Place UK private rent and BRMA at the assigned area; bind Scotland's rents at BRMA grain (#1262, #1090)

[juaristi22]

Puts private rent and the Broad Rental Market Area (BRMA) at each placed household's assigned area on the local line (#1262, item 4 of #1123), and binds Scotland's PIPR rents at the grain ONS publishes them. Draft until the head rebuilds: one spine and one dense build, compared with the 15 October main build (María's ruling of 16 October). #1307, which defers the high-value-homes cells that stopped a dense build on main, has merged, so the head builds can run. The PR also waits for Max's view on the direction, asked on PolicyEngine/policyengine-uk-data#521.

Rebased onto main after #1307 and #1306 merged.

Closes #1262. Closes #1090. Folds in and supersedes #1092: its overlap table, decision note and rent dispersion receipt, rebuilt on the area-to-BRMA layer.

The Northern Ireland energy interim is not retired here. María ruled on 17 October that #1306 retires it with its share formula, and #1306 has merged; this PR's subtraction commit was dropped.

Summary

B1. BRMA from the assigned area

• The layer. uk/brma_area_layer.json places every 2021 LSOA of England and Wales and every 2022 output area of Scotland in one BRMA. tools/build_uk_brma_area_layer.py rebuilds it from Chronicle's raw registrations, pinned by sha256:
  • England and Wales use LSOA population-weighted centroids in the VOA and Rent Officers Wales polygons, with Weight FRS BRMA draws by census private-rented households policyengine-uk-data#516's overrides. Its region × BRMA support table agrees exactly.
  • Scotland uses output-area centroids corrected by Rent Service Scotland's postcode lookup (9 filled, 59 corrected), with Measure Scotland's BRMA × council overlap; keep BRMA rents out of council targets #1092's 2 Balloch overrides. Its 63 BRMA × council pairs agree exactly.
• The node. uk.full.geography.brma (--brma-localisation area, the default on the atomic assignment) reads each placed Great Britain clone's BRMA off its LSOA or output area, so its LHA rate describes the same place as its local authority and constituency.
  • Northern Ireland and the national line keep the spine's frs_brma draw.
  • A terminal, release-blocking gate, uk_local_brma_enum_domain, rechecks the column; an off-domain value stops even a development build. The full and dense battery digests move.

B2. Private rents from the lists of rents

• The quantile table. uk/brma_rent_quantiles.json holds 25 knots per GB BRMA and bedroom category B to E. Its sources:
  • England: VOA's April 2026 list;
  • Wales: Rent Officers Wales, October 2023 to September 2024 (ATISN 25142 Table 2);
  • Scotland: the Scottish Government's 2025 quartiles with the shape of Rent Service Scotland's 2016–21 lettings.
• The node. uk.full.private_rent (--private-rent-localisation brma_list) redraws each GB private renter's positive rent, in both channels.
  • Each renter keeps its rank among its region and bedroom category's FRS-channel renters. SPI rows are ranked against that reference, and all clones of a household share a rank.
  • It reads that quantile off its assigned BRMA's list.
  • One scale per region holds the region's FRS mean rent on the census private-renter mix of its BRMAs (María's ruling, kept on 17 October). Placement follows all households, not renters, so at design weights a region's level moves by the gap between the two mixes until calibration's tenure rows restore it; the receipts report that level as a named check.
  • A sampled development rung ranks and scales a region whose sample holds fewer than ten FRS renters on Great Britain's renters, counted in the receipt. A full-scale build uses each region's own renters and refuses a region with none.
  • Category A is never used. Owners, social renters, zero rents and Northern Ireland are untouched, and the boarder, lodger, subletting, service-charge and reported Housing Benefit columns stay as reported. The receipt carries rounded record counts only.

B3. Scotland's rents at BRMA grain (#1090)

• The rows. ons.rent.private_rent_brma binds PIPR's 18 Scottish BRMA rows. They sit at the crosswalk's new statistical_scope level and on a partial brma area type: households outside Scotland belong to no BRMA area.

• The values. Each row is 12 × the BRMA's calendar-2025 mean × its private renters.

  • The renters are derived after reconciliation from the councils' tenure/private_rent cells, through Census 2022's council shares of each BRMA's renters (UV404 by output area).
  • So the BRMA counts add up to Scotland's reconciled renters.

• The 32 council cells stay signed deferred; councils have no published rent. The parity concern cross_grain_private_rent_scotland_brma resolves to the BRMA rows.

• Selection. --target-geographies can name brma. Without area BRMAs the rows would bind on the spine's regional draw, so they are left out and counted; naming brma without area BRMAs is refused.

• Weighting. The rows weigh inside the local-authority grain under every rule, grain_equal included.

• Diagnostics. A BRMA row's model figures leave a build only when at least 10 distinct source households of positive weight carry them (María's ruling of 17 October). A thinner row keeps its place and its failure. Its estimates and errors are blank in:

  • the calibration diagnostics and the targets table;
  • the gate report, whose fit gates name it by key only;
  • the manifest and the scorer;
  • the sign-off drafter and Orrery's summaries.

  It stays out of every rollup, the family fit gate included. While any row is withheld, the diagnostics' totals cover the published rows, labelled in uk_diagnostics.model_figures_withheld (María's ruling of 18 October). The release contract accepts such a row only if all its model fields are null and the block counts it.

• Active local references move from 22,986 to 23,004.

Texts and receipts

• The rent declarations say that placed renters' rents now come from the rent-officer lists, which PIPR shares as an evidence base: the rows stay bound, but their pattern is not independent validation.
• The Northern Ireland deferral says PIPR has a country mean only. "A17-uprated" becomes "SPREE-drifted in England, reconciled to the nation household controls".
• The local compile-parity receipt is regenerated on this head (27,267 rows, 23,004 compiled). The release-blocking preflight gate uk_ledger_compile_parity_local_incumbent_2025 reads it. 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 UK local release: publish a K=15, 150,000-household dense dataset in at most two builds (#1301) #1307 have merged, so whichever of Place UK private rent and BRMA at the assigned area; bind Scotland's rents at BRMA grain (#1262, #1090) #1305 and Port the uk-data wave to policyengine-uk 2.134.4: stamp duty, Move to UC, national home values (sixth #1095 PR) #1308 lands second regenerates the receipt again on the combined tree, with the full-battery and dense digests. The support floor's two authorities carry the floor's own signed rationale.

For review

• The rent rows bind PIPR's level; the node holds the FRS level, which sits below it. With tenure rows bound, calibration can close the gap only by moving weight toward higher-rent renters within each area. That is the level trade-off of UK private rent at the assigned area (#1123 item 4, "PR B"): what is in place and the pieces still missing #1262's A3/A4 rulings, declared on both cross-grain routes rather than modelled. The 18 BRMA rows' fit in the head build shows its size.
• A withheld row stays out of the family fit gate's decision. uk_local_per_family_fit decides on the published rows, so a family share never covers a row whose figures are withheld; uk_local_target_fit still decides on every row. For a problem that binds BRMA rows, Orrery's size summary publishes no total over every row, since it cannot tell which rows a build withholds.
• The BRMA rows weigh inside the local-authority grain's private-rent family. A grain of their own would take a fifth of the loss for 18 rows (a quarter under grain_equal) and push the energy and bus families below their loss-share floors.
• Cross-border BRMAs. Wales's list is a year older than England's, and 16 LSOAs read the other nation's list under their own region's scale: West Cheshire's 11 Welsh LSOAs, and Monmouthshire's 4 and Brecon and Radnor's 1 English LSOAs.
• Calibration diagnostics gain a brma geography level without a schema bump, as additive levels have before. Reports written before it may leave it out, and no out-of-repo reader validates the level set.
• The local candidate scorer counts the 18 BRMA rows as candidate-only rows and labels them in its output.
• Certification. The full-manifest pins, the declared gate count (73) and the exclusions move, so a national or dense release certified before this merges fails validate_release after it. Sequence the merge against the staged national bundle and the frozen build.
• The planned uk_local_brma_area_layer gate was dropped. The BRMA node fails closed on any area it cannot place, or places in another FRS region.

Not yet run

• The head spine and dense builds, and the receipts they feed:
  • rent fit by nation, region and category, and the design-weight regional level check;
  • the 316 + 18 rent rows;
  • PIPR by district and bedrooms;
  • the moves in the UC housing element (all, and by tenure), Housing Benefit (obr.housing_benefit), and AHC poverty on the national and local-authority rows;
  • the mean LHA rate, and UC renters above their LHA rate against DWP (indicative: a hand-copied CIH table).
• These follow before the PR leaves draft. The receipts will carry aggregates only. The receipt tool counts carriers by distinct source household of positive weight, FRS-channel only for spine values. It withholds any rent row, cell, group or share resting on fewer than ten, with complements until the withheld cells rest on ten together. It refuses to write inside any git work tree.

Tests

• The tools' --check modes reproduce their committed resources: the area-to-BRMA layer, the Scottish overlap and the rent quantiles, run locally on 17 October. The quantiles were rebuilt for a text change after their check, and all three resources are byte-identical at this head.
• The touched tests pass locally, and the full suite runs in CI.
• Since the review, the tests also cover:
  • the summary writers' disclosure floors;
  • a Monte-Carlo check that redrawn rents placed on the census mix keep the FRS level;
  • the BRMA enum gate's structural stop;
  • grain_equal shares with the BRMA rows;
  • both measurement modes' release refusals, and the override flags in none mode;
  • the size identity moving with a rent pin;
  • the sampled-rung fallback;
  • the BRMA codes and metric;
  • a pinned digest of the supports' area-to-region pairs, which CI checks against the layer;
  • the crosswalk's layer pin without the ladder.
• Since the re-review, they also cover:
  • the production fit bindings and the terminal gate kernel on a withheld BRMA row, and the kernel failing closed when a binding writes its error;
  • an attempt to recover a withheld row from the diagnostics' totals;
  • a dense CLI run whose targets table carries a withheld row;
  • the scorer, the release contract, the sign-off drafter and Orrery's summaries with a withheld row;
  • the receipt tool's main end to end, its per-row carrier rule and every writer's work-tree guard;
  • a thin sampled region ranked on Great Britain's renters, strict_cells at both ends, and the Monte-Carlo level check at 1%;
  • build_uk_full_local_problem with a Scottish BRMA row.

