Set UK road-fuel and energy levels in the LCFS stage on a calendar-2024 basis (#1113)

[juaristi22]

Summary

Closes #1113. Also covers item 3 of #1123. The UK LCFS stage now sets household road fuel and domestic energy at their published levels, at calendar-2024 prices, instead of leaving the calibration weights to reach them. The rows whose gap is not a weight question become measured diagnostics. This follows María's rulings of 2026-10-08 (D1–D3), plus two of 2026-10-09: the ONS road-fuel row is bound, and every LCFS spend column sits on one calendar-2024 basis.

• Calibration rows. obr.vat, ons.household_electricity_expenditure, ons.household_gas_expenditure and obr.fuel_duties_cars move to the measure-exclusion register. They are still measured on every evaluation, and the entries expire 2027-01-08. The contract binds a new household-scope row in their place: ons.household_road_fuel_expenditure (ONS Consumer Trends 07.2.2, calendar 2025, measured as petrol plus diesel spend). The class also carries other motor fuels and lubricants, so the row reads about 0.4% low from scope alone; the binding notes record that offset.
• What no longer holds what. After this PR no bound row and no release-blocking check holds calibrated domestic energy, the VAT base or fuel duty. The stage gates hold the energy and fuel levels at prior weights, the bound ONS row holds household road-fuel spend, and the report-only consumption_drift shows the rest. That follows from D1–D3, and the gate and register notes say so.
• LCFS stage. Three declared steps:
  • redraw_zero_road_fuel gives every flagged fuel-car household a positive fuel spend;
  • level_road_fuel scales petrol and diesel to ONS 07.2.2 for 2024, less the donor's other-fuels share;
  • recompose_from_remainder writes housing around the levelled electricity and gas, and transport around the levelled petrol and diesel and the priced bus fares.
• Calendar-2024 basis.
  • The diary columns are carried on the engine's CPI index from the April-to-March 2023-24 diary year to calendar 2024. That is three quarters of a year's step (1.0186), not the full year's (1.025), which would land a quarter past it.
  • Energy is priced at QEP's calendar-2024 averages and levelled to the DESNZ volume over the four quarters of 2024.
  • Each area's bus yield is re-priced to 2024 by its fares index.
  • policyengine-uk can then project every column from the H5's year.
• Gates. Four new release-blocking stage-health gates check the receipts: road_fuel_incidence, road_fuel_level, recomposed_totals and consumption_basis. The last reads each diary column's declared window. The housing and transport totals also keep their committed support bounds, which the terminal support gate applies to the emitted frame.
• Diagnostics (report-only; a block records why it has no figures when a published row is missing, and does not stop the stage):
  • uk_diagnostics.consumption_drift: the levelled totals and shares at design and final weights;
  • division_capture in the stage receipt: each COICOP division against ONS Consumer Trends;
  • the litres audit: each fuel against its cars share of road use.
• Feed. The UK Chronicle feed is re-pinned to 1ee7dfe (UK consumption-tax facts: Consumer Trends for every COICOP division, and road fuel by vehicle type and fuel (PolicyEngine/microcosm#1113, PolicyEngine/microcosm#1123) chronicle#322 via Expand UK consumption-tax expenditure and road fuel facts chronicle#323, 352,549 rows), which brings Consumer Trends by division and DESNZ road fuel by vehicle type and fuel. The re-pin moves no compiled target value on either surface and leaves the three compile-parity receipts as they were. The new road-fuel row adds one ledger-only row to each national receipt.

Receipts, findings and rulings are in experiments/1113-consumption-tax-levels-receipts.md. The measurement scripts and their aggregate outputs are in docs/evidence/uk-consumption-taxes-1113/.

Why the levels belong to the stage

On #1121's head calibration (policyengine-uk 2.122.2, 1,167 rows):

• obr.vat reads +24.7% at design weights, against a 25% bound. Its level is the engine's microdata_vat_coverage factor (0.38, with no recorded derivation; policyengine-uk#1996), not a weight question.
• The ONS electricity and gas rows sit below the unit cost of the DESNZ volume at QEP prices, so they are a published-source residual.
• obr.fuel_duties_cars reads −34.5% at design weights. It is a fiscal-year duty total that includes business use of cars. The LCFS diary records about a quarter less road fuel than ONS, and a third of fuel-car households show no fuel purchase in the two-week diary.
• Closing these rows through the weights distorts the frame. From design to final weights the gas-connected share fell from 84.5% to 78.9%, petrol rose 21% and diesel 34%.

