported. The concern states the difference: uk-data applies the pension-age regulations, while Microcosm's rows read the deprecated eldest_adult_age. Re-binding goes with UK spine: housing benefit £2.7bn against OBR £12.1bn — pensioner and supported-accommodation HB missing after the UC migration #867.
  • Working-age Housing Benefit, routed to UK spine: housing benefit £2.7bn against OBR £12.1bn — pensioner and supported-accommodation HB missing after the UC migration #867.
  • Pension Credit, ported.
  • LFS employment, a blocked source.
• Moved citations. The five obr.py line citations follow the functions that moved.
• Corrected verdicts.
  • The Band D origin now names the housing/council_tax_net multiplier.
  • The council-tax outlier guard is a ported compile contract. It is the one classification corrected by hand; the re-extraction changes none.
• The local employment-income count. Its concern notes that uk-data#529 made it validation-only, outside the inventory's scope.
• Next. A second re-pin follows uk-data#482 and Declare target parity against the N regime: v9.2 feed, CD surface on #484.

I. Northern Ireland's energy share comes from DESNZ.

• Factor. Great Britain's part of the Energy Trends UK level is now its share of DESNZ's sub-national domestic consumption with Northern Ireland's row added (all domestic meters, 2024). So each factor is the UK level over the UK sub-national total.
• Departs from the plan. The plan of record subtracted Northern Ireland's row from the Energy Trends total. The sub-national statistics run about 9% (electricity) and 12% (gas) above Energy Trends, so subtracting would take Northern Ireland out at a higher basis than Great Britain's part. The share keeps both nations on one basis (María's ruling of 2026-10-17).
• Periods. Northern Ireland's electricity runs on financial years from April and its gas on Northern Ireland gas years, both labelled 2024. DESNZ calls them not directly comparable.
  • The basis text, the receipt and the signed desnz_subnational_consumption_basis adjudication and census fence say so. María signed off one added sentence on 2026-10-17, with dates unchanged.
  • The census lists DESNZ's two Northern Ireland publications as sources.
• Resource. desnz_local_energy_facts.json vendors the two Northern Ireland rows in place of the four LFS household totals; its 697 meter rows are unchanged.
• Retired. The desnz_northern_ireland_interim_share adjudication and its census fence.
• Receipts. The local compile-parity receipt's 694 kWh rows (348 electricity, 346 gas) move by one factor per fuel: 0.99901 for electricity, 1.01582 for gas. So do the same cells in the local membership report. Nothing else in either file moves.

Review fixes (on top of the item commits)

• R1, items A and B: suppression, OBR row selection, the dropped ratio, terminal_code, and the UK local surface, consumption and high-value homes: remaining items (consolidation) #1278 citations.
• R2, item C: two hold reasons, and the two SLC exclusions moved to UK local surface, consumption and high-value homes: remaining items (consolidation) #1278.
• R3, item G: the baselines tool.
• R4, items J1 and H: distinct sources and differencing in the tails, and the fuel-flag shares.
• R5, item E: no import-time assert, the lapse clock and the fold test.
• R6, item D: the Housing Benefit and employment-count notes, the TV-licence citation and the fragment.
• R7, item I: the receipt and membership regenerated, the fence sentence, the census sources and the docstring.

From the second review:

• R8, items J1 and H: the receipts' floors on groups and spenders, unnamed copies refused, and the stage wiring tested.
• R9, items A and B: every unreadable OBR row reported, terminal_code hashed once per process with its scope stated, and two comments corrected.
• R10, item D: the fragment's file count.

Remaining fixes from the third review of #1307

• N1. The two below-floor region rows are worded as an expectation from their published stock, not a count of frame records, in the small-cell rule and its fragment. The local membership, which copies the rule into 601 rationales, is regenerated; nothing else in it moves.

• L6. The re-pin procedure says to regenerate the receipts one --surface at a time with nothing else running, and when the local-selector cut used here holds.

• T1. The publisher's --preflight-only verdict and printed pointer carry the contract's warnings. The tag of a release published past them records how many, never their text. Nothing refuses.

• T2. The sign-off drafter accepts digits only in citations of a fixed list of repositories, refuses a citation next to a noun a figure counts or belongs to, accepts review anchors and file URLs, bounds dates to a year of the drafting date, and accepts digits only in listed logins.

• T3. The disclosure walk catches:

  • order statistics above a figure, figure maps and pandas split tables;
  • any field holding an area code, in any case;
  • numbers with thousands separators, and the missing figure names.

  The release pre-flight reads files as the contract does and walks the incumbent files too, and the tests now drive both ends.

• T4. The runbook says how to retry build 1, and the full-build doc says the contract warns.

Regenerated

• uk/uk_local_target_census.json (B, I).
• uk/release_input_coverage_manifest.json and the H2 spine fixture (J1).
• uk/uk_data_target_inventory.json and uk/uk_data_target_parity.json (D).
• uk/desnz_local_energy_facts.json (I).
• uk/ledger_compile_parity_local_incumbent_2025_signed_differences.json and uk/local_target_reference_membership.json (I), on abf0921. uk/local_target_references.json regenerates byte-identical.
  • How. Both generators hold the whole feed in memory, so they ran on a cut of the pinned artifact (digest checked against the pin): the constituency and local-authority rows matching a local contract selector.
  • Why the cut is complete. No local reference reads operands outside its own selector, so the cut holds every candidate. Every row other than the 694 energy cells came out byte-identical to Re-pin the UK Chronicle feed to abf0921 (#1300) #1302's full-feed regeneration.

Merge order with the open PRs

• Place UK private rent and BRMA at the assigned area; bind Scotland's rents at BRMA grain (#1262, #1090) #1305. It has dropped its energy commit; the retirement lives here.
  • Whichever lands second regenerates the receipt and membership on the combined tree.
  • The local-selector cut must be widened for Place UK private rent and BRMA at the assigned area; bind Scotland's rents at BRMA grain (#1262, #1090) #1305's BRMA references at statistical_scope (the re-pin procedure says how).
• Port the uk-data wave to policyengine-uk 2.134.4: stamp duty, Move to UC, national home values (sixth #1095 PR) #1308. It has dropped its OBR-leg commit, test included; this PR's tests cover the vendored rows.
  • It conflicts with this PR on the coverage manifest's source_manifest_sha256 and on uk_spine.json. Regenerate both after both land, under Port the uk-data wave to policyengine-uk 2.134.4: stamp duty, Move to UC, national home values (sixth #1095 PR) #1308's policyengine-uk 2.134.4 lock.
• UK local release: publish a K=15, 150,000-household dense dataset in at most two builds (#1301) #1307. Merged; its four conflicts are resolved above.
• Whichever of these lands second re-runs:
  • the coverage manifest;
  • tools/graph_uk_spine_fixture.py under the locked engine;
  • the census tool;
  • the local receipt and membership, if its change moves local targets.

Rulings owed

• SPI 9422. Whether frame outputs count as analyses of the tape, for the HMRC condition.
• uk-data#529 (on UK: track the defects uk-data is fixing that microcosm still has #1095). Whether Microcosm follows uk-data in making the local employment-income count validation-only.
• Before UK local release: publish a K=15, 150,000-household dense dataset in at most two builds (#1301) #1307's build 1. The disclosure question its third review raised about five more high-value homes region rows.
• By design. Marking SPI-dependent outputs pending the SN 9422 ruling, and a floor on the distinct donor diaries behind a frame top group.

Not run locally

• The feed-gated tests. They need the whole abf0921 feed in memory: the vendored-resource round trip, the English region council-tax sum, and test_committed_surfaces_regenerate_from_pinned_feed[local].
  • That last test regenerates the local references and membership from the whole feed. Here they came from the cut described under Regenerated. It reproduced Re-pin the UK Chronicle feed to abf0921 (#1300) #1302's whole-feed output on every untouched row and, after the rebase, the committed union byte for byte.
• The full compile on abf0921. It belongs to the frozen build.

Verification (local)

• Touched surfaces: 103 test files covering every touched surface, the engine lane and the release tooling included, with 4,921 tests.
  • Results. 4,903 passed, none failed and 18 were skipped.
  • Every skip is environment-gated: the pinned feed or national artifact absent (14), the licensed LCFS/ETB tabs, the NZ artifact, the uk-data tree and the rebuilt OA ladder.
  • The uk-data tree re-hash. Run separately with MICROCOSM_UK_DATA_TREE on a43e8c5: the parity file's 14 tests passed.
• Checks: tools/ci_test_plan.py verify and tools/graph_acceptance_burndown.py --verify are ok. ruff check and ruff format --check are clean on the 33 edited Python files.
• Generators: census_uk_local_targets.py --check, census_uk_data_target_parity.py --check and build_uk_release_input_coverage_manifest.py --check all report current.

Disclosure

Commit messages, changelog fragments and this body carry no survey-derived figures, only:

• published figures (OBR, DESNZ, UKDS conditions);
• counts of tests, register entries, rows and areas.

H's fragment names, without figures, the divisions that drew above their donor.