🤖 Generated with Claude Code

[juaristi22]

Independent adversarial review of this PR at 29e5c5f

Run with Claude Code on 16–17 October 2026 at 29e5c5f: 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.

Leave it draft. It is not ready to merge. The BRMA layer, the rent quantile table, the Scottish overlap, the B3 composer and the rent node all check out against their sources, and every touched test passes at head 29e5c5f. Three things block it:

• The head spine and dense build has not run. Neither have its receipts. This is the PR's own condition for leaving draft.
• The Northern Ireland energy interim is retired twice. 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 both retire it, with different formulas, and María needs to pick one.
• The local compile-parity receipt is only correct on Place UK private rent and BRMA at the assigned area; bind Scotland's rents at BRMA grain (#1262, #1090) #1305's own tree. Whatever head is built on 20-21 October must carry a receipt regenerated on that head, or the release-blocking preflight refuses.

Before the receipts run, close the disclosure gaps in the receipt tool and fix the region-scale claim. Otherwise the first numbers will either be misread or be unsafe to quote.

Blocking

• The Northern Ireland energy interim is retired twice, with two formulas.

  • Verdict: CONFIRMED by 2/2 skeptics on each of 1305-c01, X-m01 and X-s02.

  • Where: packages/microcosm-build/src/microcosm/build/uk_runtime/domestic_energy_targets.py:268. This PR sets GB = UK level − NI. UK local surface before the Budget build: high-value homes OBR leg, per-vehicle fuel flag, Northern Ireland energy share and housekeeping (#1278) #1306 (domestic_energy_targets.py:281-282) sets GB = UK level × GB_sub / (GB_sub + NI).

  • What goes wrong: both PRs cite María's ruling of 16 October and neither body names the other. On the committed resources the factors are:

    • Place UK private rent and BRMA at the assigned area; bind Scotland's rents at BRMA grain (#1262, #1090) #1305: electricity 0.913257, gas 0.888649.
    • UK local surface before the Budget build: high-value homes OBR leg, per-vehicle fuel flag, Northern Ireland energy share and housekeeping (#1278) #1306: electricity 0.915590, gas 0.889821.
    • Re-pin the UK Chronicle feed to abf0921 (#1300) #1302's interim: electricity 0.916501, gas 0.875963.

    So every local-authority electricity row moves 0.26% and every gas row 0.13%, depending on which formula survives.

  • The second merge conflicts in five files: domestic_energy_targets.py, desnz_local_energy_facts.json, ledger_fact_vendor_selections.json, test_uk_domestic_energy_targets.py and test_uk_local_rowwise.py. local_binding_adjudications.json is byte-identical in both PRs and merges clean.

  • Three things merge silently:

    • uk_population_targets.json takes 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 is kept.
    • Both changelog fragments ship, and they contradict each other: 1262-uk-energy-northern-ireland-level.changed.md and 1278-northern-ireland-energy-share.changed.md.
    • Place UK private rent and BRMA at the assigned area; bind Scotland's rents at BRMA grain (#1262, #1090) #1305's receipt (87b3eac) scales all 348 electricity rows by 0.996460 and all 346 gas rows by 1.014483. Those factors belong to the subtraction formula. UK local surface before the Budget build: high-value homes OBR leg, per-vehicle fuel flag, Northern Ireland energy share and housekeeping (#1278) #1306 never touches the receipt.

  • Which formula: the intent skeptic found that the approved plan of record for UK local surface, consumption and high-value homes: remaining items (consolidation) #1278 PR 2 specifies subtraction (its item I), 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 plan assigns item I 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, but UK local surface before the Budget build: high-value homes OBR leg, per-vehicle fuel flag, Northern Ireland energy share and housekeeping (#1278) #1306 implemented the share formula instead and recorded no reason for the change. The share formula is arguably more consistent on basis: the ratio stays inside DESNZ's sub-national figures, whereas subtraction takes NI's figures off an Energy Trends level measured at actual temperature. So this is María's ruling to make, not a bug in either PR.

  • Fix:

    • Choose the formula and the carrier.
    • Drop the energy commit from the other PR whole (6928d08 here, or cf3aae4 there), rather than resolving it hunk by hunk.
    • The surviving PR keeps its own selection label and resource header as a pair, updates both energy selector_notes, and deletes the other changelog fragment.
    • Regenerate the receipt on the chosen formula.
    • Cross-reference the choice in both bodies.

• The local compile-parity receipt is right only on Place UK private rent and BRMA at the assigned area; bind Scotland's rents at BRMA grain (#1262, #1090) #1305's tree; a frozen-build head without it refuses at preflight.

  • Verdict:
    • 1305-10 CONFIRMED by 1/1; the skeptic raised it from low to medium.
    • X-s01 CONFIRMED by 3/3, high.
    • X-m02 PLAUSIBLE, 1/2.
  • Where: packages/microcosm-build/src/microcosm/build/uk/ledger_compile_parity_local_incumbent_2025_signed_differences.json.
  • What goes wrong:
    • Re-pin the UK Chronicle feed to abf0921 (#1300) #1302 moves 3,508 UC area cells to paid claims, does not regenerate the receipt, and says "Nothing reads it at build time". That is false.
    • gates.json:43-54 declares uk_ledger_compile_parity_local_incumbent_2025 as a release-blocking preflight gate. It runs in the full and dense build's preflight node (graph_terminal.py:1022-1048) and in the release-cut certifier (certify_uk_release_cut.py:270-278, release_certification.py:332-345). It checks each signed ledger_value to abs_tol 1e-9 (gates.py:500-546).
    • The gate's own _signed_difference_check_failures, run 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, returns 4,202 failures: 3,507 changed UC ledger values, 1 stale UC row, and 694 kWh rows that 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 735820c receipt merges over Re-pin the UK Chronicle feed to abf0921 (#1300) #1302 with no conflict and fails the same way.
    • Place UK private rent and BRMA at the assigned area; bind Scotland's rents at BRMA grain (#1262, #1090) #1305 stays draft until after the ~21 October build. So a dense build from main carrying Re-pin the UK Chronicle feed to abf0921 (#1300) #1302 (with or without UK local surface before the Budget build: high-value homes OBR leg, per-vehicle fuel flag, Northern Ireland energy share and housekeeping (#1278) #1306 or Port the uk-data wave to policyengine-uk 2.134.4: stamp duty, Move to UC, national home values (sixth #1095 PR) #1308) stops with candidate_blocked_at_preflight (full_build_cli.py:1951-1966). It fails closed and before the solve. The certifier refuses only after the national build has finished.
  • Why X-m02 is only plausible: the reproduce skeptic showed that the mismatch comes from Re-pin the UK Chronicle feed to abf0921 (#1300) #1302 alone and is disclosed there. Merging with UK local release: publish a K=15, 150,000-household dense dataset in at most two builds (#1301) #1307 adds nothing new.
  • Fix:
    • Regenerate the local receipt once on abf0921, at the head that will actually be built, using the chosen NI formula.
    • Land it in Re-pin the UK Chronicle feed to abf0921 (#1300) #1302, or in a small PR merged just before build 1.
    • Correct Re-pin the UK Chronicle feed to abf0921 (#1300) #1302's "Nothing reads it" sentence, and docs/uk-chronicle-feed-repin.md:213-215.
    • 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.
    • Change line 44 of this PR's body ("regenerated once") to say so.

Should fix

• The region scale assumes the census renter mix, but placement follows the all-household mix, so the "FRS level" claim fails at design weights.

  • Verdict (1305-01): CONFIRMED by 2/2 skeptics, medium.

  • Where: private_rent_localisation.py:297-340, where scale = mean_rent / level and the level comes from census private-renter BRMA counts. Placement is in atomic_area_support.py:420-427, which draws areas by all-household count within the FRS region and has no tenure term.

  • What goes wrong: from committed public resources only, Scotland's realised over assumed level is:

    • by band: B 0.909, C 0.940, D 1.004, E 0.966;
    • overall about 0.953, with a total variation distance of 0.122 between the two mixes;
    • Lothian holds 19.3% of renters but 13.1% of households, and Greater Glasgow 15.9% against 13.0%.

    The LSOA-count proxy elsewhere gives London −3.7%, North West −1.8%, Wales −1.4% and East of England +1.8%. A toy run through localise_rents, with 50/50 placement against an 80/20 renter split, gives 0.833 of the FRS mean; placing rows on the renter mix gives 1.000.

  • Texts that state the false consequence: graph_population.py:707, the module docstring at private_rent_localisation.py:9-12, docs/uk-full-build-graph.md:31, changelog.d/1262-uk-private-rent-from-area.added.md and the body. The receipt's region_mean_rent_after_over_before (tools/uk_private_rent_receipt.py:730-741) will show about 0.95 for Scotland, and readers will take that as the effect of the lists.

  • What limits it: weights are untouched, and the local-authority tenure/private_rent rows pull the mix back during calibration. The toy tests (test_uk_private_rent_localisation.py:214, :357-389) already place renters on the census mix, so they cannot show the gap.

  • Fix, either one:

    • Compute s_r on the pool at the node: m_r divided by the weighted mean of Q over FRS-channel rows at their assigned BRMAs. This is exact at design weights but K-invariant only in expectation, so it gives up what test_localise_rents_does_not_depend_on_k pins.
    • Or keep the census mix on purpose and reword the five texts to say the level holds once calibration restores the renter mix. Then add the design-weight regional level to the receipts as a named check.

• The 18 Scottish BRMA rows bind roughly the list (PIPR) level that the rent node removes.

