Bind the UC element targets on paid claims and an OBR UC total on the chronicle 474a0ae feed (#882)

[juaristi22]

Step 2 of the revised UC gap plan (#882, sequencing ruled 2026-09-11): bind the DWP Universal Credit element counts on paid claims now that chronicle#260 carries the Payment Indicator crosses and chronicle#263 carries the dimension labels main's schema-8 hierarchy needs. Rebuilt on main after #904 and #906 merged, so it rides #904's single 474a0ae pin for both grains and leaves #906's region tier untouched; national scope only. The local grain stays where main left it (chronicle#266 / #920) and can be regenerated on this branch if #266 lands first.

What lands

• Seven dwp_universal_credit rows on the calendar-2025 paid-claim basis of the ten broad rows (monthly_window_average, payment_indicator: Yes, GB): LCWRA (publisher group value LCWRA, both rates), carer, housing any tenure, housing social rented and private rented (via benunit_tenure_type), childcare (the ODS all-claims count, caveat declared), deductions (March–December 2025, the published history). Model side: universal_credit > 0 and the element > 0.
• obr.universal_credit: the sum of the two OBR welfare-cap cells (expected_member_count: 2, FY2025/26, UK) against the model's total award. The in/outside-cap split stays excluded; the model cannot attribute spend to the welfare-cap concept.
• Register (three entries, tracking Align UK UC claim, child and period definitions #882, signed by María in review on 2026-09-15, expiring 2026-10-15): carer and childcare as measured-but-not-fit rows (model concept gaps, not weight questions), and the any-tenure housing row as a structural-bias diagnostic: DWP's "Yes" is social + private + "Other or unknown" tenure, the model has no other/unknown tenure, so its any-tenure count is identically the two tenure rows and the published total exceeds what the model can reach by 112,518 (2.8 %). The two tenure rows carry the housing element in the fit.
• LCWRA stays in the objective by the rule its binding note states: a concept mismatch goes to the register only when it puts the row outside the 25 % release bound at the baseline (+15 % / +20 % here). Whether the is_disabled_for_benefits passthrough warrants an engine change is a Step 3 question.
• Regenerated surfaces on the Bind the Chronicle transport and energy facts on the UK national surface (#890 PR-T) #904 pin (chronicle 474a0ae, 141,400 facts): national references and membership 595 → 603 active, 7 deferred, 7 signed out, no existing reference changes; the 2025 compile-parity receipt gains the eight rows and the 2023 production receipt seven (the OBR total does not compile at 2023, the March 2026 EFO starts at FY2024), nothing else moves; the vendored fact resources check clean. No feed re-pin in this PR.
• The three --baseline-pi-floor commits from the 10 September rebuild (the Q50f baseline was built with them; changelog.d/355-baseline-pi-floor.added.md covers the knob), docs/uk-uc-baseline-2026-09-10.md (the Q50f/D2 UC readout recorded on Align UK UC claim, child and period definitions #882, both feed pins, the compile-time targets, the element sizing) and tools/diagnose_uk_uc_elements.py, which sizes element support conditioned on paid awards.

Measured at the baseline

Calendar-2025 paid targets against Q50f / D2 at their calibrated weights:

row	target	Q50f	D2
LCWRA	2,272,425	+15.3 %	+20.2 %
carer (register)	1,078,514	−44.8 %	−44.1 %
housing, any tenure (register)	4,037,650	−0.8 %	+0.6 %
housing, social rented	2,354,588	−0.3 %	+5.1 %
housing, private rented	1,570,545	+5.6 %	+1.2 %
childcare (register)	178,667	+155.0 %	+152.6 %
deductions	3,130,000	−4.5 %	−4.4 %
OBR total	£79.29bn	+1.4 %	+1.9 %

The carer gap is the engine's carer condition (Carer's Allowance receipt) against DWP's 35-hours-a-week concept, with the engine's care_hours input never filled by the spine; the childcare gap is an element paid to every working family with reported childcare costs and no take-up lever. Both repairs are Step 3 of the plan.