On the 2026-09-30 national problem (1,090 rows, before this PR's stage levels), a re-solve with the four rows masked together left the other 1,086 rows fitting no worse: 98.5% within 10%, against 98.4%. No arm masks one row alone.

Stage-only measurements

These re-run lcfs_consumption on #1121's head-r3 spine (63,956 households). All seven lcfs stage-health gates pass. A re-run redraws most households, so the comparison is with main's stage re-run on the same spine, not with the spine's stored columns.

• Total consumption moves from £969.6bn to £966.8bn (−0.3%):
  • the ten diary divisions move −0.7%, the calendar bridge;
  • housing moves from £180.1bn to £172.1bn;
  • transport moves from £173.0bn to £182.4bn.
• Road fuel. No flagged fuel-car household is left at zero fuel (33.1% of them by weight before the redraw). Petrol plus diesel is levelled to £36.89bn: ONS 07.2.2 for 2024 (£37.03bn) less the donor's 0.37% other-fuels share.
• Energy. Electricity is £29.5bn and gas £19.2bn at calendar-2024 prices, with 84.6% of households connected to gas.
• Totals. No household's housing total is below its electricity and gas, and none's transport total is below its petrol, diesel and bus fares.
• Division capture. The frame's twelve divisions are 72.4% of ONS's 2024 spending in survey scope; the donor's own diary is 58.5%.
  • Housing reads 1.051. D1's energy level is £8.0bn of its £8.3bn excess.
  • Furnishings read 1.242: the frame draws 61% more per household than the donor. That is not investigated here.
• Litres audit. Against HMRC clearances times DESNZ's cars share of each fuel, petrol reads 1.008 and diesel 0.713.
• The bound row at design weights (desk figure). ONS's own move from 2024 to 2025 is −5.2%; policyengine-uk 2.122.2 carries petrol and diesel −2.7%. With the scope offset the row reads about +2.3% on that version. R1 records the measured value.

Not in this PR

#1207 collects these open items, under the two deferral registers #736 and #796. The three measure exclusions that are not VAT track it.

• The national remeasure (R1). One spine build and one national line run next, on the policyengine-uk version Microcosm locks then. The VAT, fuel-duty and energy engine fixes are open in Re-average fuel duty and other blended rates after mid-year reforms policyengine-uk#2185, Put road-fuel litre divisors on DESNZ pump prices policyengine-uk#2186, Split VAT grossing into household share and survey coverage policyengine-uk#2187 and Model the temporary 0% VAT on GB domestic electricity policyengine-uk#2189.
• Rebinding obr.vat. That waits until policyengine-uk splits its VAT factor (policyengine-uk#1996).
• The furnishings and health over-draw surfaced by the division capture, and the extreme energy values that the housing total now contains for a small number of households. Those values predate this PR.
• Who holds the zero-emission cars. The fuel flag is a draw at the car stock's non-zero-emission share, random with respect to income. The stage notes record it as a known residual.

Merge order with #1198 and #1145

• One consistent UK target surface: enforced cross-grain reconciliation, declared uprating holds, nation household controls and the 942a1fa re-pin (#1123 PR A) #1198 pins the feed to 942a1fa, which descends from this PR's 1ee7dfe.
  • If One consistent UK target surface: enforced cross-grain reconciliation, declared uprating holds, nation household controls and the 942a1fa re-pin (#1123 PR A) #1198 lands first, this PR drops its header restamps, re-vendors its two resources at 942a1fa and regenerates the membership, coverage manifest and fixture.
  • If this lands first, One consistent UK target surface: enforced cross-grain reconciliation, declared uprating holds, nation household controls and the 942a1fa re-pin (#1123 PR A) #1198 re-vendors desnz_road_fuel_by_vehicle.json; a mechanical merge would leave it on the old pin.
• UK landlords: bind the SPI net property amounts, leave property_wealth to the engine, add receipts and finance costs #1145 overlaps on the gates, digests, fixture, parity receipts and membership. Those are regenerated, not hand-merged.