🤖 Generated with Claude Code

[juaristi22]

Independent adversarial review of this PR at cf3aae4

Run with Claude Code on 16–17 October 2026 at cf3aae4: seven independent review lenses (correctness, method, generated artifacts, tests, claims against the diff, disclosure, a second fresh-eyes pass), then one to three skeptics per finding who tried to refute it; (k/n) is how many agreed. No builds and no feed-gated tests were run. The head has not moved since the review.

Keep this in draft; it should not merge yet. Item by item, the code mostly holds up: the targeted tests pass, the generators report current, and the vendored Northern Ireland rows match the abf0921 feed byte for byte. Two things block it:

• Item I applies the 16 October Northern Ireland ruling with a different formula from Place UK private rent and BRMA at the assigned area; bind Scotland's rents at BRMA grain (#1262, #1090) #1305 and from the approved plan of record for UK local surface, consumption and high-value homes: remaining items (consolidation) #1278 PR 2.
• No PR planned before the frozen build regenerates the local compile-parity receipt, so the release-blocking local parity preflight will refuse the 20-21 October build.

Blocking

• Item I copies Place UK private rent and BRMA at the assigned area; bind Scotland's rents at BRMA grain (#1262, #1090) #1305's Northern Ireland energy change with a different formula, and departs from the approved plan without saying so. CONFIRMED (3/3; one skeptic rated it high, two medium). packages/microcosm-build/src/microcosm/build/uk_runtime/domestic_energy_targets.py:281-283.

  • The two formulas:
    • UK local surface before the Budget build: high-value homes OBR leg, per-vehicle fuel flag, Northern Ireland energy share and housekeeping (#1278) #1306 sets the GB level to UK level × GB/(GB+NI). That gives factors of 0.9155904 for electricity and 0.8898214 for gas.
    • Place UK private rent and BRMA at the assigned area; bind Scotland's rents at BRMA grain (#1262, #1090) #1305 (6928d08, same base 4c51843) sets it to UK level − NI. That gives 0.9132567 and 0.8886493.
  • Whichever wins moves every LA electricity cell by 0.26% and every gas cell by 0.13%. For example, Hartlepool E06000001 (a published DESNZ figure) is signed at 112,337,019.72 kWh in Place UK private rent and BRMA at the assigned area; bind Scotland's rents at BRMA grain (#1262, #1090) #1305's receipt and compiles to 112,624,082.16 under UK local surface before the Budget build: high-value homes OBR leg, per-vehicle fuel flag, Northern Ireland energy share and housekeeping (#1278) #1306's code.
  • The plan of record for UK local surface, consumption and high-value homes: remaining items (consolidation) #1278 PR 2 specifies subtraction, which is Place UK private rent and BRMA at the assigned area; bind Scotland's rents at BRMA grain (#1262, #1090) #1305's formula. The UK local surface before the Budget build: high-value homes OBR leg, per-vehicle fuel flag, Northern Ireland energy share and housekeeping (#1278) #1306 body (its Base paragraph and item I's Factor bullet) presents share-scaling as the ruling. No later ruling for share-scaling is on record, and neither PR body mentions the other.
  • A trial merge conflicts in domestic_energy_targets.py, test_uk_domestic_energy_targets.py, desnz_local_energy_facts.json, ledger_fact_vendor_selections.json and test_uk_local_rowwise.py, so git forces a choice in code. The receipt, the selector notes and the changelog do not conflict (see Cross-PR).
  • Fix: María rules between the two formulas.
    • For share-scaling: it stays inside DESNZ's sub-national basis. GB+NI sub-national runs about 9% (electricity) and 12% (gas) above Energy Trends, so subtraction removes 9-12% more NI than share-scaling does.
    • For subtraction: it is what the plan approved.
    • Then keep the retirement in one PR. Drop the other PR's energy commit whole (cf3aae4 or 6928d08) rather than resolving hunk by hunk, and state the overlap and the merge order in both bodies. If share-scaling stands, add a line to this body saying it departs from the plan, and why.

• Item I changes all 694 DESNZ kWh rows of the release-blocking local compile-parity receipt, but the body blames the stale receipt only on Re-pin the UK Chronicle feed to abf0921 (#1300) #1302/UK private rent at the assigned area (#1123 item 4, "PR B"): what is in place and the pieces still missing #1262. CONFIRMED (2/2; rated medium and low). packages/microcosm-build/src/microcosm/build/uk/ledger_compile_parity_local_incumbent_2025_signed_differences.json:29882 (Hartlepool electricity) and :34740 (gas).

  • What the receipt holds: each of its 348 electricity and 346 gas ledger_only rows is the raw DESNZ value times the retired LFS interim factor (0.916500968 / 0.875963200). The head now computes 0.915590 / 0.889821.
  • How the gate checks it: gates.py:500-544 compares ledger_value with rel_tol 0 and abs_tol 1e-9 (the default is set at gates.py:191-192). uk/gates.json:43-52 declares the gate release_blocking at preflight.
  • The body's "Not run locally" note on the receipt points to the "UK private rent at the assigned area (#1123 item 4, "PR B"): what is in place and the pieces still missing #1262 change". That is Place UK private rent and BRMA at the assigned area; bind Scotland's rents at BRMA grain (#1262, #1090) #1305, which is a sibling rather than stacked on UK local surface before the Budget build: high-value homes OBR leg, per-vehicle fuel flag, Northern Ireland energy share and housekeeping (#1278) #1306, and it regenerated the receipt on the other formula.
  • No test that runs without the feed catches this: test_uk_target_references.py:1487-1527 skips the local surface.
  • Combined with X-s01 (Cross-PR, high, 3/3), this blocks the frozen build.
  • Fix: say in the body that item I changes the 694 kWh rows. After the NI ruling, regenerate the receipt once with tools/build_uk_ledger_compile_parity_signed_differences.py --surface local on abf0921, at the head that will actually be built.

Should fix

• The local membership report still pins the retired LFS factor on 694 DESNZ holds. CONFIRMED (2/2 medium). packages/microcosm-build/src/microcosm/build/uk/local_target_reference_membership.json:147782 (E06000001 factor); its resolved_value is at :497033.

  • The report carries factor 0.916500968101422 on 348 electricity holds and 0.875963199869941 on 346 gas holds, with resolved values computed at those factors. The generator at head would write 0.915590 / 0.889821.
  • It was last regenerated in cb45632 (Re-pin the UK Chronicle feed to abf0921 (#1300) #1302) and is missing from the body's "Regenerated" list. test_uk_target_references.py:109-182 (test_committed_surfaces_regenerate_from_pinned_feed[local]) sha256-compares it, so that test will fail the first time it runs with the feed. CI has no feed.
  • No build numbers are wrong: the compile re-applies the uprating applier itself. local_target_references.json holds no values and does not change.
  • Fix: after the NI ruling, regenerate the membership in the same pass as the receipt and list it under "Regenerated".

• Item G says the baselines tool no longer asks about a threshold of 30, but only the docstring changed. CONFIRMED (2/2 medium).

  • The 30-case advice remains in three places in tools/measure_uk_weighted_integrity_baselines.py:
    • :327-330: caller_obligations, which is written into every receipt the tool produces.
    • :61-63: the constant comment.
    • :95-96: the --sdc-minimum-count help text.
  • c5a0a6f changed only docstring lines 23-28. The body (item G's "Tool" bullet), the commit message and changelog.d/1278-disclosure-floor-study-conditions.changed.md all say the tool no longer asks.
  • The receipt's caller_obligations also lack the SN 9422 HMRC condition that item G records elsewhere.
  • Fix: bring all three places in line with the docstring, and add the HMRC condition to caller_obligations.

• High-value homes band suppression never withholds the total, and it adds a complementary suppression only when exactly one band is suppressed. Item A makes the OBR leg live and repeats the total twice. CONFIRMED (2/2 medium). packages/microcosm-build/src/microcosm/build/uk_runtime/high_value_home_gates.py:127-154 (_suppress_small_cells, line 141); the new OBR fields are at :286-312 and the gate details at :680-692.

  • Synthetic probes show both gaps:
    • One band on 5 sources, the rest empty: the band ships as None, but the total (500.0) equals it.
    • Bands on 40/4/3/0 sources: the two thin bands are suppressed, and the total minus the shipped bands gives back their combined count, which rests on 7 sources.
  • frame_over_all_properties_ratio['total'] and frame_over_owner_occupied_sensitivity multiply back to the total, because their denominators are public OBR figures.
  • Regional shares use the same helper and sum to 1.
  • The terminal gate report is copied into the release dir as terminal_gates.json (tools/assemble_uk_release_dir.py:450).
  • The logic predates this PR (ee7f1bb, Place high-value homes where the Land Registry sees them (#1240) #1270). The test at tests/engine_free/uk/test_uk_high_value_home_gates.py:121-128 deliberately pins the leaky case. Whether a real England build hits it is unverified.
  • Fix:
    • Withhold the total when it rests on fewer than 10 distinct sources.
    • When the suppressed bands together rest on fewer than 10 distinct sources, also suppress the next-smallest band or the total.
    • Apply the same rule to regional shares, and flip the test to expect total=None.
    • Check the frozen build's high-value homes evidence before any gate report goes public.