Verification

• Full UK lane with the pinned feed available (CHRONICLE_UK_FACTS), at fb6b1674 (pre-UK region-grain activation: fan out the 18 two-level targets over the region tier and make region a cross-grain control (#905) #906): 2,098 passed, 22 skipped, 1 failed — test_committed_surfaces_regenerate_from_pinned_feed[local], the UK local target surface cannot compile since #855: sub-national geographies have no hierarchy label (needs chronicle geography.name) #920 local-generator refusal (LedgerHierarchyMetadataError, no geography.name) that fails identically on main and skips in CI. The national hermetic regeneration and compile-parity tests pass. Round 1 (c088792c) and round 2 changed no generated surface; after the rebuild on UK region-grain activation: fan out the 18 two-level targets over the region tier and make region a cross-grain control (#905) #906 the twelve touched and neighbouring suites re-run green with the pinned feed (result in the rebase comment).
• ruff check . clean; CI green on all 23 checks at fb6b1674, c088792c and 2afa2869; re-running on the rebuilt head.
• Contract accounting: targets 244 → 252, registry scope 210 → 218, unmapped declarations 27 → 35, register 50 → 53, monthly UC references 111 → 118.

Review

Approved by Vahid at c088792c (2026-09-15); the pre-merge should-fix (the register tranche assertion) landed in 2afa2869. Rebuilt on main after #906 merged; the diff is the same content on the new base. Round 1 addressed in c088792c: the any-tenure housing row moved to the register as a structural-bias diagnostic, the three register entries adjudicated to #882 and the committed baseline doc and signed by María with a one-month window, the LCWRA fit-versus-register rule stated, and the body corrections above (2023 receipt count, deductions rounding, both feed pins, the #355 fragment).

Follow-ups

• Chronicle-side: chronicle#261 (labels, shipped as Build the Belgian calibration surface by recalibrating the US release recipe #263) done; chronicle#266 (sub-national geography names) blocks the local grain only.
• Step 3 issues after this merges: carer care hours, childcare element take-up, whether the LCWRA passthrough warrants an engine issue.
• UK region-grain activation: fan out the 18 two-level targets over the region tier and make region a cross-grain control (#905) #906 merged first; this branch was rebuilt on it (regenerable surfaces and count assertions only, plus the register where both append entries).

🤖 Generated with Claude Code

[vahid-ahmadi]

Review at fb6b1674 (Claude Code, high effort; worktree checkout of the four commits on main 18271b28; without the engine or the pinned feed: 480 passed, 5 skipped, 0 failed across the thirteen touched and neighbouring suites; ruff check . and ruff format --check clean; both census tools, the coverage manifest, the spec-engine coverage check and ci_test_groups --verify current; CI green on all 23 lanes; merge-tree clean against main and #927).

The bindings are clean where they are unambiguous. Every DWP selector matches exactly one series — LCWRA the group package rather than LCWRA - Higher/Lower, Yes rather than the tenures, source_measure_id: benefit_units separating childcare's count from the mean-amount fact — and each resolved value is the twelve-month mean of the chronicle #260 rows to the float (deductions over March–December). The window is the same calendar-2025 basis the ten broad rows already use. All five element variables exist on the engine at benefit-unit entity and uc_deductions is driven by a spine input from #850, not zero by construction. The tenure mapping (council plus housing association → social, private renting → private) is consistent on both sides for the two tenure rows. The OBR row's two-member sum has precedent in the NITHC, Wales and Translink rows, the FY2025-26 → 2025 rule is the shared one, and the 424 untouched references are byte-identical to main with membership moving only by the eight new pins. The childcare all-claims caveat is declared in the binding note and pinned by test. The body table is in docs/uk-uc-baseline-2026-09-10.md with Q50f and D2 pinned, and carer and childcare are labelled as register rows there and in the gate, which passes.

Should-fix

1. The three housing rows are jointly inconsistent by DWP's "Other or unknown" tenure cell. From the same chronicle rows, paid, January–December 2025: social 2,354,588 + private 1,570,545 + other/unknown 112,518 = 4,037,651, which is the any-tenure target to the unit. On the model side uc_housing_costs_element pays only for council, housing-association and private renting and TenureType has no other/unknown value, so the model's any-tenure count is identically social plus private — your own baseline shows it (Q50f 4,006,395 = 2,348,245 + 1,658,150 exactly). Three active rows, two model degrees of freedom, contradictory by 112.5k or 2.8%: the solver has to over-fill the tenure rows to reach the total. The binding note says never to sum the tenure rows into the total, but does not say the model cannot reach the total by construction. Either drop the any-tenure row as redundant, or keep it and declare the structural bias, as a register entry or a note the gate can see.

2. The carer and childcare register entries are not adjudicated to anything a reader can open. adjudication: "uk-uc-gap-plan-2026-09-11 (María's sequencing ruling …)" resolves to nothing in the repo, and María's 2026-09-11 comment on #882 records measurements and sequencing but no ruling that these two stay out of the objective. Every other entry cites microcosm#NNN or a committed receipt. approved_by: juaristi22 and approved_on: 2026-09-11 are pre-filled while the body says the signature is María's to confirm in review, and the author is the approver. The window is 88 days to 2026-12-08 where every other entry runs exact months; it looks borrowed from the #355 tranche's expiry. Point the adjudication at an issue comment or a committed file, and let María's sign-off set the dates.