Testing

• Targeted engine-free tests for each commit's surface, and a broad funnel at the head (counts in the review replies).
• During the rebase onto 7f235941c, a check ran after every commit: each loads the UK spec, has a current release-input coverage manifest, and has current gate digests mirrored in microcosm-data. The review commits pass the same check.
• The charter-H2 spine fixture is regenerated by each commit that moves it. The commits before 3ed090d03 fail test_h2_uk_spine_parity on their own.

🤖 Generated with Claude Code

[vahid-ahmadi]

Automated review pass (Claude Code, high effort) — round 1 at 2e32d7bd

Verdict: nothing blocks this. The stage steps are sound, declared and gated, and the four rows it moves out of the fit are justified and signed. Two should-fixes on the register wording and the bound row, and one to note for the engine side before the national remeasure.

What checks out

• level_road_fuel is one factor on petrol and diesel at prior weights, so the drawn mix and each household's relative spend are kept (lcfs_consumption.py:1151-1182). The other-fuels share comes from the donor's own c72213, and the step refuses a fiscal-year level or a share outside [0, 1).
• redraw_zero_road_fuel is seeded, identity-keyed and redraws only flagged households that drew zero, so households without a fuel car stay at zero.
• recompose_from_remainder refuses a declaration that differs from the parents and components it recomposes, and floors the remainder at zero.
• Calendar 2024 is applied everywhere: energy sums the four Energy Trends quarters of 2024 (energy_pricing.py, period_quarters); road fuel levels to ONS 07.2.2 for 2024; bus yields are re-priced to 2024. The consumption_basis gate holds every level step to that year.
• Four new release_blocking gates (road_fuel_incidence, road_fuel_level, recomposed_totals, consumption_basis) are in gates.json, with digests mirrored.
• The four exclusions are signed, carry a tracking issue, an adjudication with receipts and an expiry (2027-01-08), and stay measured through tools/diagnose_uk_consumption_taxes.py.
• The Chronicle pin 1ee7dfe is the merge commit of chronicle#323 on main. The selectors resolve 07.2.2 as text, and the 217 restated rows keep their values under the new keys.
• Tests: the 17 touched test files pass locally (778 passed, 6 skipped). Against the base 7f235941, the new stage-health tests fail and the new diagnostics test doesn't import, so they cover the change.

Should

1. The electricity and gas exclusion reasons describe the old basis (calibration_measure_exclusions.json:590, :599). They say the stage sets spend from "Energy Trends FY2024-25 domestic volume at QEP FY2024-25 average prices", but the stage now uses calendar 2024 (the 2026-10-09 ruling), and the quoted ratios (0.90, 0.81) were measured on the fiscal-year basis. Restate both reasons on calendar 2024 and re-quote the ratios, so the register matches what the stage does.
2. The bound row compares petrol plus diesel with all of 07.2.2 (uk_population_targets.json, ons.household_road_fuel_expenditure). The stage levels to 07.2.2 less the donor's other-fuels share (0.37%), but the target is the full class, so calibration pulls about 0.4% toward fuel buyers to close a gap that is only scope. Either scale the target by the same share the stage uses, or record it as a known 0.4% offset in the binding notes and the receipts.
3. Non-residents' spending inside the bound row. The notes say 07.2.2 is on the domestic concept and no bridge is applied. That's fine as a choice, but the frame is residents only, so the row is an upper bound. Add one line with the size of the effect (or that ONS gives no split), so a later reader knows which way it leans.

Nits

4. The model year: policyengine-uk reads the H5's 2024 as fiscal 2024-25, while the frame is priced at calendar 2024, about a quarter apart. Worth one sentence in the stage notes, since every levelled column inherits it.
5. If the LCFS diary has a line for vehicle electricity inside 07.2.2, it should sit with c72213 in the other-fuels share; otherwise a sentence saying it doesn't exist.