Minor

Low:

• The re-pin to uk-data a43e8c5 misses uk-data#529's local validation-only switch. CONFIRMED (2/2, both low; parts refuted, see below). packages/microcosm-build/src/microcosm/build/uk/uk_data_target_parity.json:336-349 and data_target_parity.py:272-282.

  • uk-data#529 made local hmrc/employment_income/count validation-only. It did this in create_datasets.py, which is outside the inventory's fixed scope.
  • Microcosm still binds that count: the census marks it bound_in_code, under fence hmrc_spi_frame_model_proxy, which runs to 2027-01-14. The new national_ons_lfs_employment fence quotes the same part-year reason without mentioning uk-data#529.
  • The plan of record for UK local surface, consumption and high-value homes: remaining items (consolidation) #1278 PR 2 lists this as a ruling owed, but no public issue or PR body records it.
  • Fix: add a uk-data#529 note to local_contract.hmrc.employment_income.count and list the owed ruling in the body.

• The SPI tape's HMRC pre-distribution condition is recorded but nothing enforces or tracks it. CONFIRMED (2/2 low). packages/microcosm-build/src/microcosm/build/uk_runtime/receipt_counts.py:12-19.

  • git grep 'working days' finds only the docstring and the changelog fragment. No gate, marker, caller obligation, checklist item or follow-up refers to it.
  • Whether frame outputs (such as item H's tails) count as "analyses of the tape" is María's SPI 9422 ruling, still owed.
  • The gap predates the PR.
  • Fix: confirm the condition's wording, settle the SPI 9422 ruling, add the condition to caller_obligations, and list the ruling as a follow-up.

• The division-tail "<10" floor counts frame rows, not distinct sources. CONFIRMED (1/1). packages/microcosm-build/src/microcosm/build/uk_runtime/lcfs_consumption.py:795-822 (line 817).

  • high_value_home_support splits each selected owner household into 6 copies (high_value_home_support.py:75, 228-301), and spi_support_channel stacks synthetic copies.
  • A synthetic repro: 2 sources split into 6 copies each print an exact 0.5013 share, where the unsplit frame prints "<10".
  • No real-data UK path runs the stage on a small frame.
  • Fix: pass source_household_id into the tail and count distinct sources.

• The top-1% and top-10% shares can be differenced. CONFIRMED (1/1). lcfs_consumption.py:825-846.

  • A synthetic skewed-weights case recovers the exact weighted spend of the 2 households between the groups, and of a single household in a variant. Realistic lognormal weights leave gaps of 435-548 households.
  • Evidence runs copy division_capture whole, so both shares would be committed together.
  • Fix: withhold the top-10% share when the two group sizes differ by 1 to 9, and add a skewed-weights test.

• The OBR leg's row selection is under-specified. CONFIRMED (1/1). high_value_home_gates.py:175, :187-198, :201-226, :284, :302.

  • Silent cases:
    • A missing total gives status available with a sensitivity of 0.0.
    • Rows relabelled as calendar_year are accepted.
    • obr_period='2030' uses the FY2028 share without warning.
  • Cases that raise inside the terminal phase (graph_terminal.py:873-879), after the solve: per-year share rows, or an unknown England band.
  • The vendoring row counts and this PR's own tests catch most bad re-vendors before a build runs.
  • Fix: select by period type, geography and basis; require every band plus the total; report a status instead of raising.

• COMMITTED_BINDING_NOW asserts at import. CONFIRMED (1/1). test_support/microcosm_build/uk_local_rowwise.py:21-40.

  • A partial renewal breaks it: six entries renewed on 2027-01-05, with the high-value homes fence (expired 2026-12-31) left dormant. The real build review passes on that register, but the module raises AssertionError at import.
  • There is also an off-by-one: the assert uses a strict <, while an entry stays in force on its expiry date.
  • 7 test modules fail collection (test_uk_graph_terminal imports the helper inside one function). With no --continue-on-collection-errors, one collection error interrupts the whole pytest run.
  • Fix: compute the date from the families the fixtures actually bind, or make it a lazy fixture.

• Item E removes CI's only signal that the local adjudications are about to lapse. CONFIRMED (1/1). Same file, :40.

  • With the review clock forced to 2027-01-20, 66 tests in the five touched files still pass. A control run shows a build is refused on that date.
  • The DESNZ fences lapse 2026-12-15, the high-value homes fence 2026-12-31, and six others 2027-01-14. Only the engine-uk test_uk_full_target_graph.py still goes red, from 2027-01-15 and for the census fence only.
  • Fix: add test_no_committed_local_adjudication_is_about_to_lapse, following test_uk_uprating_holds.py:263.

• DESNZ's "not directly comparable" caveat for NI has no fence, and the family's sources leave out the NI publications. CONFIRMED (1/1). packages/microcosm-build/src/microcosm/build/uk_runtime/local_target_census.py:266-276; fence rules at :960-1021; stale note at :726-727.

  • Retiring the interim also removed the family's only NI fence. The three fences that remain (basis, meters, stage share) say nothing about NI, and require_adjudicated_uk_local_binding reviews only those (local_rowwise.py:813-840).
  • Fix: add an NI comparability sentence to the basis fence before its 2026-12-15 renewal, list the NI publications as family sources, and fix the electricity source note.

• The module docstring still describes the retired LFS interim. CONFIRMED (1/1). domestic_energy_targets.py:13-20.

  • The function docstrings at :193-199 and :246-251 are correct, so the module docstring contradicts them.
  • Fix: rewrite the module docstring, and update the comment at local_target_census.py:133.

• Commit D says "No existing concern's classification changes", but credibility_local_council_tax_outliers changes. CONFIRMED (1/1). changelog.d/1278-uk-data-parity-a43e8c5.changed.md:1; data_target_parity.py:591-597.

  • The guard moves from routed / validation_check to ported_local_declared / compile_contract. The same text lists it a few lines later as a corrected verdict, so it contradicts itself.
  • Fix: write "the re-extraction changes no classification; one verdict is corrected by hand".

• J1's commit message says "The expected zero-emission share of cars stays the VEH1103 share", which holds only for the hidden per-car draw. CONFIRMED (1/1; the company-car sub-claim is unsubstantiated). lcfs_consumption.py:2904-2907; fuel zeroing at :618-621.

  • The frame stores only a household fuel flag, so a zero-emission car in a mixed household keeps the household's fuel spend.
  • Synthetic run with a 56/34/10 split of 1/2/3+ cars: 2.24% of vehicle households are fuel-free and 1.49% of cars sit in them, against z = 3.85% (the old rule gave about 3.84%).
  • Fix: reword the message, and optionally record the fuel-free share of cars in the receipt.

• The parity register calls pension-age Housing Benefit "ported" and hides a definitional difference. CONFIRMED (1/1). data_target_parity.py:247-252 (line 251).

  • Microcosm's bindings filter only on eldest_adult_age >= 66. uk-data also excludes units on IS, income-based JSA or income-related ESA.
  • policyengine-uk has housing_benefit_pension_age_regulations_apply in both 2.132.6 and 2.134.5, and it encodes the uk-data rule. policyengine-uk marks eldest_adult_age as deprecated.
  • Fix: bind that variable, or state the difference in the concern's evidence.

• The owner-occupied ratio equals the private-sector ratio by construction. CONFIRMED (1/1). high_value_home_gates.py:323-337; tools/build_uk_high_value_home_stock.py:895-908.

  • The stock build scales the owner-occupied count to exactly 60% of the private-sector count, using the same OBR share. On the committed facts both ratios are 1.1709173, so the second one is not an independent check.
  • Fix: say so in the body, the changelog and a code comment, or drop the ratio.

• The body's Disclosure section says "Nothing is survey-derived", which is too broad. CONFIRMED (1/1).

  • changelog.d/1278-division-capture-tail-shares.added.md:1 and the 65dd810 message name which divisions drew above the LCFS donor. That is a qualitative, national survey result, not a breach.
  • Fix: write "no survey-derived figures".

• Item A's hash change ties the dense solve's cache key to the report-only high-value homes gates module. CONFIRMED (1/1). packages/microcosm-build/src/microcosm/build/uk_runtime/graph_terminal.py:717-747 (line 741).

  • The preflight node (:1045-1061) runs the same kernel and feeds the dense solve's key. A verifier reproduced it: a comment-only edit to high_value_home_gates.py changes UKFullGateKernel.implementation_hash.
  • Consequence: any later fix to that module during the build window (for example for 1306-05, the band suppression, or 1306-12, the OBR row selection) re-solves the dense calibration.
  • Fix: keep the high-value homes gates module out of the preflight key. In UK local release: publish a K=15, 150,000-household dense dataset in at most two builds (#1301) #1307's split, override the hash in UKFullPreflightGateKernel; otherwise use a terminal-only kernel ref. At minimum, say it in the body.