  • Verdict (1305-03): PLAUSIBLE, 1/2 skeptics.
  • Where: ledger_targets.py:2272-2408, where total = months * mean * renter_households (:2359) with no levelling, against private_rent_localisation.py:332.
  • What goes wrong:
    • Against the SG 2025 published means, PIPR 2025 S33 means range from −16.5% to +16.4% per BRMA. The mean gap is about 0, so the list level is about 0.99 of PIPR.
    • The node holds Scotland at the FRS level. With tenure counts bound, calibration can close the gap only by shifting Scottish weight toward high-rent renters within each BRMA.
  • Why only plausible: the intent skeptic found the trade-off on record, in UK private rent at the assigned area (#1123 item 4, "PR B"): what is in place and the pieces still missing #1262 A3/A4, María's uk-data#521 comments of 13 October (Q4) and 16 October, and commit 609ebfd. The 316 England and Wales LA rows carry the same tension. What remains is that nothing in Place UK private rent and BRMA at the assigned area; bind Scotland's rents at BRMA grain (#1262, #1090) #1305 says these rows sit above the node's level.
  • What would settle it: the fit of the 18 rows in the head dense build, plus the shift in Scottish renter weights and in UC/HB exposure.
  • Fix: add one line to "For review" and to the cross-grain routes (cross_grain_declarations.json:612, :648) declaring the expected level miss. Or bind the rows as a pattern, rescaling the 18 targets to Scotland's FRS-level total, and decide the same for the LA rows.

• Under --target-weight-rule grain_equal, "brma" becomes a grain of its own and takes 25% of the loss.

  • Verdict (1305-02): CONFIRMED by 2/2 skeptics, medium.
  • Where: local_rowwise.py:1273-1276, which takes raw area_type labels, and :1298-1302, which passes them unmapped and skips the floor check. The same unmapped path appears at :2074-2077 (holdout) and at target_weights.py:507-510 (uk_rule_loss_weights). UK_LOCAL_WEIGHT_GRAIN is applied only in _rule_grains (target_weights.py:381-392).
  • What goes wrong: with realistic counts (650 constituency, 361 la, 18 brma and 364 national labels), each grain takes 0.25 of the loss, and a BRMA row weighs about 20 times an LA row. The national share also drops from 1/3 to 1/4, so a grain_equal arm on the head is not comparable with main's. The dense posture allows grain_equal (rowwise_posture.py:270-274). Release candidates refuse non-default rules (rowwise_cli.py:676-678), so no release is affected. The doc at docs/uk-scotland-private-rent-brma-1090.md:159-162 and the body state the mapping without that condition.
  • Fix: map the labels through UK_LOCAL_WEIGHT_GRAIN in all three places, and keep the raw labels for the target_rows.grain alignment check. Add a test that grain_equal weights do not change when the 18 brma rows are added.

• The structural stop on the new gate uk_local_brma_enum_domain is untested.

  • Verdict (1305-04): CONFIRMED by 2/2 skeptics, medium; one skeptic rated it low.
  • Where: full_gates.py:336-340 (structural_local).
  • What goes wrong: with the gate patched out of structural_local in memory, the result was unenforced_release_failures at f0.1 and exportable_blocking at f1.0, with artifact_permitted=True in both, and all 7 existing classification cases still passed.
    • The parametrize list at test_uk_full_gates.py:333-352 leaves the gate out.
    • test_uk_battery_bindings.py:209-214 stubs enum_domain to pass, so the assertion at :978-981 proves nothing.
  • Narrowed:
    • Broken terminal evidence fails closed (gate_battery.py:645-722).
    • A release would still be blocked by the dense contract's _UK_DENSE_RELEASE_BLOCKING_IDS (contract.py:753-764, 4697-4702).
    • So the worst case is a diagnostic candidate exported with an off-domain BRMA.
  • Fix: add the gate to the parametrize list, and add a terminal-phase test that runs the unstubbed binding on a frame with an off-domain BRMA and asserts FAILED and a structural stop.

• The functions that write the quotable private_rent_summary.json have no tests.

  • Verdict (1305-07): CONFIRMED by 2/2 skeptics, medium.
  • Where: tools/uk_private_rent_receipt.py:655-679 (measure_households), which calls rents_by_cell (:358-385), pipr_comparison (:475-534), weight_concentration (:582-596) and coherence (:599-618).
  • What goes wrong: the docstring (:31-41) says this file is the only one that may be quoted in a PR or a committed receipt, and that every cell rests on at least 10 distinct source households. But test_uk_private_rent_receipt_tool.py tests only the building blocks, and nothing sanitises the output before the write at :800-801. Synthetic probes reproduced three defects, listed in the next four bullets. rents_by_cell's suppression and pipr_comparison's cell exclusion behaved correctly in the probes, but no test pins them.
  • Fix: write synthetic-frame tests on measure_households' summary covering:
    • a region × category cell under 10 source households reads "<10" and withholds a complement;
    • a district cell under 10 is excluded;
    • weight_concentration and coherence withhold figures below 10 source households;
    • a group with fewer than 3 cells is withheld;
    • a mixed-channel spine-values cell.

• top_1pct_weight_share has no record or household floor.

  • Verdict (1305-08): CONFIRMED by 1/1, low.

  • Where: tools/uk_private_rent_receipt.py:582-596, where top = max(1, ceil(0.01 * n)). :670 puts it in the summary.

  • What goes wrong:

    • With 41 rows it prints a single row's weight share.
    • With 900 rows it rests on 9 rows.
    • With 5,000 rows whose top 50 are clones of 2 households, it still prints.

    It bites only on small or subsampled runs, but nothing enforces that limit.

  • Fix: print "<10" unless the top slice has at least 10 rows spanning at least MIN_SOURCE_HOUSEHOLDS distinct source_key values. Or move the figure to the audit file.

• The coherence unweighted share floors on clone rows, not source households.

  • Verdict (1305-23): CONFIRMED by 1/1, low.
  • Where: tools/uk_private_rent_receipt.py:614, floored_share(mask.sum(), mask.size). weighted_share at :144-160 counts distinct keys instead.
  • What goes wrong:
    • 8 of 800 households × 25 clones gives weighted "<10" but unweighted 0.01.
    • A complement case (all but 3 of 100 households) gives 0.97.
  • Fix: pass the carriers, and print "<10" when either side has 1-9 distinct source keys.

• pipr_comparison summarises groups of one or two district cells.

  • Verdict (1305-24): CONFIRMED by 1/1, low.
  • Where: tools/uk_private_rent_receipt.py:517-529 sets no minimum cell count. Only _spearman (:467-472) refuses below 3.
  • What goes wrong: in a one-cell group, median_ratio is that district's model/PIPR ratio. Multiplying it by the public PIPR mean gives back the model district mean; in a toy case, 1.25 × 800 = 1000. That breaks the docstring's "no figure describes a single district or BRMA" and the sub-GOR rule.
  • Fix: withhold the group's statistics unless at least 3 cells pass, and preferably 10.

• Spine-values cells count SPI copies under their FRS parent's key.

  • Verdict (1305-25): CONFIRMED by 1/1, low.
  • Where: tools/uk_private_rent_receipt.py:345-355 (_mean_cell counts source_key.nunique() across channels) and :711-714 (the spine block). Related code: spi_support.py:260-266, :726-746, and spi_housing_shell.py:505-552, which redraws SPI tenure, bedrooms and rent.
  • What goes wrong: a toy cell with 4 FRS-channel renters and 6 SPI copies reports 10 carriers and prints a mean. Narrowed: the QRF interpolates across many donors rather than copying a few. Still, fewer than 10 real reported rents can pass the floor.
  • Fix: for the spine block, count carriers over FRS-channel rows only, or drop spine-values rents_by_cell from the summary.

• Licensed outputs have no guard against landing in the repo.

  • Verdict (1305-26): CONFIRMED by 1/1, low (defensive).
  • Where: tools/uk_private_rent_pool.py:117-119 writes pool.npz by default to <run>/audit-1262/. It holds per-household rent, subrent, reported HB, weights and source keys. Also tools/uk_private_rent_receipt.py:745-747, :774, :800-801, and engine_rent_diagnostics.py:208, :268-269.
  • What goes wrong: .gitignore covers *.h5, out/ and build/, but not *.npz, audit-*/ or runs/. A run directory inside the checkout plus a later git add -A would stage unit-record files. The guard pattern already exists at tools/build_uk_efrs_parity_reference.py:565, :623.
  • Fix: refuse output paths under the repo root, and add *.npz and audit-*/ to .gitignore.

• "Not yet run" omits UK private rent at the assigned area (#1123 item 4, "PR B"): what is in place and the pieces still missing #1262's acceptance measures.

  • Verdict (1305-06): CONFIRMED by 2/2 skeptics, low.
  • Where: the body's "Not yet run" section, against UK private rent at the assigned area (#1123 item 4, "PR B"): what is in place and the pieces still missing #1262's Acceptance and D3. docs/evidence/uk-private-rent-1262/scripts/engine_rent_diagnostics.py:1-15, :132-141 already computes these measures. The LA-grain AHC rows (local_targets.py:72-81, 440-447) move as well.
  • What goes wrong: the PR closes UK private rent at the assigned area (#1123 item 4, "PR B"): what is in place and the pieces still missing #1262 without listing the moves in the UC housing element, Housing Benefit (obr.housing_benefit) and AHC rows as owed. The changelog does say the boarder, lodger, subletting, service-charge and reported HB columns stay as reported; only the body is silent.
  • Fix: add those moves to "Not yet run", covering both national and LA-grain AHC, and repeat the "stay as reported" line in the body.

Minor

• The toy-graph test comment contradicts its own data, and the scale test re-derives the formula.

  • Verdict (1305-05): CONFIRMED by 2/2 skeptics, low.
  • Where: test_uk_private_rent_localisation.py:372-375 says the region mean is unchanged, but the toy renter goes from 9,000 to 5,271.02, which is 9,000 × (0.5 + u).
  • What goes wrong: a kernel that writes the rent back unchanged passes this test, though the London test catches it. A category-blind π passes the scale and redraw tests, because the toy gives every category the same mix (uk_atomic_support_fixtures.py:319).
  • Fix: correct the comment, assert that the rent changes, give the toy category-specific mixes, and add a property test that the weighted redrawn mean holds within Monte-Carlo tolerance.