3. LCWRA is bound active while its own note declares a concept mismatch. The note says the engine's LCWRA flag is the is_disabled_for_benefits input, not a work-capability assessment, and the row fits at +15% to +20%. Carer went on the register for exactly "model concept gap, not a weight question". If magnitude is what decides between fit and register, say so where María can rule on it; the body's own follow-up asks whether the passthrough warrants an engine issue.

Body accuracy

• "Both national compile-parity receipts gain the new rows": the 2023 production receipt gains seven, not eight — obr.universal_credit does not compile at 2023 because the March 2026 EFO reaches back only to FY2024. The 2025 receipt gains all eight.
• The deductions note says counts are "rounded to 100,000"; the chronicle package header says only "rounded counts" with the caseload derived at 10,000.
• docs/uk-uc-baseline-2026-09-10.md:11 pins feed ec7169b while the paid targets at lines 78–87 come from 474a0ae; pin both.
• Three #355 commits ride this #882 PR (--baseline-pi-floor, size evaluation, receipts across dataset_size.py, size_evaluation.py, local_rowwise.py, build_uk_rowwise_candidate.py), and the new knob has no changelog fragment because the #355 fragment predates the branch.

Could not verify

The deductions value 3,130,000 (ODS parse), the hermetic regeneration and compile-parity tests (feed-gated; CI is the evidence), and the "provisional March 2026 month" in the OBR note.

Ordering

Merge-tree against #906 conflicts in eight files, all regenerable surfaces and count assertions (432 / 49 / 218 here) with no content dependency, so whichever lands second regenerates and re-bumps. Clean against #927 and current main.

---

1 is the one to settle before this is used as a fit basis, since the surface asks the solver for something the model cannot produce; 2 is governance and cheap; 3 is María's call.

[juaristi22]

Thanks — round 1 addressed in c088792c (no generated surface moves: 432 references and both receipts byte-identical; register 49 → 50).

Should-fix 1 (housing rows). Agreed on the arithmetic: DWP's "Yes" is social + private + "Other or unknown" (112,518 in calendar 2025, 2.8 %), and the model's any-tenure count is identically the two tenure rows because uc_housing_costs_element pays only council, housing-association and private renters. The any-tenure row now sits on the measure-exclusion register as a structural-bias diagnostic (dwp/uc/elements/housing, reason states the mechanism and the 112,518 gap); the two tenure rows carry the housing element in the fit. The binding note says the same, so the contract and the gate agree.

Should-fix 2 (register adjudication and dates). The three entries now cite microcosm#882 and docs/uk-uc-baseline-2026-09-10.md (the committed sizing), and their signature is María's in review of this PR: approved_on: 2026-09-15, expires_on: 2026-10-15, a one-month window she set rather than the #355 tranche's date.

Should-fix 3 (LCWRA). The rule is now stated in the binding note and in the baseline doc: a concept mismatch goes to the register only when it puts the row outside the 25 % release bound at the baseline; LCWRA at +15 % / +20 % stays in the objective, carer and childcare do not. Whether the is_disabled_for_benefits passthrough warrants an engine change is deferred to Step 3 under #882, where María rules on it.

Body accuracy. Corrected: the 2023 production receipt gains seven rows (the OBR total does not compile at 2023 because the March 2026 EFO starts at FY2024), the deductions note now says DWP publishes rounded counts, the baseline doc pins both feeds (ec7169b for the build, 474a0ae for the element targets), and changelog.d/355-baseline-pi-floor.added.md covers the --baseline-pi-floor knob riding this branch. The "provisional March 2026 month" wording follows docs/uk-uc-paid-target-contract.md, which records FY2025/26 as including provisional March 2026.

Could not verify. The deductions target is the mean of the ten published monthly counts in the 474a0ae feed (dwp.uc_households_with_deduction, units_with_deduction, 2025-03 to 2025-12), 3,130,000.0; the feed-gated regeneration and compile-parity tests run locally with CHRONICLE_UK_FACTS set and pass on the national surface (the local one fails with the #920 refusal as on main).

Ordering. Noted on #906: regenerable surfaces and count assertions only, whichever lands second regenerates.