Nit:

• A parity citation points at the wrong lines. CONFIRMED (1/1). data_target_parity.py:221 (mirrored at uk_data_target_parity.json:1210) cites "obr.py:667-756 (_parse_tv_licence …)". At a43e8c5, _parse_tv_licence is at obr.py:620-664.
• Tests leave three paths untested. CONFIRMED (1/1).
  • Deleting the OBR geography filter (high_value_home_gates.py:179-180) or the concept filter (:172-173) leaves every test green.
  • The share refusals at :219-226 are never exercised.
  • test_uk_local_holdout.py:78-85 raises on fold 1, so later folds are never shown to receive now.
• The new fuel-flag share skips the small-cell floor. CONFIRMED (1/1). lcfs_consumption.py:2491-2493. It adds no new information: it equals flagged_share / with_vehicles_share, both already in the receipt raw.
• Two small text slips. CONFIRMED (1/1).
  • The body's Checks bullet says "22 edited Python files"; the diff edits 20.
  • uprating_holds.json:242 calls a ruling open under closed UK local build (K=25, 2026-10-06): data-quality defects on the target surface and the spine #1123, while the tracking field at :243 already says UK local surface, consumption and high-value homes: remaining items (consolidation) #1278. :286 has "(a UK local build (K=25, 2026-10-06): data-quality defects on the target surface and the spine #1123 follow-up)".
• Item B points the open coverage decision at closed UK: data fixes to model the High Value Council Tax Surcharge (mansion tax) correctly, and the related housing and council-tax gaps #1240. CONFIRMED (1/1).
  • high_value_home_gates.py:87, local_target_census.py:246 and the main-residence changelog fragment all cite "UK: data fixes to model the High Value Council Tax Surcharge (mansion tax) correctly, and the related housing and council-tax gaps #1240 fix 6". UK: data fixes to model the High Value Council Tax Surcharge (mansion tax) correctly, and the related housing and council-tax gaps #1240 closed into UK local surface, consumption and high-value homes: remaining items (consolidation) #1278 at 10:31Z on 16 October, before B was authored.
  • The two slc.borrowers.plan_2_* exclusions still track UK spine: SLC plan-2 borrower stocks at a third of SLC's — the FRS flag captures repayers, not the liable stock #868, which is closed. Commit C keeps UK spine: SLC plan-2 borrower stocks at a third of SLC's — the FRS flag captures repayers, not the liable stock #868 on the SLC hold deliberately, but nothing says whether the two exclusions should follow. That one is María's call.

Cross-PR

• X-s01: No PR planned before the frozen build regenerates the local compile-parity receipt on abf0921. CONFIRMED (3/3 high).

  • The receipt is read in two places:
    • The release-cut certifier: tools/certify_uk_release_cut.py:270-278 and release_certification.py:332-345.
    • The full/dense build preflight: full_gates.py:76-88, graph_terminal.py:1036-1060 and full_build_cli.py:1951-1966.
  • So Re-pin the UK Chronicle feed to abf0921 (#1300) #1302's line "Nothing reads it at build time" is false.
  • A verifier ran the gate's own _signed_difference_check_failures, with Re-pin the UK Chronicle feed to abf0921 (#1300) #1302's receipt as the signed side and Place UK private rent and BRMA at the assigned area; bind Scotland's rents at BRMA grain (#1262, #1090) #1305's abf0921 regeneration as the live side. It found 4,202 failures:
    • 3,507 changed UC values and 1 stale UC row, which come from Re-pin the UK Chronicle feed to abf0921 (#1300) #1302 alone;
    • 694 kWh rows, which come from the NI change.
  • UK local release: publish a K=15, 150,000-household dense dataset in at most two builds (#1301) #1307's receipt was regenerated on 735820c. It merges over Re-pin the UK Chronicle feed to abf0921 (#1300) #1302 without a conflict and gives the same 4,202.
  • The dense path refuses at preflight, before the solve, so "a third build" overstates the cost. The certifier refuses only after the national build has finished.
  • Fix:
    • Regenerate the receipt once, at the head that will actually be built, and land it in Re-pin the UK Chronicle feed to abf0921 (#1300) #1302 or in a small follow-up right before build 1.
    • Correct Re-pin the UK Chronicle feed to abf0921 (#1300) #1302's "nothing reads it" text.
    • Regenerate again after Place UK private rent and BRMA at the assigned area; bind Scotland's rents at BRMA grain (#1262, #1090) #1305 and UK local release: publish a K=15, 150,000-household dense dataset in at most two builds (#1301) #1307 land.

• X-m01 and X-s02: The NI formula, beyond 1306-01. CONFIRMED (2/2 medium each).

  • uk_population_targets.json merges automatically to UK local surface before the Budget build: high-value homes OBR leg, per-vehicle fuel flag, Northern Ireland energy share and housekeeping (#1278) #1306's selector-note wording, even if Place UK private rent and BRMA at the assigned area; bind Scotland's rents at BRMA grain (#1262, #1090) #1305's code wins.
  • Place UK private rent and BRMA at the assigned area; bind Scotland's rents at BRMA grain (#1262, #1090) #1305's own notes (:31688, :31727) still say "share of UK households until UK local energy, area classification and Welsh bus facts: DESNZ Northern Ireland and sub-local consumption, rural-urban lookups, BUS04 country rows (PolicyEngine/microcosm#1207) chronicle#356 lands".
  • Both changelog fragments survive the merge: 1278-northern-ireland-energy-share and 1262-uk-energy-northern-ireland-level.
  • X-s02's suggestion to keep UK local surface before the Budget build: high-value homes OBR leg, per-vehicle fuel flag, Northern Ireland energy share and housekeeping (#1278) #1306 because it "matches the 'reads the share' wording" does not hold. The approved plan specifies subtraction, and the ruling text on Re-pin the UK Chronicle feed to abf0921 (#1300) #1302 (local_binding_adjudications.json:76) fits either formula.
  • Fix: pick one formula and drop the other PR's energy commit and fragment. Make both selector notes describe the chosen formula. Keep each PR's vendor-selection label with its resource header, since test_committed_vendored_resources_carry_the_national_pin_and_declared_counts checks them together.

• 1306-04, X-m06 and X-s04: Port the uk-data wave to policyengine-uk 2.134.4: stamp duty, Move to UC, national home values (sixth #1095 PR) #1308 also fixes item A's OBR-leg bug. CONFIRMED (1306-04: 2/2, final severity low; X-m06 and X-s04: 1/1 each).

  • Port the uk-data wave to policyengine-uk 2.134.4: stamp duty, Move to UC, national home values (sixth #1095 PR) #1308's fix is now 908f6a3 at head 306bc11, the same patch as 21cb5e0.
  • There is one conflict hunk, the loop body of _obr_band_counts. The test file merges cleanly, and Port the uk-data wave to policyengine-uk 2.134.4: stamp duty, Move to UC, national home values (sixth #1095 PR) #1308's test passes when run against UK local surface before the Budget build: high-value homes OBR leg, per-vehicle fuel flag, Northern Ireland energy share and housekeeping (#1278) #1306's code.
  • Port the uk-data wave to policyengine-uk 2.134.4: stamp duty, Move to UC, national home values (sixth #1095 PR) #1308's version reads the non-owner share from facts.parameters.get("obr_non_owner_occupied_share") (high_value_home_gates.py:221 at 306bc11). The shipped stock facts never set that key, so in production Port the uk-data wave to policyengine-uk 2.134.4: stamp duty, Move to UC, national home values (sixth #1095 PR) #1308 reports no owner-occupied sensitivity. It also leaves the high-value homes modules out of the kernel hash.
  • Two changelog fragments ship for one fix. Under UK local surface before the Budget build: high-value homes OBR leg, per-vehicle fuel flag, Northern Ireland energy share and housekeeping (#1278) #1306's code, Port the uk-data wave to policyengine-uk 2.134.4: stamp duty, Move to UC, national home values (sixth #1095 PR) #1308's line "now reads either form" becomes false.
  • Fix: keep UK local surface before the Budget build: high-value homes OBR leg, per-vehicle fuel flag, Northern Ireland energy share and housekeeping (#1278) #1306's loop. Drop Port the uk-data wave to policyengine-uk 2.134.4: stamp duty, Move to UC, national home values (sixth #1095 PR) #1308's code hunk and changelog.d/1095-hvcts-obr-leg.fixed.md, and keep Port the uk-data wave to policyengine-uk 2.134.4: stamp duty, Move to UC, national home values (sixth #1095 PR) #1308's test.

• 1306-17 and X-m05: Generated digests and fixtures must be recomputed, not picked from one side. CONFIRMED (1/1 each, low).