• The London toy test's ratio check is conditional.

  • Verdict (1305-19): CONFIRMED by 1/1, low.
  • Where: test_uk_private_rent_localisation.py:411-424. The if len(by_brma) == 2 branch and the subset assertions all hold on an empty frame.
  • What goes wrong: with all London clones in one BRMA, a write-back kernel passes everything, and so does the Lothian test. Seed 7 splits them 2/4 today.
  • Fix: assert len(london) == 6 and len(by_brma) == 2 unconditionally, and that the rents differ from 20,000.

• A stale signed UC rationale is carried through the regenerated receipt.

  • Verdict (1305-09): CONFIRMED by 1/1, low.
  • Where: tools/build_uk_ledger_compile_parity_signed_differences.py:428-447.
  • What goes wrong:
    • The text claims a uniform fixture/ours of 0.8925-0.9003 across 982 rows. At the head the 980 rows span 0.793-1.048, with p5-p95 0.937-0.973 and 3 rows above 1.
    • For children rows the text claims p5-p95 of 0.77-1.19; the actual is 0.815-1.237.
    • The text was already wrong at the base; this PR re-signs it.
  • Fix: restate the rationale from the regenerated rows, or compute the band in the tool, and say what residual gap the paid-claims basis leaves.

• "Certification part digests do not move" understates the effect.

  • Verdict (1305-11): CONFIRMED by 1/1, low.
  • Where: the body's line 15 and the changelog, against contract.py:3191-3200.
  • What goes wrong: the check compares spec.{gates_manifest_sha256, policy_sha256, spec_fingerprint} with the full-manifest pins this PR moves (:450-458). It also checks declared_entry_count, now 73, and the exclusions, which gain uk_local_brma_enum_domain (:811-814). Any national or dense release certified before the merge therefore fails validate_release afterwards (:6115-6130).
  • Fix: say so in the body, and sequence the merge against the staged national bundle and the frozen build.

• --target-geographies cannot name brma.

  • Verdict (1305-12): CONFIRMED by 1/1, low.
  • Where: full_build_cli.py:399-411, graph_build.py:181-190, and graph_targets.py:124, :194, :405-407. target_selection.py:83-110 drops the rows silently.
  • What goes wrong: "country,region,constituency,la" used to equal "all" and now drops the 18 rows. The release check (rowwise_cli.py:701-708) still passes. The default is unaffected.
  • Fix: accept brma in the parser, config and help text, or imply it from la. At minimum, receipt the dropped rows.

• Cross-border BRMAs mix list vintages under one region scale.

  • Verdict (1305-13): CONFIRMED by 1/1, low.
  • Where: tools/build_uk_brma_rent_quantiles.py:22-24, :556-559, and private_rent_localisation.py:297-339, :388-405.
  • What goes wrong:
    • 11 Welsh LSOAs in WEST_CHESHIRE read the newer English list under the Wales scale.
    • 4 South West LSOAs in MONMOUTHSHIRE and 1 West Midlands LSOA in BRECON_AND_RADNOR read the older Welsh list under English scales. These 5 are the VOA overrides at tools/build_uk_brma_area_layer.py:120-124.
    • A synthetic 5% vintage gap passes straight through. About 16 LSOAs are affected.
  • Fix: uprate the Welsh list to England's window, or qualify the "a nation's factor cancels" text and name these BRMAs.

• --brma-localisation none still binds the BRMA rows, on spine labels unrelated to placement.

  • Verdict (1305-14): CONFIRMED by 1/1, low.
  • Where: full_problem.py:89-99. The spine BRMA comes from frs_brma.py:79-91 (keyed on benunit id, region × category only), there is no node in none mode (graph_population.py:1236-1239), and rowwise_cli.py:695-697 refuses none only for release candidates.
  • What goes wrong: the docstring at local_targets.py:134-147 is wrong in this mode. The CLI default is "area", and only the config falls back to none on the legacy assignment.
  • Fix: drop the brma grain, or fail closed, when the mode is none.

• The private-rent node can refuse small dev rungs (f001).

  • Verdict (1305-15): CONFIRMED by 1/1, low; the skeptic had medium confidence.
  • Where: private_rent_localisation.py:269-280, :333-337, :385-387. Sampling stratifies by region only (national_sampling.py:236-245), and SPI renters are redrawn independently.
  • What goes wrong: a synthetic demo shows both refusals. How often this happens at real f001 needs licensed counts.
  • Fix: below f100, fall back to a GB-wide pool with a receipt, or resolve f001/f010 to none with a receipt. Keep f100 strict.

• The layer-vs-supports region test always skips in CI.

  • Verdict (1305-17): CONFIRMED by 1/1, low.
  • Where: test_uk_brma_area_layer.py:314-337. build/ is gitignored and no CI job mounts the supports.
  • Mitigation:
    • The provenance sha pins run in CI (:94-108).
    • Release candidates check support digests (full_build_cli.py:1013-1020).
    • The invariant passes today on the pinned supports.
    • A drift fails closed at brma_area_layer.py:276-281 during the head build.
  • Fix: commit a compact region-by-area digest that an engine-free test can compare, or run the test in a job that has the supports.

• The body claims the tools check their committed resources against a rebuild.

  • Verdict (1305-18): CONFIRMED by 1/1, low.
  • Where: the body's line 56.
  • What goes wrong: only test_uk_scotland_brma_la_overlap.py:274 runs a rebuild, and it skips without the feed. No test runs the layer or quantile --check, and no record shows one was run.
  • Fix: reword to say the tools have a --check mode run locally on a given date against a given commit, and record that run.

• New CLI, size-identity and weighting paths are untested.

  • Verdict (1305-20): CONFIRMED by 1/1, low.

  • What goes wrong: a line trace over about 30 test files shows these lines never run:

    • the refusal raises at full_build_cli.py:909, :968;
    • the rowwise_cli.py:696, :699 refusals;
    • size_checkpoint.py:149-155, :158-165;
    • the brma branch at full_problem.py:93;
    • local_targets.py:144-147 and :353-360.

    _rule_grains never saw a brma label.

  • Fix: tests for the release-candidate override refusal, both "none" refusals, the size identity changing when a rent pin changes, brma grains in the target weights, and a local problem with a brma grain.

• The Scotland doc promises per-area correction records the tool does not write.

  • Verdict (1305-22): CONFIRMED by 1/1, low.
  • Where: docs/uk-scotland-private-rent-brma-1090.md:21-27.
  • What goes wrong: the tool writes only counts (tools/build_uk_brma_area_layer.py:533-539): 9 filled, 59 corrected, 2 overridden. Measure Scotland's BRMA × council overlap; keep BRMA rents out of council targets #1092's 61 is probably 59 + 2.
  • Fix: state the counts and where they live, or write the list of areas into the provenance. Area codes are geography, so listing them raises no disclosure problem.

• The 18 new BRMA rows need a carrier floor before any BRMA-level diagnostics leave the build. This one is María's call.

  • Verdict (1305-27): CONFIRMED by 1/1, low; the skeptic had medium confidence.
  • Nobody has measured carriers per BRMA yet. See the Disclosure check below.

• Nits, each CONFIRMED by 1/1:

  • No schema bump for the new diagnostics level (1305-16). schema.py:22, :236-258: older readers reject brma rows. All in-repo readers are updated, and there is precedent for additive changes without a bump (21a4d3a). Say in the body that no out-of-repo reader validates the level set.
  • The scorer counts the 18 BRMA rows as candidate-only without saying so (1305-21). tools/score_uk_local_candidate.py:41-47, :331-333. The count move is disclosed and the scorer's output labels the rows. Add one line to "For review".
  • A cross-grain reason names the wrong composer (1305-28). cross_grain_declarations.json:648 cites uk_private_rent_mean_to_total; it should cite uk_private_rent_brma_mean_to_total (ledger_targets.py:2272).
  • Override and sha flags are silently ignored when the mode is none (1305-29). full_build_cli.py:898-900, :939-951. The property path already behaves this way. Refuse the flags in none mode.
  • The crosswalk's area_layer_sha256 is checked only by a test that skips in CI (1305-30). local_area_crosswalk.json:2136-2138; the test is at test_uk_local_area_crosswalk.py:34-43 and needs the ladder. The pin matches today. Add a ladder-free assert, or drop the field.
  • Two English UC fixture_only cells carry the Northern Ireland rationale (1305-31). tools/build_uk_ledger_compile_parity_signed_differences.py:548-549, :842-843. They belong to the support-floor deferrals, not the Northern Ireland class. This predates the PR. Route support-floor deferrals before the per-metric maps.
  • The parity text names the PIPR September edition, but the selector resolves the July record set (1305-32). data_target_parity.py:369, uk_data_target_parity.json:947. All 2025 values are identical across the two editions, and the error predates this PR (04f7202). Correct the wording.
  • The DWP above-LHA shares are hand-copied from a quote on uk-data#521 (1305-33). engine_rent_diagnostics.py:55-61; no Chronicle fact exists. Register the table in Chronicle, or cite the primary table and mark the comparison indicative.
  • "Rounded counts only" is not quite true (1305-34). At graph_population.py:709-710, two counts in the receipt are exact (:778-785), but they count cells and regions, not records. Reword to "rounded record counts only".

Cross-PR

• The Northern Ireland energy formula (Place UK private rent and BRMA at the assigned area; bind Scotland's rents at BRMA grain (#1262, #1090) #1305 × 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 the receipt merge order (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). Both are under Blocking.

• Battery digests and the local receipt collide with UK local release: publish a K=15, 150,000-household dense dataset in at most two builds (#1301) #1307 and Port the uk-data wave to policyengine-uk 2.134.4: stamp duty, Move to UC, national home values (sixth #1095 PR) #1308.