Knock-on for policyengine-uk

• #2186 (DESNZ litre divisors): still holds. Its divisors are DESNZ calendar-year averages, the same basis this stage prices road fuel on, and its spending proxy is derived to keep litres on the volume path whatever the level. What moves is the level: petrol plus diesel for 2024 goes from £34.81bn on the certified release to £36.89bn here (+6.0%), so litres and fuel duty rise by the same 6%, about 25.6bn litres in 2024 and about £13.3bn of duty in 2025. That matches the register's −8.1% against the OBR cars line. The new ONS row will read about −1% at design weights, because the engine's 2024-to-2025 spending growth (−6.2%) is a little below ONS's (−5.2%).
• #2187 (VAT coverage split): the 0.67 coverage was derived on the certified release. Here total consumption is 0.8% lower and energy is levelled separately, which by itself moves coverage to about 0.665, still 0.67 to two places, but it should be re-derived on the R1 remeasure. Once #2187 is in the locked version, the obr.vat exclusion's own retire condition is met ("rebind when a locked policyengine-uk splits the factor"), and on the certified release the split factor put VAT within 0.4% of OBR.
• #2189 (electricity and gas VAT base): energy here is priced at QEP calendar-2024 average prices including VAT, so the 5/105 extraction is still right. #2189's description says the bills are at Ofgem cap rates; that should say QEP average prices paid. With electricity at £29.5bn and gas at £19.2bn, domestic energy VAT is about £2.32bn, against about £2.02bn on the certified release and HMRC's implied £2.33bn, and the 2026-27 zero-rate cost rises from about £0.62bn to roughly £0.70bn.

[juaristi22]

Round 1 addressed at 3523109, one commit on top of 2e32d7bd, so the delta reviews on its own. The branch is not rebased: main has not touched the spec, gates, feed or generated files since the base. The PR now closes #1113.

Should

1. The energy exclusions describe calendar 2024. Both reasons now name the DESNZ volume over the four quarters of 2024 at QEP calendar-2024 prices, and quote that year beside 2025: ONS 04.5.1 is £26.10bn against the stage's £29.52bn (0.88), and ONS 04.5.2 is £14.64bn against £19.23bn (0.76). One correction: the 0.90 and 0.81 ratios were not on the fiscal-year basis. They compare ONS for calendar 2025 with Energy Trends 2025 volume at QEP 2025 prices, so they stay beside the 2024 ones. María approved the restatement; each adjudication records it and keeps the 2026-10-08 decision date.
2. The 0.4% scope gap is recorded, not scaled (María's ruling of 2026-10-10). The binding notes now say the row reads about 0.4% low from scope alone: the stage levels petrol plus diesel to 07.2.2 less the donor's other-fuels share (0.37%), and the target is the full class. Scaling the target would put a diary-derived constant into the contract. The receipts carry the offset too.
3. The domestic concept is sized as far as ONS allows. The notes give both flows for 2024: £64.7bn of non-residents' spending in the UK (3.8% of the domestic total) and £78.9bn of residents' spending abroad. Consumer Trends publishes them for total spending only (sheet TOURCN) and splits them by no class, so their size for road fuel is not known. One nuance: the row is not strictly an upper bound. It includes non-residents' purchases here and leaves out residents' purchases abroad, and across all spending the second flow is the larger.

Nits

4. Model year. The stage notes now say policyengine-uk applies the fiscal 2024-25 rules to model year 2024 (it samples each parameter at 30 April), so the levelled columns sit about a quarter before the midpoint of that policy year. They also say the calibration binds calendar-year facts as published and fiscal-year facts at the calendar window (María's rule of 2026-09-22).
5. Vehicle electricity. The 2023-24 derived household file has three 07.2.2 lines (c72211, c72212, c72213) and none for vehicle electricity, so nothing joins the other-fuels share. The stage notes say so.

policyengine-uk knock-ons. Noted for the national remeasure, which has not run yet. It will re-read the ONS row at design weights and the VAT coverage. The obr.vat exclusion's retire condition gets checked once a version with PolicyEngine/policyengine-uk#2187 is locked. The wording point on PolicyEngine/policyengine-uk#2189 (QEP average prices paid, not Ofgem cap rates) is for that PR.

Verification at 3523109: 870 tests pass across the register, contract, source-stage, LCFS stage, stage-health, coverage, gate, country-spec and microcosm-data contract suites, with test_h2_uk_spine_parity and test_uk_graph. load_country_spec("uk") loads, the coverage manifest --check passes and the mirrored gate digests are current. The stage notes feed source_stages.json, the coverage manifest and the H2 fixture bundle, which are regenerated; no stage contract hash or gate digest moves.