  • The coverage manifest's 20 source_manifest_sha256 entries (release_input_coverage_manifest.json:473 and the 19 after it) end up wrong whichever side is kept. UK local surface before the Budget build: high-value homes OBR leg, per-vehicle fuel flag, Northern Ireland energy share and housekeeping (#1278) #1306's value is 7bfd67bb…, Port the uk-data wave to policyengine-uk 2.134.4: stamp duty, Move to UC, national home values (sixth #1095 PR) #1308's is 5f8a2bc0…, and the auto-merged source_stages.json hashes to 1d3e618b…, matching neither.
  • uk_spine.json is a single line. UK local surface before the Budget build: high-value homes OBR leg, per-vehicle fuel flag, Northern Ireland energy share and housekeeping (#1278) #1306 changes one stage_contract_sha256. Port the uk-data wave to policyengine-uk 2.134.4: stamp duty, Move to UC, national home values (sixth #1095 PR) #1308 adds 3 nodes and changes the policyengine-uk pin: the reviewed Port the uk-data wave to policyengine-uk 2.134.4: stamp duty, Move to UC, national home values (sixth #1095 PR) #1308 tree locked 2.134.5, but a verifier read 2.134.4 in uv.lock at Port the uk-data wave to policyengine-uk 2.134.4: stamp duty, Move to UC, national home values (sixth #1095 PR) #1308's current head 306bc11. Use whatever Port the uk-data wave to policyengine-uk 2.134.4: stamp duty, Move to UC, national home values (sixth #1095 PR) #1308 locks when it lands.
  • CI catches a wrong pick, but only once the PRs are retargeted to main.
  • Fix: whichever lands second reruns build_uk_release_input_coverage_manifest.py and tools/graph_uk_spine_fixture.py under the locked engine, recomputes the gate digests, reruns the census tool, and then runs the matching tests.

• 1306-c02: Item B conflicts in text with UK local release: publish a K=15, 150,000-household dense dataset in at most two builds (#1301) #1307. CONFIRMED (1/1 low).

  • The conflict is in local_target_census.py:243-246 and the high_value_homes description in uk_local_target_census.json. Re-pin the UK Chronicle feed to abf0921 (#1300) #1302 alone merges cleanly with UK local release: publish a K=15, 150,000-household dense dataset in at most two builds (#1301) #1307, so the conflict comes from B.
  • UK local release: publish a K=15, 150,000-household dense dataset in at most two builds (#1301) #1307 is ready for review, based on main and MERGEABLE.
  • Fix: name UK local release: publish a K=15, 150,000-household dense dataset in at most two builds (#1301) #1307 in the body, keep both sentences when resolving, and regenerate with tools/census_uk_local_targets.py.

• X-s06: Stale NI interim text survives in each retirement. CONFIRMED (1/1 low).

  • The stale text: Place UK private rent and BRMA at the assigned area; bind Scotland's rents at BRMA grain (#1262, #1090) #1305's selector notes, UK local surface before the Budget build: high-value homes OBR leg, per-vehicle fuel flag, Northern Ireland energy share and housekeeping (#1278) #1306's module docstring (1306-16), and docs/uk-chronicle-feed-repin.md:199-202. The doc is a dated record of the re-pin, so it is the weakest of the three.
  • Fix: update them in whichever PR keeps the retirement.

• Review date. UK local release: publish a K=15, 150,000-household dense dataset in at most two builds (#1301) #1307 adds a graph review_date, but graph_targets.py:648 and the holdout kernel still call without now. The UK local surface before the Budget build: high-value homes OBR leg, per-vehicle fuel flag, Northern Ireland energy share and housekeeping (#1278) #1306 body already names this follow-up for uk.full.problem, but not for the holdout kernel at graph_terminal.py:1220.

• A workable order, from the findings' suggestions:

  1. María rules on the NI formula.
  2. Drop the duplicate energy commit and the duplicate OBR-leg commit wholesale.
  3. Land the pre-build set.
  4. Regenerate the receipt, the membership and the references once, at the built head, on abf0921.
  5. Regenerate again after Place UK private rent and BRMA at the assigned area; bind Scotland's rents at BRMA grain (#1262, #1090) #1305 and UK local release: publish a K=15, 150,000-household dense dataset in at most two builds (#1301) #1307.

Disclosure check

• The disclosure lens found the PR's own outward text clean. It checked all 9 commit messages, all 9 changelog fragments, the body, the new parity text, the exclusions (a word-level diff shows only the tracking id changed), the holds, the vendored DESNZ NI rows (OGL admin data), the selector notes, the census text, the manifest and fixture, and the test fixtures (synthetic, or published OBR figures). The only figures are OBR-published figures, UKDS study conditions, and counts.
• Latent floors, shown only with synthetic data:
  • 1306-05: the high-value homes total and differencing.
  • 1306-10: distinct-source counting in the division tails.
  • 1306-11: differencing the two tail shares.
  • 1306-25: the fuel-flag share.
  • Of these, 1306-05's output goes to the release dir (terminal_gates.json). Whether a real England build triggers it is unverified.
• The HMRC condition (1306-09) is a gap in process, not a figure in this PR; its scope waits on María's SPI 9422 ruling.
• The PR's self-certification (1306-22) is worded too broadly, but it is not a breach.

Tests and tools run

• Pytest, by surface:
  • test_uk_domestic_energy_targets.py: 13 passed.
  • High-value homes gates, energy and parity: 42 tests, 0 failed, 1 skipped (the uk-data tree re-hash). The LCFS fuel and tail subset (-k): 27 passed.
  • Vendoring, energy, census, parity, coverage manifest and high-value homes gates: 90 tests, 0 failed, 2 skipped (the feed round trip and the uk-data re-hash).
  • The touched test files: 156 tests, 155 passed, 1 skipped.
  • The local row-wise, holdout, weight-wiring and CGT observation-period files: 64 passed.
  • The same four files plus full-solve scope and graph terminal, with the clock forced to 2027-01-20: 81 passed.
  • The generated-surface, register and spec tests: 354 tests, 353 passed, 1 skipped.
  • The touched files plus census and high-value homes, with the clock forced: 199 tests, 198 passed, 1 skipped.
  • A clock control confirmed a default-date solve is refused on 2027-01-20.
  • Claims checks: 44 tests, 0 failed, 1 skipped. Disclosure checks: 18 passed. Second pass: 44 tests, 0 failed, 1 skipped.
  • The 1306-14 verifier: 66 passed with the clock forced, and 2 passed in the control run.
• Tools: census_uk_local_targets.py --check (current), census_uk_data_target_parity.py --check (current), build_uk_release_input_coverage_manifest.py --check (current; 144 required, 1 reviewed exclusion), ci_test_plan.py verify (ok), and graph_acceptance_burndown.py --verify (ok).
• Ruff: ruff check and ruff format --check --no-cache are clean on the 20 edited .py files.
• Probes:
  • gb_energy_level_factors() at Re-pin the UK Chronicle feed to abf0921 (#1300) #1302, Place UK private rent and BRMA at the assigned area; bind Scotland's rents at BRMA grain (#1262, #1090) #1305 and UK local surface before the Budget build: high-value homes OBR leg, per-vehicle fuel flag, Northern Ireland energy share and housekeeping (#1278) #1306.
  • Re-vendoring the DESNZ resource: 699 rows, byte match.
  • A uk_spine.json diff against the live graph.
  • Probes of _suppress_small_cells, weighted_top_share, _obr_band_counts, assign_recipient_has_fuel, the import assert, the kernel hash and Port the uk-data wave to policyengine-uk 2.134.4: stamp duty, Move to UC, national home values (sixth #1095 PR) #1308's test body on UK local surface before the Budget build: high-value homes OBR leg, per-vehicle fuel flag, Northern Ireland energy share and housekeeping (#1278) #1306's code.
  • The gate's own _signed_difference_check_failures across the stack's receipts.
• Read-only checks: fixed-string greps of the abf0921 feed (NI 2024 rows, Hartlepool rows, MGRN/MGRQ); GitHub API reads of uk-data a43e8c5 files (57/57 inventory hashes match) and of uk-data#529; trial merges (git merge-tree) of UK local surface before the Budget build: high-value homes OBR leg, per-vehicle fuel flag, Northern Ireland energy share and housekeeping (#1278) #1306 against Place UK private rent and BRMA at the assigned area; bind Scotland's rents at BRMA grain (#1262, #1090) #1305, Port the uk-data wave to policyengine-uk 2.134.4: stamp duty, Move to UC, national home values (sixth #1095 PR) #1308 and UK local release: publish a K=15, 150,000-household dense dataset in at most two builds (#1301) #1307.
• Not reproduced: the body's "2,522 passed / 5 skipped on 53 files". The engine-lane H2 parity test, the uk-data tree re-hash and the feed round trip were not run.
• Late in the run, some verifiers could not run tests and read the code or worked by hand instead (1306-16, 1306-21, 1306-24, 1306-25 and one 1306-02 voter).

Checked and found sound

• A: row selection matches the vendored Table 1.1 rows (static 2028: 71k/54k/25k/15k, total 165k; 40% share; 99,000 sensitivity). Per-band ratios use the suppressed shipped bands. The kernel hash now covers both high-value homes modules.
• B: the scope dict reaches the evidence summary and both gates' details, and the census text was regenerated.
• C: exactly 22 entries move to UK local surface, consumption and high-value homes: remaining items (consolidation) #1278. No approved_on or expires_on changed, and no signature or digest covers these registers.
• D:
  • The register follows its own rules: two concerns may share dwp_housing_benefit, the cited target ids exist, and the status counts move as the body says.
  • Four of the five obr.py citations are exact.
  • The Band D and working-age HB verdicts match uk-data.
  • abf0921 has no LFS employee rows, so ons_labour_market is correctly blocked.
• E: now reaches the binding review from all three entry points. Builds still default to today, so expiry stays fail-closed.
• G: no floor value changed.
• H:
  • weighted_top_share picks the smallest group that reaches the fraction and handles ties correctly.
  • The donor side has one row per LCFS household, so its floor is sound.
  • H adds no new crash path.
• I:
  • The two NI rows match the feed, the 697 meter rows are unchanged, and _one_value fails closed on a missing or bad NI row.
  • Only local DESNZ kWh references move.
  • No leftover callers or constants remain.
• J1: 1 − z^n is computed against the same keyed uniforms, a one-vehicle household reproduces the old threshold, and the rename is consistent across the YAML, the stage manifest, the schema and the H2 fixture.
• Committed hashes: no committed file pins the base sha256 of any changed resource.