  • Verdict (1305-c02): CONFIRMED by 1/1, low.
  • Where: contract.py:450-458 and :766-770.
  • What goes wrong:
    • Place UK private rent and BRMA at the assigned area; bind Scotland's rents at BRMA grain (#1262, #1090) #1305, UK local release: publish a K=15, 150,000-household dense dataset in at most two builds (#1301) #1307 and Port the uk-data wave to policyengine-uk 2.134.4: stamp duty, Move to UC, national home values (sixth #1095 PR) #1308 each rewrite the three full-battery pins from the same old values to different new ones.
    • 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 both rewrite the dense digests.
    • All three PRs edit gates.json from the same base blob, so neither side's pin is right on the combined tree. test_gate_battery_contract_pins.py:70-81 would fail, which is fail-closed.
    • 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 both regenerate the local receipt from the same base, with compiled counts of 25,327 and 22,986. UK local release: publish a K=15, 150,000-household dense dataset in at most two builds (#1301) #1307's is built on 735820c.
  • Fix: agree a merge order for 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, UK local surface before the Budget build: high-value homes OBR leg, per-vehicle fuel flag, Northern Ireland energy share and housekeeping (#1278) #1306, UK local release: publish a K=15, 150,000-household dense dataset in at most two builds (#1301) #1307 and Port the uk-data wave to policyengine-uk 2.134.4: stamp duty, Move to UC, national home values (sixth #1095 PR) #1308 before 20 October. The last PR to land regenerates the digests and the receipt on the combined tree.

• 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: the right merged counts match neither side.

  • Verdict (X-m04): CONFIRMED by 1/1, low.
  • What goes wrong:
    • local_target_references.json auto-merges to 23,004: 25,327, minus UK local release: publish a K=15, 150,000-household dense dataset in at most two builds (#1301) #1307's 2,323 high-value-homes cells.
    • The membership conflicts only in its header hunk (lines 4-16); the body merges correctly.
    • The correct merged values are:
      • membership: active 23,004, candidate 27,133, contract 59;
      • battery pins: 23,004 / 27,133, with 4,129 exclusions;
      • population pins: 418 / 359 / 59.
    • A header taken from one side passes CI, and the scorer then refuses every candidate (tools/score_uk_local_candidate.py:406).
    • Taking UK local release: publish a K=15, 150,000-household dense dataset in at most two builds (#1301) #1307's side of score_uk_local_candidate.py drops Place UK private rent and BRMA at the assigned area; bind Scotland's rents at BRMA grain (#1262, #1090) #1305's statistical_scope and brma entries from _LOCAL_GRAINS.
  • Fix: regenerate the membership and the receipt on abf0921 with the merged generators. Keep UK local release: publish a K=15, 150,000-household dense dataset in at most two builds (#1301) #1307's dynamic count together with Place UK private rent and BRMA at the assigned area; bind Scotland's rents at BRMA grain (#1262, #1090) #1305's grains.

• Some digests must be recomputed, not taken from either side.

  • Verdict (X-m05): CONFIRMED by 1/1, low.
  • What goes wrong: in 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 × Place UK private rent and BRMA at the assigned area; bind Scotland's rents at BRMA grain (#1262, #1090) #1305, gates.json auto-merges into a third file, so the full-battery pins must be recomputed. The dense digests need recomputing only in 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, because Port the uk-data wave to policyengine-uk 2.134.4: stamp duty, Move to UC, national home values (sixth #1095 PR) #1308 touches only uk_export_surface, which is outside the dense scope. Stacked PRs get no CI until they are retargeted.
  • Fix: whoever lands second recomputes from the merged gates.json and runs test_gate_battery_contract_pins.py locally.

• Stale Northern Ireland interim text survives each retirement.

  • Verdict (X-s06): CONFIRMED by 1/1, low.
  • What goes wrong:
    • On Place UK private rent and BRMA at the assigned area; bind Scotland's rents at BRMA grain (#1262, #1090) #1305 alone, uk_population_targets.json:31688 and :31727 still say Northern Ireland is "taken out at its 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", and no generator rewrites them.
    • 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 (domestic_energy_targets.py:13-20) is stale in the other direction.
    • docs/uk-chronicle-feed-repin.md:199-202 is unchanged in every tree. It reads as a dated record, so it is the weakest of the three.
  • Fix: whichever PR keeps the retirement updates these texts.

Disclosure check

What the disclosure lens checked:

• the body;
• all 8 commit messages;
• the 5 changelog fragments;
• docs/uk-scotland-private-rent-brma-1090.md and the experiments/1090 script and JSON;
• every committed resource: the layer and its provenance, the quantiles, the overlap, the compile-parity receipt, references, membership, population targets, cross-grain, parity, census and the DESNZ facts;
• kernel receipts and error messages;
• all new test fixtures, which are synthetic.

Every figure comes from Census/NRS, DESNZ/ONS, VOA, Rent Officers Wales, SG, RSS, PIPR, DWP admin counts or target-registry counts. The "30.9%" in 6421fe4 comes from Rent Service Scotland FOI lettings: admin data, outside UKDS scope. No committed content carries a FRS/LCFS/WAS/SPI/NTS figure, a record id, a per-record value or an extreme. The FRS level is applied only inside the kernel, the region scales and ranks never leave it, and the receipts round record counts.

Gaps, all in tooling that has not run yet, so nothing has been published:

• 1305-07: no tests on the summary writers.
• 1305-08: top_1pct_weight_share has no floor.
• 1305-23: the unweighted share counts clone rows.
• 1305-24: one-cell PIPR groups are summarised.
• 1305-25: SPI copies are counted as carriers.
• 1305-26: no repo-path guard on licensed outputs.

Close these before the receipts run. The PR plans to quote rent fit by region and category, and PIPR by district and bedrooms, from exactly this file.

For María to decide (1305-27): the 18 new BRMA rows need a carrier floor before any BRMA-level diagnostics leave the build.

The figures in this review are synthetic test values, counts of targets, areas or tests, or come from public census, rent-officer list, PIPR, SG, DESNZ and DWP sources.

Tests and tools run

All runs were at head 29e5c5f.

• Correctness lens:
  • Rent localisation, BRMA layer, BRMA graph, quantile and Scottish overlap tests (test_uk_private_rent_localisation, test_uk_brma_area_layer, test_uk_brma_area_graph, test_uk_brma_rent_quantiles, test_uk_scotland_brma_la_overlap): 95 tests, 0 failures, 3 skipped (supports not mounted, feed absent).
  • The other touched test files (test_uk_domestic_energy_targets, test_uk_local_rowwise, test_uk_diagnostics, test_uk_local_target_census, test_uk_population_targets, test_uk_pipr_deferrals, test_uk_local_gate_battery, test_uk_private_rent_receipt_tool, test_uk_brma_area_layer_tool): 167 tests, 0 failures.
  • Build-preparation tests (test_uk_full_build_preparation, test_uk_size_l2, test_uk_rowwise_candidate, test_uk_local_candidate_scorer): 89 tests, 0 failures.
  • A check of the layer against the pinned supports: 0 missing, 0 extra, 0 region mismatches.
  • A check of the region levels: every GB region × category cell resolves.
  • A synthetic grain_equal weighting probe.
• Method lens:
  • Rent localisation, quantile, layer, domestic energy and Scottish overlap tests: 100 tests, 0 failures, 3 skipped.
  • The engine test tests/engine/uk/test_uk_brma_area_layer.py under policyengine-uk 2.132.6: 3 tests, 0 failures.
  • test_uk_ledger_targets, test_uk_local_rowwise and test_uk_pipr_deferrals, filtered with -k 'brma or partial or pipr or private_rent': 14 tests, 0 failures.
  • Read-only checks on committed public resources: the NI energy level factors, the Scottish and England and Wales tenure-mix level ratios, and PIPR against SG means by BRMA.
• Artifacts lens:
  • Resource tests (the BRMA layer and layer tool, quantiles, Scottish overlap, rent localisation, domestic energy, local target census, PIPR deferrals and population targets, plus shared/test_country_spec): 291 tests, 0 failures, 3 skipped.
  • Pin tests (shared/test_gate_battery_contract_pins, test_uk_local_gate_battery, test_uk_local_small_cell_rule): 37 tests, 0 failures.
  • The engine layer test under policyengine-uk 2.132.6: 3 tests, 0 failures.
  • tools/census_uk_local_targets.py --check and tools/census_uk_data_target_parity.py --check: both current.
  • tools/ci_test_plan.py verify: ok (735 files).
  • A consistency check of the quantile table and the layer (Scottish facts against the feed; monotone knots, means and bases; the layer's BRMA set against the quantile table's): 288 checked, 0 mismatches.
  • A check of the layer's BRMA names against BRMAName and LHA rates under policyengine-uk 2.134.5: passes.
• Tests lens:
  • New BRMA and rent test files (test_uk_brma_area_graph, test_uk_brma_area_layer, test_uk_brma_area_layer_tool, test_uk_brma_rent_quantiles, test_uk_private_rent_localisation, test_uk_private_rent_receipt_tool, test_uk_scotland_brma_la_overlap): 112 tests, 0 failures, 3 skipped.
  • Modified test files, first set (test_uk_diagnostics, test_uk_domestic_energy_targets, test_uk_ledger_targets, test_uk_local_candidate_scorer, test_uk_local_gate_battery, test_uk_local_rowwise, test_uk_local_target_census, test_uk_pipr_deferrals, test_uk_population_targets, test_uk_rowwise_candidate, test_uk_full_gates): 312 tests, 0 failures.
  • Modified test files, second set (test_uk_battery_bindings, shared/test_country_spec, test_uk_orrery_contract, test_uk_size_l2, test_uk_full_target_graph, test_uk_take_up_gate): 239 tests, 0 failures.
  • Tests that cover the changed code (test_uk_target_weights, test_uk_target_weight_wiring, test_uk_full_measure, test_uk_local_targets, test_uk_cross_grain_declarations, test_uk_calibration_run, test_uk_release_certification, test_uk_local_small_cell_rule, test_uk_full_population_graph, test_uk_graph_terminal, test_uk_property_localisation): 157 tests, 0 failures.
  • microcosm-data shared/test_contract.py and microcosm-diagnostics shared/test_schema.py: 343 tests, 0 failures.
  • Engine tests (test_uk_brma_area_layer.py, test_uk_full_graph_admission.py): 4 tests, 0 failures.
  • A synthetic toy-node probe: 2 passed.
  • ruff check and ruff format --check on the 63 changed .py files: clean.
• Claims lens: test_uk_domestic_energy_targets.py, test_uk_scotland_brma_la_overlap.py, test_uk_brma_rent_quantiles.py, test_uk_pipr_deferrals.py and shared/test_gate_battery_contract_pins.py: 77 tests, 0 failures, 1 skipped (the feed-gated overlap rebuild).
• Disclosure lens: test_uk_private_rent_receipt_tool.py and test_uk_private_rent_localisation.py: 29 tests, 0 failures, plus synthetic disclosure probes.
• Second fresh-eyes pass: the engine layer test (3 tests, 0 failures) and the engine-free rent localisation, BRMA graph, layer, quantile and domestic energy tests (80 tests, 0 failures, 2 skipped).
• Skeptics:
  • 1305-04: 27 tests, 0 failures, plus a mutation check (all 7 classification cases pass against the mutant).
  • 1305-05: 7 passed.
  • 1305-17: 26 tests, 2 skipped.
  • 1305-18: 9 tests, 1 skipped.
  • 1305-21 (the scorer): 5 tests, 0 failures.
  • 1305-30: 1 test, skipped (it needs the ladder).
  • 1305-19: 5 tests, 0 failures, plus 1 further test, 0 failures.
  • A line trace over about 30 test files: 4 batches, all exit 0.
  • git merge-tree simulations of the cross-PR merges, and fixed-string searches of the feed.