[juaristi22]

Independent adversarial review (Claude Code) at 352310965

Static review only: I read the code, the registers and experiments/1113-consumption-tax-levels-receipts.md against the base 7f235941c, and ran nothing. I've tried not to repeat round 1. Where a round-1 point was only partly fixed, I say so.

Verdict. The three stage steps do what the body says:

• redraw_zero_road_fuel is identity-keyed and touches no global RNG, so later stages don't move.
• level_road_fuel is one factor at prior weights, with a weighted c72213 share.
• recompose_from_remainder floors the remainder at zero.
• The re-pin changes headers only, and the restated rows keep their values.

But the headline "one calendar-2024 basis" isn't true for the diary columns, and the gate meant to hold it only checks a label. The PR also closes the issue its new exclusions point at. Several gate and exclusion texts still describe rows that no longer bind.

Should fix

1. The diary columns are not at calendar-2024 prices, and consumption_basis can't tell. Round 1's "calendar 2024 is applied everywhere" is not right.

   • The uprating step: spec/sources.yaml:2846-2849 uprates the LCFS 2023-24 diary (April to March) by one year of engine CPI, from_period: 2023 → to_period: 2024 at January-first instants. A one-year ratio carries the diary's midpoint (about October 2023) to about October 2024. That's FY2024-25 prices, a quarter-year past calendar 2024.
   • The code says so itself:
     • donor_uprating.py:3: "one year behind the FRS 2024-25 spine".
     • The stage notes (sources.yaml:3286) say the diary "is uprated to the FRS 2024-25 base year", in the same paragraph as "every household spend column is at calendar-2024 prices … the diary columns by the CPI uprating to 2024".
   • The gate checks a label: stage_health.py:2681 sets found["donor_uprating"] = ("calendar_year", uprating.get("to_period")). The period type is made up, and only to_period == 2024 is checked.
   • Affected: the ten non-levelled divisions, the housing and transport remainders, and Wales's bus fares. That's most of the frame's consumption, so a quarter-year of CPI is likely several £bn, more than the energy fiscal-to-calendar move this PR made.
   • The PR already has the right bridge: it bridges Scotland's bus index with 0.75 · FY + 0.25 · FY+1 (bus_fare_pricing.py:248).
   • Receipts: experiments/…:143 says "the diary columns … were already on calendar 2024 and don't move", which is wrong for the diary.
   • Fix: either bridge the diary the same way, or restate the claim and the gate as "FY2024-25 for the diary".

2. "Closes UK domestic electricity and gas: record why ONS 04.5.1/04.5.2 is preferred over DESNZ × QEP, reconcile the two, and declare the price basis PE-UK should project from #1113" orphans three new exclusions.

   • ons.household_electricity_expenditure, ons.household_gas_expenditure and obr.fuel_duties_cars all have "tracking": "microcosm#1113". The two energy rows retire "if the microcosm#1113 reconciliation finds a concept difference to bridge".
   • This PR rules the other way (D1) and does no reconciliation. After merge, three signed narrowings would point at a closed issue whose work was never done, and the expiry error would send people there.
   • Fix: open follow-up issues, one for the energy-concept reconciliation and one for business use of cars, and repoint tracking. Or drop "Closes".

3. Round 1's stale-wording fix reached the exclusion register but not the gate manifest or the other registers.

   • gates.json:333 (energy_rake notes): still says the level is "over the four FY2024-25 quarters".
   • gates.json:888 (uk_aggregate_admin): says "the calibrated frame is held to the bound ONS 04.5.1 electricity and 04.5.2 gas rows". Both are now measure-excluded. The notes are hashed into gates_manifest_sha256, so every signed battery report attests this.
   • target_reference_signed_exclusions.json:55: obr.fuel_duties is signed out because "obr.fuel_duties_cars binds the cars receipts instead". It no longer binds, so no fuel-duty row binds at all.
   • target_reference_signed_exclusions.json:62-63: COICOP 04.5 is "superseded_by_component_targets", the per-fuel rows that "bind per fuel". They no longer bind.
   • Also stale:
     • uk_population_targets.json:708 ("F1 binds ONS … 04.5.1/04.5.2");
     • energy_pricing.py:5 ("base-year (FY2024-25) frame");
     • lcfs_consumption.py:1800 ("FY2024-25 regional average prices").
   • The substance behind the wording: after this PR, no release-blocking check holds calibrated domestic energy or the VAT base. The report-only consumption_drift is all that sees them, and the receipts show unbound VAT moving (+35%). That is a decision to record, not just text to fix.