Refuted during verification

• 1306-01: three of the files listed as textual conflicts (local_binding_adjudications.json, uk_local_target_census.json, local_target_census.py) actually merge cleanly.
• 1306-02: one voter thought "the build reads the receipt" was overstated. X-s01 confirms the full/dense preflight does read it.
• 1306-04: "the test file conflicts", "Port the uk-data wave to policyengine-uk 2.134.4: stamp duty, Move to UC, national home values (sixth #1095 PR) #1308's tests fail on UK local surface before the Budget build: high-value homes OBR leg, per-vehicle fuel flag, Northern Ireland energy share and housekeeping (#1278) #1306's code" and "the sensitivity disappears if Port the uk-data wave to policyengine-uk 2.134.4: stamp duty, Move to UC, national home values (sixth #1095 PR) #1308 wins" are all false for a real merge.
• 1306-07: "Microcosm still binds obr.universal_credit_outside_cap" is false; it is excluded under Swap the microcosm-built UK spine into the national build and retire the incumbent feed #757. "Added a three-year carry_forward rule" is inaccurate; a43e8c5 adds only a per-target opt-out. restrict_to_countries is plumbing, not a target family. The inventory's scope gap predates the PR.
• 1306-12: an abort after the solve is unlikely, because the vendoring row counts and the PR's tests catch bad re-vendors first.
• 1306-14: "CI no longer notices" is slightly too strong. engine-uk goes red from 2027-01-15, for the census fence only.
• 1306-16: "the census provenance still describes the LFS interim" is false. The verified_on comment is merely dated.
• 1306-19: the company-car sub-claim is unsubstantiated.
• 1306-23: the salary-sacrifice wording is broadly accurate.
• 1306-c01: "works against UK local release: publish a K=15, 150,000-household dense dataset in at most two builds (#1301) #1307's aim" is overstated. UK local release: publish a K=15, 150,000-household dense dataset in at most two builds (#1301) #1307's preflight kernel already hashes terminal-only code by design.
• 1306-c03: the UK spine: SLC plan-2 borrower stocks at a third of SLC's — the FRS flag captures repayers, not the liable stock #868 part is a question for María, not a defect. Commit C keeps UK spine: SLC plan-2 borrower stocks at a third of SLC's — the FRS flag captures repayers, not the liable stock #868 on the SLC hold deliberately.
• X-s01: "a third build" is overstated. The dense path refuses before the solve.
• X-m01: the stale receipt is silent only at merge time and in CI; the gate refuses it at build time.
• X-s02: its suggested fix (keep UK local surface before the Budget build: high-value homes OBR leg, per-vehicle fuel flag, Northern Ireland energy share and housekeeping (#1278) #1306's formula because of the ruling's wording) is contradicted by the approved plan.
• X-m05: dense digests need recomputing only in the UK local release: publish a K=15, 150,000-household dense dataset in at most two builds (#1301) #1307 × Place UK private rent and BRMA at the assigned area; bind Scotland's rents at BRMA grain (#1262, #1090) #1305 merge.
• Second-pass item 1 (production kernels ignore the review date): dropped as pre-existing and disclosed in the body. Small remaining gap: the body's follow-up does not name the holdout kernel.
• Second-pass item 5 (obr.pension_credit "stamped Great Britain"): dropped, contradicted by target_references.json (K03000001).

Not covered

• The microcosm-graph H2 uk_spine parity test. It runs in the engine lane, which this review did not set up for this PR. The JSON was compared with the live graph, excluding numerical_dependencies.
• A full local compile on abf0921 with item I. Builds were out of scope, so the release-blocking refusal is inferred from the gate code and the receipt comparison, not run.
• G's study-condition claims (SN 9563, 9603, 7215, 5340, 8856, 9422) were not re-read against the UKDS catalogue.
• The feed-gated vendoring round-trip test and the English region council-tax sum test.
• Whether a real England build hits 1306-05's suppression gaps. No gate output from a real build was available.
• Outside the diff: whether UK spine: SLC plan-2 borrower stocks at a third of SLC's — the FRS flag captures repayers, not the liable stock #868's closure makes the slc.borrowers.plan_2_* exclusions (expiring 2026-11-03) stale.

🤖 Generated with Claude Code

[juaristi22]

Every finding of the review above is addressed below. The branch is rebased onto main after #1302's merge, and the PR now targets main.

History. The nine item commits are replayed on main. Only two messages change: J1 and D, which the review flagged. The fixes sit in seven commits on top, one per item, so each finding's change can be read on its own:

• R1: A and B
• R2: C
• R3: G
• R4: J1 and H
• R5: E
• R6: D
• R7: I

Blocking

• Item I's formula against Place UK private rent and BRMA at the assigned area; bind Scotland's rents at BRMA grain (#1262, #1090) #1305. María ruled on 17 October that the share-scaling formula stands: Great Britain's part of the Energy Trends level is its share of DESNZ's sub-national consumption with Northern Ireland's row added.
  • Why. The sub-national statistics run about 9% (electricity) and 12% (gas) above Energy Trends. Subtracting Northern Ireland's row from the Energy Trends total would take it out at a higher basis than Great Britain's part. The share keeps both nations on one basis.
  • Where it's recorded. The body and item I's fragment now say this departs from the plan's subtraction, and why.
  • Cross-PR. The retirement stays here. Place UK private rent and BRMA at the assigned area; bind Scotland's rents at BRMA grain (#1262, #1090) #1305 drops its energy commit (6928d08) and changelog.d/1262-uk-energy-northern-ireland-level.changed.md whole, rather than resolving hunk by hunk.
• The local compile-parity receipt (and X-s01). R7 regenerates the local compile-parity receipt on abf0921 at this head.
  • What moves. Exactly the 694 kWh rows: 348 electricity and 346 gas, ledger_value only, each moved by its fuel's new factor over the interim's (0.99901 and 1.01582). Every other row is byte-identical to Re-pin the UK Chronicle feed to abf0921 (#1300) #1302's regeneration.
  • Body. It now says item I changes those rows.
  • Merge order. Re-pin the UK Chronicle feed to abf0921 (#1300) #1302's own regeneration (d8493a3) already covers the 3,507 UC cells, so this head's receipt matches the live compile once UK local surface before the Budget build: high-value homes OBR leg, per-vehicle fuel flag, Northern Ireland energy share and housekeeping (#1278) #1306 lands. Place UK private rent and BRMA at the assigned area; bind Scotland's rents at BRMA grain (#1262, #1090) #1305 and UK local release: publish a K=15, 150,000-household dense dataset in at most two builds (#1301) #1307 regenerate again if they land after it.

Should fix

• The membership report. Regenerated in the same pass (R7).

  • What moves. Only the same 694 cells: their hold factors and resolved values.
  • Unchanged. local_target_references.json is byte-identical.
  • Body. The report is now listed under Regenerated.

• Item G's tool text. R3 brings all three places in line with the docstring: the default's comment, the --sdc-minimum-count help and the caller obligations. The obligations also carry the SPI tape's two conditions.

• High-value homes band suppression. R1 makes three changes:

  • The total. It is withheld when it rests on fewer than ten distinct source households.
  • The withheld bands together. Further bands are withheld until the withheld ones rest together on ten, so the total less the shipped bands cannot give back their sum.
  • Regions. The regional shares (which sum to one) follow the same rule, and so does their shipped frame total.

  The ratios read the shipped counts, so they go with them. The toy-pool test now expects the total withheld. New tests cover the review's two probes: a single band on five sources, and bands on 40, 4 and 3 sources.

  The frozen build's high-value homes evidence still gets a disclosure check before any gate report goes public.