Checked and found sound

• Layer: it covers the pinned supports exactly (35,672 England and Wales LSOAs, 46,363 Scottish OAs), with 0 region mismatches. All 192 GB BRMA names are in BRMAName under policyengine-uk 2.132.6 and 2.134.5, and all have LHA rates.
• BRMA roster: the layer, the quantile table and the count table agree. The 232 region × BRMA pairs match exactly.
• Quantiles: knots are monotone. Each stored mean equals knot_mean. Scottish knots equal the published quartiles. The 288 Scottish published values match the abf0921 facts. Medians rise B ≤ C ≤ D ≤ E in every BRMA. No cell is modelled with fewer than 20 lets.
• Rent node mechanics:
  • Units run weekly to annual through WEEKS_IN_YEAR on both sides of the scale.
  • The rank formula and the tie-break are K-invariant and do not depend on row order.
  • Northern Ireland, owners, social renters and zero rents are untouched.
  • Clones share a rank.
• B3 composer: BRMA renters add up to the councils' reconciled total. All 32 council tenure cells are active. It runs after reconciliation.
• No double counting: BRMA rows are Scotland-only, LA rent rows are England and Wales only, and Scottish council rent cells stay deferred.
• Partial grain: the −1 area index is only ever compared with ==, never used to index, so nothing wraps onto the last BRMA.
• Default weight rule: BRMA rows weigh inside the LA private-rent family, and the energy and bus floors are unaffected.
• NI energy change: the units are consistent, and the level moves (−0.354% electricity, +1.448% gas) match the body.
• Membership, references and receipt counts are internally consistent.
• Resources: country_package.json lists all four new resources, and the retired items leave no dangling references.
• Test changes: no assertions were deleted or loosened, no new test reads the clock, and the CI test plan verifies.

Refuted during verification

No finding was refuted outright. These parts were dropped or narrowed:

• 1305-03 "undeclared": the level trade-off is on record in UK private rent at the assigned area (#1123 item 4, "PR B"): what is in place and the pieces still missing #1262 A3/A4, uk-data#521 and 609ebfd. What is missing is only a statement in Place UK private rent and BRMA at the assigned area; bind Scotland's rents at BRMA grain (#1262, #1090) #1305.
• 1305-04 "broken evidence fails open": it fails closed. A release is still blocked through the dense contract.
• 1305-05 "a no-op kernel passes the suite": the London test catches it, though only that one test does.
• 1305-02 "about 890×": that came from synthetic counts. With realistic counts it is about 20×, though the 25% grain share holds either way.
• 1305-06 "the reported-column incoherence is never stated": the changelog states it; only the body is silent.
• 1305-21, the validation-level selector re-point: nothing reads that selector, and the old record set is absent from the feed.
• 1305-25 "the same few donors": QRF draws interpolate across many FRS records.
• 1305-14: the CLI default is "area", not "none".
• 1305-16: no out-of-repo reader validates the level set.
• 1305-c01's conflict list: local_binding_adjudications.json does not conflict; test_uk_local_rowwise.py does.
• X-s02's 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): the approved plan of record for UK local surface, consumption and high-value homes: remaining items (consolidation) #1278 PR 2 specifies Place UK private rent and BRMA at the assigned area; bind Scotland's rents at BRMA grain (#1262, #1090) #1305's subtraction formula.
• X-m02: the receipt mismatch is Re-pin the UK Chronicle feed to abf0921 (#1300) #1302's alone and is disclosed; the UK local release: publish a K=15, 150,000-household dense dataset in at most two builds (#1301) #1307 merge adds nothing.
• X-s01 "a third build": the dense preflight refuses before the solve, so no solve is wasted.
• X-m04 "no CI cross-check": taking either side's whole file is caught. Only a wrong header count slips through.
• X-m05: dense digests need recomputing only in 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.

Not covered

• The claim that each commit's touched tests pass. Every lens checked the head only.
• build_uk_brma_area_layer.py --check and build_uk_brma_rent_quantiles.py --check were not run (they need geopandas, raw registrations and a full feed stream).
• The logic of experiments/1090-scotland-brma-la-rent-dispersion.py. Only its JSON receipt was checked.
• Whether the rent-officer lists exclude tenancies paid by HB or UC. If they do, the within-region pattern the node imposes leans toward lets without a claimant.
• How the redrawn rents interact with reported HB/UC amounts and take-up inside policyengine-uk, beyond confirming that benunit_rent derives from household rent.
• The head-build receipts. That includes the real Scottish weight shift behind 1305-03, f001 refusal rates (1305-15) and carriers per BRMA (1305-27).
• A wider claims-lens test run, which never ran.
• An untraced lead from one skeptic, not a finding: on Place UK private rent and BRMA at the assigned area; bind Scotland's rents at BRMA grain (#1262, #1090) #1305, the regenerated receipt's 694 DESNZ kWh ledger_values differ from the membership's resolved_values, by about ×0.986 for gas and ×1.004 for electricity. On main the two agree. It is worth checking whether the membership carries levelled energy values and needed regenerating after 6928d08. The feed-gated currency test that would catch this skips in CI.
• Whether all 59 Balloch corrections really are around Balloch (1305-22).

🤖 Generated with Claude Code

[juaristi22]

Addressed in c5ec9f6, rebased onto main after #1302 merged. Item by item:

Blocking

• Northern Ireland interim retired twice (1305-c01, X-m01, X-s02). María ruled on 17 October: 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 share formula, carried by UK local surface before the Budget build: high-value homes OBR leg, per-vehicle fuel flag, Northern Ireland energy share and housekeeping (#1278) #1306. This PR drops its subtraction commit (6928d08) whole, so the five-file conflict, the duplicate changelog fragment and the receipt's 694 kWh factors go. The membership pointer is moot here for the same reason.
• Local compile-parity receipt (1305-10, X-s01, X-m02). Re-pin the UK Chronicle feed to abf0921 (#1300) #1302 regenerated it before merging. This head regenerates it again: 25,327 compiled cells, and the 18 BRMA rows are ledger-only. The body says that whichever of Place UK private rent and BRMA at the assigned area; bind Scotland's rents at BRMA grain (#1262, #1090) #1305, UK local surface before the Budget build: high-value homes OBR leg, per-vehicle fuel flag, Northern Ireland energy share and housekeeping (#1278) #1306, UK local release: publish a K=15, 150,000-household dense dataset in at most two builds (#1301) #1307 and Port the uk-data wave to policyengine-uk 2.134.4: stamp duty, Move to UC, national home values (sixth #1095 PR) #1308 lands last regenerates it, and the battery digests, on the combined tree.
• Head builds. Still owed; the PR stays draft. On current main a local dense build stops at the problem step on high-value-homes cells that UK local release: publish a K=15, 150,000-household dense dataset in at most two builds (#1301) #1307 defers, so the head build follows UK local release: publish a K=15, 150,000-household dense dataset in at most two builds (#1301) #1307.

Should fix

• 1305-01. María kept the census-mix scale. The five texts now say the FRS level holds on the census renter mix, and that at design weights a region moves by the mix gap until calibration's tenure rows restore it. The receipt reports region_level_at_design_weights as a named check.
• 1305-03. Both private-rent routes, and the body, declare the level miss.
• 1305-02. UK_LOCAL_WEIGHT_GRAIN moves into local_doctrine, and uk_local_target_loss_weights maps through it. That covers the solve, the holdout and uk_rule_loss_weights; the alignment check keeps the raw labels. A test holds each grain's share when the 18 rows are added.
• 1305-04. The gate is in the classification parametrize. A new test runs the unstubbed binding on an off-domain BRMA, which fails and stops at both rungs.
• 1305-07, -08, -23, -24, -25, -26.
  • Carriers come from one column, FRS-channel only for spine values.
  • The top 1% needs ten rows over ten households, and the unweighted coherence share counts households on both sides.
  • District and BRMA groups need ten cells.
  • The summary, the audit, the pool table and the engine report refuse paths inside the repository, and .gitignore covers *.npz and audit-*/.
  • Synthetic tests cover every summary writer.
• 1305-06. "Not yet run" lists the UC housing element, obr.housing_benefit, and national and local-authority AHC. The body repeats the stay-as-reported columns.