4. Two of the new gates can't fail on a real receipt, and they replace the frame bounds that could.

   • recomposed_totals: recompose_from_remainder sets rest = max(drawn_total − drawn_parts, 0) and total = rest + parts (lcfs_consumption.py:1313-1315). So rows_below_components is always 0, minimum_remainder is always ≥ 0, and the weighted checks are identities of the same receipt.
   • The removed bounds: the PR deletes housing_water_and_electricity_consumption [-20000, 300000] and transport_consumption [0, 200000] from lcfs_consumption_support_bounds.json. The terminal uk_support battery applied those to the emitted frame. The gate note says this check "holds them instead", but it doesn't.
   • road_fuel_level: checks the step's own frame_after, recomputed from the same vendored row. A unit error in that row, or a later stage moving these columns, passes.
   • Fix: keep frame-level bounds on the two totals. Dropping petrol and diesel follows the existing raked-columns convention, so that part is fine.

5. The redraw now guarantees fuel spend to battery-electric households.

   • The flag is num_vehicles > 0 & Bernoulli(1 − ZEV share) (lcfs_consumption.py:1973-1975), blind to fuel type.
   • Before this PR, a flagged EV owner could draw zero; redraw_zero_road_fuel now forces them positive.
   • Small at today's ZEV share, but it grows each year. It's worth conditioning the flag on fuel type, or at least recording it as a known residual.

Minor

6. Bus fares are levelled but transport isn't recomposed around them.

   • bus_fare_spending is in UK_LCFS_RAKED_COLUMNS and is re-priced here (−2.05% outside London).
   • UK_LCFS_RECOMPOSED_PARENTS["transport_consumption"] nets and adds back only petrol and diesel (:287-296).
   • So transport can sit below bus + petrol + diesel, and "no household's … transport total is below its levelled parts" holds only if bus doesn't count. The end-to-end test checks only transport >= road_fuel.

7. "Consumption moves from £979.0bn to £971.4bn (−0.8%)" isn't this PR's effect.

   • The only divisions the steps rewrite are housing (180.1 → 173.5) and transport (173.0 → 181.5), which net +£1.9bn.
   • The other −£9.5bn is re-run drift; the receipts concede "differ by the re-run as well".
   • Round 1's VAT-coverage knock-on (0.67 → 0.665) rests on that −0.8%.

8. Exclusion reasons mix bases.

   • Electricity: "cut electricity 8.6 percent … (82.9 TWh … against 93.0 TWh)", but 82.9/93.0 is −10.9%.
   • Gas: "18.9 percent" is the masked-arm attribution, while 260.1 → 205.2 TWh is −21.1%.
   • Cars: petrol "+21 / diesel +34" are UK: port the remaining uk-data fixes tracked in #1095, engine items included #1121's total design-to-final moves. The isolating arm gives +18%/+36%.
   • "Masked one at a time": each entry says "a re-solve with [the row] masked changed no other row's fit". Only a joint four-row mask was run, on the 2026-09-30 problem without this PR's stage levels.
   • "1,086 other rows": that's 1,090 − 4 from that older problem, not the 1,167-row UK: port the remaining uk-data fixes tracked in #1095, engine items included #1121 head the body places it under.

9. Question: is the bound row's design-weight gap only the 0.4% scope gap?

   • ons.household_road_fuel_expenditure binds the published calendar-2025 value (35.099bn), not uprated. The engine projects the levelled 2024 frame forward with its own growth.
   • ONS's own 2024-to-2025 move is −5.2% (37.025 → 35.099bn). Any difference between that and the engine's fuel growth adds to the 0.37% scope gap, and the calibration closes the combined gap by moving weight toward fuel buyers.
   • The binding notes size only the scope part. It's worth reading the row at design weights in R1 and recording the total.