[vahid-ahmadi]

Closing pass at c088792c (Claude Code, high effort; worktree checkout, the response commit read against the dispositions — five files, 23 lines added, 12 removed; without the engine or the feed: 373 passed, 5 skipped, 0 failed across the nine named suites; ruff check and ruff format --check clean; both census tools, the coverage manifest, the spec-engine coverage check and ci_test_groups --verify current; CI 21 lanes green with the 3.13 wheels lane still running when I posted).

All three should-fixes closed as described, and the body corrections are in.

1. Verified. dwp/uc/elements/housing is on the measure-exclusion register as a structural-bias diagnostic with the mechanism and the 112,518-of-4,037,651 gap in the reason, and the row stays bound and active in the references and membership (432 active, twelve matched facts) so it is measured on every evaluation. The binding note states the model-side identity in the same words, so the contract and the gate agree.

2. Verified. The three entries now cite microcosm#882 and the committed baseline doc, with approved_on: 2026-09-15 and expires_on: 2026-10-15; the unresolvable plan reference is gone. One thing I will say plainly rather than re-raise: the approver login is the author's, and an author cannot post a review on her own PR, so the signature is the commit rather than an independent review record. The adjudication text says María signed in review; if that is the convention, fine, but the register cannot show it.

3. Verified and consistent. The rule — a concept mismatch goes to the register only when it puts the row outside the 25% bound at baseline — is on the LCWRA note and in the baseline doc, and it holds across all six element rows against the table: LCWRA at +15 / +20 in, carer and childcare out, the two tenure rows and deductions in. The any-tenure row is registered under a separate structural-bias class, so no contradiction.

Body. The 2023 receipt is now "seven rows" with the reason, the deductions note says "rounded counts", the baseline doc pins both feeds, and the --baseline-pi-floor fragment exists. No generated surface moved: references, membership, both parity receipts and the uk-data parity register are byte-identical to fb6b1674.

One residual, should-fix before merge. The 2026-09-15 tranche is unasserted. test_uk_measure_simulation.py pins the total at 50 and asserts the 08-26 and 09-03 tranches by name, date, window and tracking, but the three dwp/uc/elements/* entries have no tranche block and no census marker, so a drift in any of their fields passes the test the docstring says exists to catch exactly that. One block in the shape of the A16 one. Nit: the register says 4,037,651 where the baseline table publishes 4,037,650 — rounding of a monthly mean, but the register is the signed record.

Ordering. Clean against main and #927. Against #906 the eight conflicts are the same regenerable surfaces as before plus the register, where both PRs append entries; whichever lands second regenerates and re-bumps.

---

Approving, with the tranche assertion asked for before merge.

[juaristi22]

Thanks for the closing pass. The residual is in 2afa2869:

• Tranche assertion. test_uk_measure_simulation.py now carries a dwp/uc/elements/ census marker (3) and a 2026-09-15 tranche block in the shape of the A16 one: the three names, tracking: microcosm#882, the approver, an adjudication that cites microcosm#882, the committed baseline doc and microcosm#921, and the 2026-10-15 expiry; it also asserts the four element rows that stay in the objective (LCWRA, both tenure rows, deductions) and the OBR total are absent from the register. Any drift in those fields now fails the test.
• Nit. The register and the binding note quote the published any-tenure mean, 4,037,650, with a note that the three tenure cells reproduce it to the unit.
• On the approver. Understood, and stated plainly: the register cannot carry an independent review record, so the signature is the commit plus this thread, where María set the window in review; her merge is the sign-off the register points at.

No generated surface moved; the register-consuming suites pass (108 tests).

[juaristi22]

Rebuilt on main after #906 merged (head ecfefee1 on 3f8f1d0d), same content on the new base: the eight rows and three register entries were merged into main's contract and register rather than hand-resolved in the generated files, and everything regenerated on the 474a0ae feed. Main's region tier is untouched.

• References and membership: 595 → 603 active, 7 deferred, 7 signed out; no existing reference changes.
• Compile-parity receipts: the 2025 one gains the eight rows, the 2023 one seven; nothing else moves. Vendored fact resources check clean.
• Counts: targets 252, registry scope 218, unmapped declarations 35, register 53, monthly UC references 118; the 2026-09-15 tranche assertion carries over.
• Local run of the twelve touched and neighbouring suites with the pinned feed (CHRONICLE_UK_FACTS): 408 passed, 2 skipped, 1 failed — the UK local target surface cannot compile since #855: sub-national geographies have no hierarchy label (needs chronicle geography.name) #920 local-generator refusal (test_committed_surfaces_regenerate_from_pinned_feed[local]), as on main.
• GitHub reports the head mergeable and clean; CI is running on it.