Minor

• uk-data#529's validation-only switch. R6 adds a note to local_contract.hmrc.employment_income.count: uk-data#529 made the count validation-only in create_datasets.py, outside the inventory's scope, and Microcosm still binds it under hmrc_spi_frame_model_proxy. The LFS concern points to it. The body lists the ruling owed on UK: track the defects uk-data is fixing that microcosm still has #1095.
• The SPI tape's HMRC condition. It is now in the baselines tool's caller obligations (R3). Whether frame outputs count as analyses of the tape is María's SPI 9422 ruling, listed in the body as owed.
• The tail floor counted rows. It now counts distinct source households (R4). The stage passes source_household_id, which the SPI support channel writes before the LCFS stage runs.
• Differencing the two tail shares. Both fractions now come from one ordering (R4). A top-10% share is "<10" when its households beyond a shipped top 1% rest on one to nine sources. A skewed-weights test covers it.
• The OBR leg's row selection. R1 selects by concept, England, fiscal-year period type, year and basis, and requires every band plus the total, once each.
  • Otherwise. The leg reports no_matching_rows, incomplete_rows or malformed_rows and never raises.
  • The share. It is read for the leg's year where the rows state one, else the note's one year (2028-29), and the leg names that year.
  • A share it cannot read. The leg reports why and keeps the band comparison.
• COMMITTED_BINDING_NOW asserted at import. R5 drops it.
  • The date. COMMITTED_BINDING_NOW is still the register's latest approval.
  • What the review checks. On that date it is a build's review: a lapsed fence a fixture binds fails, and a dormant one does not.
• CI's signal that the local adjudications lapse. R5 adds test_no_committed_local_adjudication_is_about_to_lapse. Following the holds' clock, it fails fourteen days before the earliest local expiry.
• Northern Ireland's comparability caveat. R7 adds one sentence on Northern Ireland's periods and comparability to the desnz_subnational_consumption_basis adjudication and its census fence. María signed off the rewording on 17 October; the dates are unchanged.
  • Sources. DESNZ's two Northern Ireland publications are now family sources, verified 17 October.
  • The electricity note. It now says Northern Ireland is published separately and enters the factor only.
• The module docstring. Rewritten in R7, along with the census's verified-on comment.
• Commit D's "no classification changes". The fragment now says the re-extraction changes no classification and one is corrected by hand. D's commit message is reworded the same way.
• J1's message. Reworded: the draw keeps the share per car, but the frame stores one flag per household. R4 adds the share of vehicles in fuel-free households to the receipt, beside the per-car share.
• Pension-age Housing Benefit called "ported". R6 states the difference in the concern's evidence. uk-data applies the pension-age regulations, which policyengine-uk 2.132.6 encodes as housing_benefit_pension_age_regulations_apply. Microcosm's four rows filter on the deprecated eldest_adult_age alone. Re-binding moves national targets, so it goes with UK spine: housing benefit £2.7bn against OBR £12.1bn — pensioner and supported-accommodation HB missing after the UC migration #867 (PR 4) rather than here, and the body says so.
• The owner-occupied ratio equals the private-sector one. Dropped (R1), with a comment saying why: the stock build sets the owner-occupied share from the same OBR share.
• The Disclosure section's wording. The body now says "no survey-derived figures".
• Item A's hash and the dense solve's key. R1 moves the high-value homes code out of the full-gates kernel's hash. The calibrated node now carries that code's source hash in its parameters (terminal_code), and the kernel checks it before running.
  • Effect. An edit to that code re-keys the calibrated gates and what follows them, never the preflight or the solve.
  • Composes with UK local release: publish a K=15, 150,000-household dense dataset in at most two builds (#1301) #1307. The split doesn't depend on UK local release: publish a K=15, 150,000-household dense dataset in at most two builds (#1301) #1307's preflight kernel. Once that lands, the code could move into the terminal-only kernel's hash instead.

Nits

• The parity citation. It now reads obr.py:620 (_parse_tv_licence) and 668-684 (the static rows before get_targets).
• Untested paths. New tests cover:
  • the OBR concept, geography and period-type filters;
  • a missing total;
  • the share's three failure modes;
  • all five holdout folds receiving now.
• The fuel-flag share's floor. It now carries the floor.
• The two text slips.
  • The two hold reasons now point at UK local surface, consumption and high-value homes: remaining items (consolidation) #1278.
  • The body's file count is corrected.
• Item B cited closed UK: data fixes to model the High Value Council Tax Surcharge (mansion tax) correctly, and the related housing and council-tax gaps #1240. The scope note, its comment, the census family text and B's fragment now cite the coverage decision as tracked on UK local surface, consumption and high-value homes: remaining items (consolidation) #1278.
• The SLC exclusions on UK spine: SLC plan-2 borrower stocks at a third of SLC's — the FRS flag captures repayers, not the liable stock #868. María ruled on 17 October that they move to UK local surface, consumption and high-value homes: remaining items (consolidation) #1278 too (R2). Their adjudications record the move, and the plan-2 and plan-5 uprating hold keeps UK spine: SLC plan-2 borrower stocks at a third of SLC's — the FRS flag captures repayers, not the liable stock #868.

Cross-PR

• Port the uk-data wave to policyengine-uk 2.134.4: stamp duty, Move to UC, national home values (sixth #1095 PR) #1308. It keeps its test and drops its _obr_band_counts hunk and changelog.d/1095-hvcts-obr-leg.fixed.md. It also doesn't need the non-owner share from facts.parameters.
• UK local release: publish a K=15, 150,000-household dense dataset in at most two builds (#1301) #1307. Its family-text conflict with B is resolved by keeping both sentences and regenerating with tools/census_uk_local_targets.py. Its UKFullPreflightGateKernel composes with R1's terminal_code.
• Generated files. Whichever PR lands second re-runs:
  • build_uk_release_input_coverage_manifest.py;
  • tools/graph_uk_spine_fixture.py under its locked engine;
  • the census tool.
• The review date. The follow-up (the uk.full.problem kernel and the holdout kernel taking the build's date) is named in the body and in E's fragment.
• The re-pin doc. Its history entry gets a forward pointer to this retirement.

Verification

Local, since this is the first CI run on main. See the body for the counts:

• the touched-surface funnel;
• the generator checks;
• ci_test_plan.py verify and graph_acceptance_burndown.py --verify;
• ruff.

The receipt and membership were regenerated on a cut of the pinned abf0921 artifact (digest checked against the pin), not the whole feed: both generators hold the whole file in memory. The cut keeps the constituency and local-authority rows that match a local contract selector. No local reference reads operands outside its own selector, so the cut holds every candidate. Every row other than the 694 energy cells came out byte-identical to the committed full-feed outputs.

🤖 Generated with Claude Code

[juaristi22]

Re-review of this PR at 567d383

Run with Claude Code on 17 October 2026, after the rebase onto main (post #1302), the seven fix commits (R1–R7) and the response. Three independent reviewers re-checked every first-round finding and every claim in the response, attacked the new code with synthetic probes, checked the new public text against the UK Data Service rules, and simulated the merges with #1305, #1307 and #1308. No builds and no feed-gated tests were run.

Verdict

Every first-round blocker and should-fix item is fixed, and the reviewers verified each one in code. CI is 10/10 green at this head.

• Regenerated receipt. It moves exactly the 694 DESNZ kWh rows, on ledger_value only, at uniform per-fuel factors of 0.99901 and 1.01582. Every other row is byte-identical to main's.
• Membership. It moves the same 694 cells' hold factors and resolved values. The references are byte-identical, and the gate replay passes with 0 failures.
• The feed cut was complete for this head. No local reference reads facts outside its own selector.
• terminal_code is sound. It composes with UK local release: publish a K=15, 150,000-household dense dataset in at most two builds (#1301) #1307's live head.

What is left is low.

Low

• The fuel-flag share is floored on the wrong side.
  • Where: lcfs_consumption.py:2561-2575.
  • flagged_share_among_vehicle_households is floored on the flagged (large) count. So with 1–9 fuel-free vehicle households it ships, and 1 − it gives back their share, which fuel_free_share_of_vehicles withholds.
  • The raw with_vehicles_share − flagged_share gives the fuel-free share of all households.
  • The floors count rows, including zero-weight rows and copies.
  • This only matters on narrowed or sampled runs.
  • Fix: pass _source_households in, floor both vehicle shares on the smaller of the flagged and fuel-free distinct positive-weight source counts, floor or drop the raw pair, and correct the "<10" wording in the body, the J1 fragment and the docstring.
• The top-10% share's complement is not floored.
  • Where: :836-847, :981-995.
  • Each division's frame and donor totals ship beside the shares, so total × (1 − s10) is the spend outside the top 10%.
  • The floor counts top-set rows, zero spenders included. So when 1–9 households outside the top 10% spend anything, the share still ships. A synthetic probe recovered 5 spenders' spend exactly.
  • It reaches committed files only through docs/evidence/*/scripts/stage_dev_run.py.
  • Fix: also withhold when the nonzero contributors outside the top set number 1–9.
• The distinct-source fix is untested and fails open.
  • No test passes sources through division_capture or the stage, so deleting either wiring line leaves every test green.
  • _source_households returns None when the column is missing, and the tails then count rows silently.
  • The real spine does carry source_household_id by stage 24.
  • Fix: a test with copies, and fail closed when copy columns exist without the id.
• terminal_code re-reads its source files at run time (graph_terminal.py:785).
  • Editing or checking out high_value_home_gates.py or high_value_homes.py in the build's checkout during a long build aborts uk.full.gates.calibrated after the solve. A restart re-runs only the gates.
  • Memoise the hash per process, or say so in the runbook.
• Merging with UK local release: publish a K=15, 150,000-household dense dataset in at most two builds (#1301) #1307 needs one hand resolution.
  • The calibrated node's params in graph_terminal.py must become {**gate_params, "phase": "terminal", "terminal_code": uk_terminal_only_code_hash()}.
  • Keeping UK local release: publish a K=15, 150,000-household dense dataset in at most two builds (#1301) #1307's side drops terminal_code, and only the engine-lane admission test catches that. Otherwise the terminal phase refuses after the solve.
  • Keeping this PR's side drops UK local release: publish a K=15, 150,000-household dense dataset in at most two builds (#1301) #1307's gate clock, which an engine-free test catches.
  • Name this in the body, and drop the "move into the terminal-only kernel's hash" suggestion: that subclass inherits the full hash, so the move would re-key the preflight.
• terminal_code's scope is narrower than the docstring says.
  • high_value_homes.py is also hashed by the target kernels (graph_targets.py:220-229), so editing it still re-solves. Only high_value_home_gates.py is solve-free.
  • For UK local release: publish a K=15, 150,000-household dense dataset in at most two builds (#1301) #1307's build 2: put any high-value homes disclosure fix in high_value_home_gates.py only, and don't move these hashes between build 1 and build 2. graph_terminal.py itself is hashed by the preflight and holdout kernels.
• "Never raises" isn't quite true for the OBR leg. A matching row with a null, missing or non-numeric value raises, and so does a non-mapping period or dimensions (high_value_home_gates.py:173, :195, :213, :243). NaN or negative counts pass as "available". Low risk, since the rows come from the vendoring tool.