Minor

• 1305-05 and 1305-19.
  • The toy-graph comment is corrected, and the test asserts the rent changes.
  • London's toy categories have distinct BRMA mixes, and a Monte-Carlo test holds the weighted redrawn mean at the FRS level for renters placed on the census mix.
  • The London test asserts its six clones and both BRMAs unconditionally.
• 1305-09 and 1305-31. The parity tool reads the UC bands off the regenerated rows and says what the paid-claims basis leaves. It routes the support floor's signed deferrals first, so E06000053 and E09000001 carry their own rationale.
• 1305-11, -16, -21. Added to "For review".
• 1305-12. --target-geographies accepts brma, and the selection receipt lists what a selector drops.
• 1305-13. The quantile tool, its resource and the build-graph doc name the 16 cross-border LSOAs. The resource was rebuilt for the text after its --check reproduced it.
• 1305-14 and 1305-29. Without area BRMAs the selection leaves the BRMA rows out and counts them; naming brma is refused. The layer and quantile override and sha flags are refused in none mode.
• 1305-15. Below full scale, a region whose sample holds no FRS renter ranks and scales on Great Britain's, and the receipt counts the cells and regions. Full scale stays strict.
• 1305-17 and 1305-30.
  • A pinned digest of the supports' area-to-region pairs (35,672 LSOAs, 46,363 output areas) lets CI check the layer; the mounted test re-derives it.
  • A ladder-free test checks the crosswalk's layer pin.
• 1305-18. The layer, overlap and quantile --check runs are recorded in the body: 17 October at 3a45efd1b.
• 1305-20. New tests cover:
  • both none release refusals;
  • the release-candidate override refusals and the none-mode flag refusals;
  • the size identity moving with a rent pin;
  • brma in the target weights;
  • the BRMA codes and metric.
• 1305-22. The note states 9 filled, 59 corrected and 2 Balloch overrides, and where those counts live.
• 1305-27. María's ruling: 10 distinct source households. A thinner BRMA row keeps its place and its failure in the calibration diagnostics, the targets CSV and the gate report, without estimates or errors, and stays out of every rollup. The gates still evaluate on the full figures.
• 1305-28, -32, -33, -34.
  • The composer is named correctly.
  • The parity census names the July 2026 edition the selector resolves.
  • The DWP comparison is marked indicative.
  • The receipt says rounded record counts.

Cross-PR.

• X-s06. The DESNZ notes are 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 to update now; Re-pin the UK Chronicle feed to abf0921 (#1300) #1302 left them neutral.
• 1305-c02, X-m04, X-m05. These need an agreed merge order before 20 October. The body asks the last PR to land to regenerate the membership, the receipt and the battery digests on the combined tree.

🤖 Generated with Claude Code

[juaristi22]

Re-review of this PR at c5ec9f6

Run with Claude Code on 17 October 2026, after the rebase onto main (post #1302) and the nine new commits. Three independent reviewers re-checked every first-round finding, attacked the new code with synthetic probes, checked the new public text against the UK Data Service rules, and simulated the merges with #1306, #1307 and #1308. No builds and no feed-gated tests were run.

Verdict

The rebase is clean:

• The first five commits are patch-identical replays.
• The Northern Ireland energy commit and its changelog fragment are gone, with no remnant, per María's 17 October ruling that UK local surface before the Budget build: high-value homes OBR leg, per-vehicle fuel flag, Northern Ireland energy share and housekeeping (#1278) #1306 carries the share formula.
• The regenerated local compile-parity receipt passes the gate replay with 0 failures.
• Almost every first-round item is fixed in code.

But 48529f1, which withholds the model figures of BRMA rows resting on fewer than 10 distinct source households, has four high-severity problems. Two independent reviewers found them. Fix them before the head build:

• one withheld row breaks the dense build;
• three routes still give a withheld row's figures back.

The response's 1305-27 line ("without estimates or errors … and stays out of every rollup") is not yet true for the reasons below.

Blocking (before the head build)

• One withheld BRMA row breaks the dense build after export (high; fail-closed).
  • The manifest projection reads the withheld targets CSV back (graph_terminal.py:1522), and the blank cells become NaN.
  • max_abs_relative_error, median_abs_relative_error (:1614, :1791-1796) and fit.local_by_family (:1865, whose np.argmax picks the NaN row as the worst) all go NaN.
  • json_text writes with allow_nan=False (rowwise_cli.py:153-164), so full_build_cli.py:2446 raises "Out of range float values are not JSON compliant: nan". That happens after the solve and the export, leaving no manifest and no build.json.
  • The scorer (tools/score_uk_local_candidate.py:78-103) also refuses any null final estimate, and the dense release requires score_vs_incumbent.json.
  • Plausible at release size: on public census counts, proportional selection at 100,000 households would put only about 10–18 private-renter rows in the smallest Scottish BRMAs (an expectation from the census, not a count of frame records), and distinct source households can only be fewer. It is near-certain on sampled rungs.
  • Fix:
    • Compute those manifest figures over rows with finite errors, and count the withheld rows.
    • Have the scorer skip withheld rows, counting them.
    • Add a projection test and a scorer test with a withheld row.
• The real local fit gate's output keeps a withheld row's error (high; disclosure).
  • uk_withhold_gate_report_rows (uk_runtime/diagnostics.py:1017-1049) matches lines starting with ":" and details.relative_errors/worst_offenders. That is the generic target_fit_gate's shape.
  • uk_local_target_fit writes something else (terminal_gates.py:1019-1022, :1101):
    • lines like microcosm_uk_2024_25: private_rent/<BRMA>/rent/private_rent_brma relative error +x% exceeds 25%;
    • details.failing_targets;
    • rows keyed family/area/metric (battery_bindings.py:1391).
  • A synthetic probe on the production binding kept the failure line and the error after withholding.
  • uk_local_per_family_fit still counts withheld rows in its family shares and sizes. Set against weakest_families, which drops them, that gives each row's within-10% and within-25% status.
  • Where it spreads: the full gate report, the signed local report, the dense manifest's blocking_failures/diagnostic_failures, and staging.
  • The commit's test builds the generic shape by hand.
  • Fix:
    • Map each withheld row to its gate key, and remove it from every details mapping, list and failure line.
    • Publish family shares and sizes without withheld rows, with a withheld count.
    • Test against the real bindings, through UKFullGateKernel.
• The calibration diagnostics' totals give a withheld row back by difference (high; disclosure).
  • Withholding blanks per-row fields only (diagnostics.py:74-81, :1143-1203).
  • initial_loss, final_loss, loss_trajectory, fraction_within_10pct and past_cap_census.pushed_out_rows are still computed over every row (microcosm-calibrate diagnostics.py:270-350, :436-438).
  • final_loss equals the sum of the published contributions to within 1e-12, and the withheld row keeps its target, target_loss_weight_share and target_loss_scale. So |estimate − target| = (final_loss − other contributions) / share × scale.
  • On the commit's own fixture this recovered the withheld estimate exactly, and pushed_out_rows lists the withheld row with its relative errors.
  • These files ship in the dense release.
  • Fix (María's call on the form):
    • while any row is withheld, publish the losses, trajectory and fractions over the published rows only, labelled as such, or null them;
    • take withheld rows out of pushed_out_rows;
    • add a test that tries to recover a withheld row.
• The quotable rent receipt summarises the 18 BRMA rows with no carrier rule (high; disclosure).
  • tools/uk_private_rent_receipt.py:338-376 (rent_row_fit), written to summary["rent_rows"] at :798.
  • It gives the BRMA group's share_within_10pct, rows_over_25pct and median_abs_error.
  • With one of 18 synthetic rows withheld, the median's arithmetic returned its |error| (0.089 against a true 0.0891), and share × 18 minus the published count returned its within-10% status.
  • The PR plans to quote exactly these "316 + 18 rent rows".
  • Fix: count carriers per rent row as uk_withheld_brma_rows does, and leave withheld rows out of every group, with a count.

Should fix

• A cell with zero FRS carriers passes every floor (medium; receipt tool). The floors test 0 < n < 10 (:213, :230, :250), and frs_only blanks the SPI keys. So a spine-values cell made only of SPI rows prints a mean with source_households: 0. Treat any non-empty side with fewer than 10 counted carriers, zero included, as "<10".
• SPI copies count as their FRS parent for housing figures (medium; from code reading; María's call).
  • SPI copies carry their parent's lineage, but their tenure, bedrooms, rent, subrent and reported Housing Benefit are drawn from FRS donors (spi_housing_shell.py).
  • So the receipt's release and design blocks, and the BRMA carrier rule, overcount carriers in the way c4afed1 fixed for spine values only.
  • Fix: count FRS-channel rows only for housing figures.
• The Monte-Carlo test's tolerance (rel=0.03) is loose.
  • The unmutated ratios are 0.995–1.000.
  • A mutant using one category's BRMA mix for every category gives 0.970–0.974, and all 25 tests still pass.
  • About 1% would catch it.
• 1305-15 is partly fixed. A sampled region with one or two FRS renters still computes its own scale, and refuses when the scale leaves SCALE_BAND. cells_ranked_against_their_region double-counts fallback cells.

Low and nits

• The repository guard knows only this checkout (uk_private_rent_receipt.py:97, :164-174, the pool tool, the engine script). It accepts a path inside another git work tree, and on a case-insensitive volume this checkout's own path written in different case. Refuse when any ancestor of the resolved path holds a .git entry.
• Two thin cells get no complementary withholding. suppress_cells withholds a complement only when exactly one cell is thin (:254).
• Zero-weight rows count as carriers in the receipt; the diagnostics rule uses positive weights only. Defensive today.
• Tests that pass with the fix reverted.
  • Nothing runs the kernel's withholding.
  • Nothing calls measure_run, main or the pool tool's guard, so reverting frs_carriers=True or the guard call passes everything.
  • Nothing pins strict_cells.
• 1305-11: changelog.d/1262-uk-brma-from-area.added.md still ends "while the certification part digests do not".
• 1305-20 is partly fixed. There is no test of the scorer's statistical_scope/brma grain, or of build_uk_full_local_problem with BRMA rows.
• Nits:
  • the floors' #: comment in local_doctrine.py:98-115 now documents UK_LOCAL_WEIGHT_GRAIN;
  • stale six-gate counts in three places (graph_national.py:1347, contract.py:4608, calibration_run.py:269);
  • local_validation_levels.json:27/47 still name the September PIPR record sets;
  • the cross-grain rent reasons lack 3b65352's design-weight qualifier;
  • the pool tool and the engine report check --out only after the expensive work;
  • max_target_scaled_change in the size receipt is a maximum over all rows, withheld ones included;
  • no changelog fragment mentions the null BRMA model fields.