Nits

• The comment at high_value_home_gates.py:388-391 still describes the old single-region rule.
• Two thresholds for "over £2m": the bands use >= 2m and the regional set uses > basis. No leak today; the :398 comment's "ships beside the shares" is inaccurate.
• The D fragment says "three source files changed"; one of the three is a test.
• Stale text in the body and response:
  • Body item I says the plan's subtraction is "as Place UK private rent and BRMA at the assigned area; bind Scotland's rents at BRMA grain (#1262, #1090) #1305 does", but Place UK private rent and BRMA at the assigned area; bind Scotland's rents at BRMA grain (#1262, #1090) #1305 has dropped it.
  • The body says Port the uk-data wave to policyengine-uk 2.134.4: stamp duty, Move to UC, national home values (sixth #1095 PR) #1308 "keeps its OBR-leg test" (the response: "keeps its test"), but Port the uk-data wave to policyengine-uk 2.134.4: stamp duty, Move to UC, national home values (sixth #1095 PR) #1308's live head (8739557) dropped the whole commit, test included. This PR's own tests cover the vendored rows.
• Still open by design (María's rulings owed):
  • 1306-09: nothing yet marks SPI-dependent outputs pending the SN 9422 ruling.
  • 1306-10's residual: no floor on distinct LCFS donor diaries behind a frame top group.

Regeneration note for #1305

The feed cut used for R7 (constituency and local-authority rows matching a local selector) was complete at this head. #1305 adds 18 ons.rent.private_rent_brma references at geography level statistical_scope, so whoever regenerates after #1305 must widen the cut.

First-round dispositions

• Fixed: every first-round finding not listed below as partly fixed (c01 to c03 included), and the cross-PR items on this PR's side.
• Partly fixed:
  • 1306-09 (pending the SPI 9422 ruling);
  • -12 ("never raises");
  • -25 (the fuel-flag floor).

Cross-PR

• This PR × Place UK private rent and BRMA at the assigned area; bind Scotland's rents at BRMA grain (#1262, #1090) #1305. One trivial conflict (the census test's geographies). The union receipt and membership pass the gate.

• This PR × UK local release: publish a K=15, 150,000-household dense dataset in at most two builds (#1301) #1307 (e119867). Four conflicts:

  • the census text in local_target_census.py and the generated uk_local_target_census.json: keep both high-value homes sentences, then re-run the census tool;
  • graph_terminal.py: as above;
  • test_uk_graph_terminal.py: keep both tests.

  The receipt and membership auto-merge to the correct union (27,249 rows).

• Port the uk-data wave to policyengine-uk 2.134.4: stamp duty, Move to UC, national home values (sixth #1095 PR) #1308 (8739557, rebased onto main, OBR-leg commit dropped). It conflicts with this PR on the coverage manifest's source_manifest_sha256 and on uk_spine.json. Regenerate both after both land, under Port the uk-data wave to policyengine-uk 2.134.4: stamp duty, Move to UC, national home values (sixth #1095 PR) #1308's policyengine-uk 2.134.4 lock.

Disclosure check

These parts of the new public text are clean:

• R1–R7's messages and the reworded J1 and D messages;
• the nine fragments, the body and the response;
• the 694 changed rows of the receipt and membership (DESNZ times a factor; no text);
• the synthetic test probes.

Tests and tools run

• At 567d383: 672 tests, 0 failures, 7 skips, plus 158 more, 0 failures.
• tools/ci_test_plan.py verify, the census and parity --checks, the coverage manifest --check and tools/graph_acceptance_burndown.py --verify are all current or ok. Ruff is clean on the 23 edited files.
• Gate replay: 0 failures.
• CI: 10/10 green, engine-uk included.

🤖 Generated with Claude Code

[juaristi22]

Every item of the re-review above is addressed. The branch is rebased onto main after #1307's merge (ff3f348). The first 16 commits are replayed unchanged apart from the four conflicts the re-review named, resolved as it described:

• The census text. The family keeps both high-value homes sentences, and the census is regenerated.
• graph_terminal.py. The calibrated node takes {**gate_params, "phase": "terminal", "terminal_code": uk_terminal_only_code_hash()}.
• test_uk_graph_terminal.py. Both sides' tests are kept.

The receipt and membership auto-merged to the union, 27,249 rows. A regeneration on the combined tree, from a re-cut of the pinned artifact identical to R7's, reproduced the committed receipt, references and membership byte for byte.

Low

• The fuel-flag floor (R8).

  • Groups. Every share on a side differences into the households with and without a vehicle, crossed with fuel or none. So each share is "<10" when any non-empty group rests on one to nine distinct positive-weight sources: the frame's source households, the donor's diaries.
  • The raw pair. It is floored the same way.
  • Wording. The docstring and J1's fragment say so.
  • Tests. A ten-plus case, a thin-group case on both sides, and copies counted as sources.

• The top-10% share's complement (R8). The tail floors now count sources that spend anything, so zero rows lend a group nothing. A share is withheld when one to nine spenders sit:

  • in its top group;
  • outside it;
  • or between it and a shipped smaller fraction's group.

  A test covers the five-spenders probe and a top group padded with zero rows.

• The distinct-source fix (R8).

  • Wiring. The stage computes the source households once and passes them to both receipts. A stage test on the synthetic frame checks that both get them, and division_capture is tested with copies.
  • Failing closed. A frame that marks copies without a source_household_id is refused: the SPI support channel's household_support_clone_index, and the high-value split's household_is_high_value_support_copy or high_value_support_copy_index.

• terminal_code re-reading its files (R9). The hash is memoised per process, so editing the files during a build cannot fail the calibrated gates after the solve.

• The merge with UK local release: publish a K=15, 150,000-household dense dataset in at most two builds (#1301) #1307. Resolved as above.

  • Dropped. The "move into the terminal-only kernel's hash" suggestion is gone from the body: the preflight kernel inherits that hash.
  • The hash test. It now covers the preflight kernel too.

• terminal_code's scope (R9). The docstring, A's fragment and the body now say that only high_value_home_gates.py is solve-free. high_value_homes.py also feeds the target kernels, so a disclosure fix between two builds that must reuse the solve belongs in high_value_home_gates.py alone.

• "Never raises" (R9). These are now reported as malformed_rows or incomplete_rows:

  • a missing, non-numeric, NaN or negative count;
  • a period, dimensions, layout or geography field that is not a mapping;
  • a row that is not a mapping.

  An unreadable share leaves the band comparison standing. Tests cover each case.

Nits

• The two comments (R9). The regional comment now describes the joint floor. The thresholds comment says the split follows the derived stock's "over £2m" and the bands follow OBR's, and that the evidence summary, not the gate details, carries the total.
• D's fragment (R10). It now says "two source files and a test".
• The body. It drops "as Place UK private rent and BRMA at the assigned area; bind Scotland's rents at BRMA grain (#1262, #1090) #1305 does" and says Port the uk-data wave to policyengine-uk 2.134.4: stamp duty, Move to UC, national home values (sixth #1095 PR) #1308 dropped its OBR-leg commit, test included.
• Still open by design. These wait on María's rulings: marking SPI-dependent outputs pending the SN 9422 ruling, and a floor on distinct donor diaries behind a frame top group.

With #1307's remaining fixes

This PR also carries the remaining fixes from the third look at #1307:

• the wording of the two below-floor region rows, with the local membership regenerated (N1);
• the receipt-regeneration memory guidance in the re-pin procedure (L6);
• the release-tooling items: the verdict JSON, the free-text allowlist, the walk's shapes, the pre-flight gaps, the build-1 retry guidance and the docs wording. See the body.

Regeneration note for #1305

The re-pin procedure now says when the local-selector cut holds, and to widen it for references at other levels, such as #1305's BRMA rows at statistical_scope.

🤖 Generated with Claude Code