First-round dispositions

• Fixed: 1305-01 to -04, -06 to -09, -12 to -14, -16 to -19, -21 to -26 and -28 to -34. The parity UC bands are recomputed from the rows: 0.937–0.973 over 980, and 0.815–1.237 over 2,527.
• Partly fixed: 1305-05 (tolerance), -11 (changelog), -15 (thin sampled regions) and -20 (two test gaps).
• The fix introduces new problems: 1305-27 (the four Blocking items).
• No longer applies: 1305-10, c01, X-m01, X-m02, X-s01, X-s02 and X-s06.
• Future sequencing: c02, X-m04 and X-m05.

Cross-PR

• This PR × UK local surface before the Budget build: high-value homes OBR leg, per-vehicle fuel flag, Northern Ireland energy share and housekeeping (#1278) #1306.
  • The only conflict is the geographies set in test_uk_local_target_census.py; resolve it to {"constituency","la","brma","msoa","country"}.
  • The receipt and membership auto-merge to the correct union (29,590 rows; 25,327 compiled), and the merged gate replay passes.
  • No regeneration is needed. Re-run the census and parity --checks on the merged tree.
• This PR × UK local release: publish a K=15, 150,000-household dense dataset in at most two builds (#1301) #1307 (e119867).
  • 10 conflicts. Both PRs add _uc_drift_rationales: keep UK local release: publish a K=15, 150,000-household dense dataset in at most two builds (#1301) #1307's UC code, plus this PR's support-floor routing and its BRMA/PIPR rationales.
  • The correct union: receipt 27,267 rows (23,004 compiled), references 23,004, membership active 23,004 and candidates 27,133.
  • Count pins to change by hand: the scorer 23,004; reference_targets 27,133; deferrals and absences 4,129; population targets 418.
  • Recompute the full and dense digests.
  • If this PR lands first, UK local release: publish a K=15, 150,000-household dense dataset in at most two builds (#1301) #1307 must redo its regeneration on the combined tree.
• This PR × Port the uk-data wave to policyengine-uk 2.134.4: stamp duty, Move to UC, national home values (sixth #1095 PR) #1308 (8739557). It conflicts on the full-battery digest triple in contract.py and test_contract.py; recompute the digests on the combined tree.
• The head build follows UK local release: publish a K=15, 150,000-household dense dataset in at most two builds (#1301) #1307.

Disclosure check

The new commit messages, added lines and body carry no survey-derived figures. The three leak routes in the new code are under Blocking.

Tests and tools run

• Engine-free, at c5ec9f6: 1,421 tests, 0 failures, 10 skips (feed, supports or ladder), plus 283 more, 0 failures, 3 skips.
• Engine (policyengine-uk 2.132.6): 10 tests, 0 failures.
• tools/ci_test_plan.py verify ok, the census and parity --checks current, and ruff clean on the 68 changed .py files.
• Gate replays:
  • head membership vs head receipt: 0 failures;
  • head membership vs main's receipt: fails on the 18 BRMA rows, so the regeneration was needed;
  • merged with UK local surface before the Budget build: high-value homes OBR leg, per-vehicle fuel flag, Northern Ireland energy share and housekeeping (#1278) #1306: 0 failures.
• The BRMA withholding problems were shown with synthetic frames only.

🤖 Generated with Claude Code

[juaristi22]

Addressed in b80b596, rebased onto main after #1307 and #1306 merged. Item by item:

Rebase

• On UK local release: publish a K=15, 150,000-household dense dataset in at most two builds (#1301) #1307: the 10 conflicts resolve as the re-review's simulation said. UK local release: publish a K=15, 150,000-household dense dataset in at most two builds (#1301) #1307's UC drift code is kept, beside this PR's support-floor routing and BRMA rationales. The count pins are 23,004 active references, 27,133 candidates, 4,129 signed deferrals and absences, and 418 contract targets. The full and dense digests are recomputed. The certification part digests don't move.
• 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: the census test's geographies resolve to {"constituency", "la", "brma", "msoa", "country"}, as your simulation said. The references, membership and census regenerate on the combined tree to exactly the auto-merged bytes (23,004 active), and the census and parity checks are current. The gate digests are unchanged.
• The local compile-parity receipt is regenerated on the combined tree: 27,267 rows, 23,004 of them compiled. Against main it adds only the 18 BRMA ledger-only rows. Reasons change on the 40 cells of the support floor's two authorities, which now carry the floor's own signed rationale. Of the four PRs, Place UK private rent and BRMA at the assigned area; bind Scotland's rents at BRMA grain (#1262, #1090) #1305 and Port the uk-data wave to policyengine-uk 2.134.4: stamp duty, Move to UC, national home values (sixth #1095 PR) #1308 are still open; whichever lands second regenerates the receipt and the battery digests on the combined tree.
• The sign-off drafter's local levels include brma, now that the full gates' do.

Blocking

• One withheld row breaks the dense build. The manifest projection reads the targets table's new model_figures_withheld flag. Every error summary and family fit covers the published rows, and the solve block counts the withheld ones. While any row is withheld, the size receipt's totals over every row (dense_loss, compact_loss, max_target_scaled_change) read null in the manifest, with a note; the stored receipt keeps them. The scorer skips a withheld row on both sides and counts it (rows_withheld). It refuses a withheld row that carries an estimate. The release contract lets a row carrying the note leave its estimates null, if it carries no other model figure and the payload's withheld block counts it. New tests: a dense CLI run whose terminal table has a withheld row, plus scorer and contract tests.

• The real local fit gate keeps the error. build_full_gate_context counts each BRMA row's carriers on the installed weights and flags the thin ones; the kernel reads that one set. uk_target_fit_gate still decides on every row, but writes a withheld one by gate key only:

  • its failure line quotes no error and comes after every other line, in name order, so its rank gives nothing away;
  • it moves from failing_targets to withheld_failing_targets;
  • an exclusion it stands on carries no error.

  uk_local_per_family_fit rests its shares, sizes and decision on the published rows and counts the withheld rows by family. The kernel now refuses any report that names a withheld row beside a figure, so a future gate that lists per-row errors fails the build instead of leaking. The first round's rewrite of the generic gate shape goes. The tests run the production bindings, and the terminal kernel on a fixture whose BRMA row is withheld. A binding that drops the flag makes the kernel fail closed.

• The diagnostics' totals give the row back. María ruled on 18 October that while any row is withheld, the totals cover the published rows only, labelled.

  • final_loss is the published rows' summed contributions, on the solve's own weights, scales and cap. initial_loss is the same at the input weights.
  • fraction_within_10pct and past_cap_census count the published rows, so a withheld pushed-out row leaves the listing.
  • loss_trajectory is empty, since it cannot be restated.
  • uk_diagnostics.model_figures_withheld counts the withheld rows and names the basis. It is an optional schema field.

  A test tries to recover a withheld row from the totals and finds nothing left. Over every row, the restatement reproduces the solve's own totals.

• The rent receipt has no carrier rule. Each rent row now rests on the distinct source households of positive weight with an entry in it, as the build's rule counts them. A row on fewer than ten stays out of every group, which counts it. Source keys come from the pool table for the design and dense weightings, and from the H5 for the release one; --control-pool gives the control's. A weighting without keys isn't summarised.

Found while fixing these. Two more readers took a withheld row's figures:

• The sign-off drafter (UK local-area release: publish a K=15, 150,000-household dense dataset in at most two builds #1301). It refused a blank error, so it could not sign such a build. It now signs a withheld cell beyond the bound by key, as the gate names it, and its reasons carry no figure. For its prediction of build 2 it stands each withheld cell in at twice the bound or at zero, as the gate found it; the stand-ins are never written.
• Orrery's result summary. A staged build publishes this. It now counts each BRMA row's carriers on that result's own weights, blanks a thin row's figures, and restates the final loss over the published rows. The size summary can't tell which rows are withheld, so for a problem that binds BRMA rows it publishes no total over every row.

Should fix

• Zero-carrier cells. A non-empty side with fewer than ten counted carriers reads "<10", including zero.
• SPI copies as carriers. María ruled on 18 October to keep lineage as the carrier key. No change.
• Monte-Carlo tolerance. It is now 1%. The mutant that uses one category's mix for every category fails on all three seeds, and the unmutated test passes.
• 1305-15. Below full scale, a region with fewer than ten FRS renters ranks against, and is scaled on, Great Britain's renters, and the receipt counts it. Full scale stays strict. Each cell is now counted once, so cells_ranked_against_their_region no longer includes Great Britain's fallbacks.

Low and nits

• The guard. outside_work_trees (receipt_counts) refuses any path with a .git entry in an ancestor. The receipt, the pool tool and the engine report run it before any work.
• Thin cells. Complements are withheld until the withheld cells together rest on ten households.
• Zero weights. A zero-weight row carries nothing, as in the build's rule.
• Tests that passed with the fix reverted. main now runs end to end on a synthetic run, so reverting frs_carriers or the guard fails. Every writer refuses a work tree before reading anything, the kernel's withholding is tested, and strict_cells is pinned at both ends.
• 1305-11. The B1 changelog says that a release certified before this change fails validate_release after it.
• 1305-20. Tests now cover the scorer's statistical_scope grain and build_uk_full_local_problem with a BRMA row.
• Nits.
  • The floors' doc comment is back on its constant.
  • Three live descriptions say seven gates.
  • The two validation rows name the July record set the rent rows resolve.
  • The cross-grain reasons and the selector notes carry the design-weight qualifier.
  • max_target_scaled_change is covered under Blocking.
  • The B3 changelog says what a build withholds.

Not yet run. The head spine and dense builds, and their receipts. Main now has #1307, so they can start.

🤖 Generated with Claude Code