10. "The three compile-parity receipts are unchanged" holds for the re-pin commit only. At PR level, the incumbent-2025 and production-2023 receipts each gain one ledger_only row (1,257 → 1,258 and 924 → 925), from 149fb61b5, and the changelog repeats the claim unscoped.

11. Reproducibility: run_stage_dev.sh, the Group B config, the M1–M4 measurements and the masked re-solve aren't committed. The outputs live under data/ukds/acceptance/… and the plan outside the repo. docs/evidence/uk-a16-v20/scripts/ is the precedent for committing them.

12. Questions and nits.

    • What does LCFS p537 cover? If it's all of 04.5, then where derive_energy_from_lcfs splits the whole of p537 into electricity and gas (:1948), oil and solid fuel are subtracted with the drawn energy. Then the housing remainder doesn't hold them, despite the comment at :284.
    • The diagnostics can abort the build: _ons_annual_value raises unless exactly one row matches (with no price-basis or seasonal-adjustment filter), and desnz_cars_share_by_fuel raises for a year outside 2022–2024.
    • The 07.2.2 target selector doesn't pin the price_basis, consumption_concept or seasonal_adjustment dimensions that chronicle#323 added, and has no row-count guard, unlike the vendored selections.
    • Northern Ireland's bus factor on England's index picks up England's January 2025 fare-cap rise (191.1 → 201.8).
    • Receipt arithmetic: "33.2%" is 16,362/48,529 = 33.7% (label which is weighted). "66% more per household" is 1.625 on the receipts' own numbers. And 84.5 → 78.9% appears for both the UK: port the remaining uk-data fixes tracked in #1095, engine items included #1121 head and the 2026-09-30 build; is that a copy error?
    • docs/uk-chronicle-feed-repin.md:124 has no blank line before the new paragraph.

Interaction with #1198 and #1145

• Feed pins: One consistent UK target surface: enforced cross-grain reconciliation, declared uprating holds, nation household controls and the 942a1fa re-pin (#1123 PR A) #1198 pins the feed to 942a1fa, which strictly descends from this PR's 1ee7dfe (13 commits ahead, 0 behind). The 42 Consumer Trends rows both PRs vendor are byte-identical.
• If One consistent UK target surface: enforced cross-grain reconciliation, declared uprating holds, nation household controls and the 942a1fa re-pin (#1123 PR A) #1198 lands first:
  • Drop this PR's header restamps. Keeping "ours" in the conflict would downgrade main to 1ee7dfe and revert chronicle#314's 2025 values.
  • Re-vendor ons_household_expenditure_facts.json (93 rows, 3 selections) and desnz_road_fuel_by_vehicle.json at 942a1fa.
  • Regenerate the membership, coverage manifest and H2 hashes, and rewrite the re-pin changelog and doc paragraph.
• If this lands first:
  • desnz_road_fuel_by_vehicle.json is a new file with no textual conflict. A mechanical merge leaves its 1ee7dfe header, and vendored_rows() then refuses at stage time.
  • So One consistent UK target surface: enforced cross-grain reconciliation, declared uprating holds, nation household controls and the 942a1fa re-pin (#1123 PR A) #1198 has to re-vendor it, not just resolve conflicts.
• UK landlords: bind the SPI net property amounts, leave property_wealth to the engine, add receipts and finance costs #1145 overlaps on 11 files (gates, contract.py digests, H2 fixture, parity receipts, membership). Regenerate those rather than hand-merge.

Checked and fine:

• The re-pin: every +4/−4 vendored file changes only the four feed-identity fields, and chronicle_feed.json shows 352,549 rows.
• Ratios: 0.88 and 0.76 (2024), 0.90 and 0.81 (2025), and £36.887bn = 37.025 × (1 − 0.00372).
• Mirrors: gate digests recompute to the mirrored values, and only stage_contract_sha256 moves in uk_spine.json.
• Fixture: lcfs_household.csv is synthetic.
• obr.vat: nothing else depends on it being bound, and its retire condition is checkable against policyengine-uk#2187.
• Round 1's Should-2 and Should-3 and the two nits are fixed.

[juaristi22]

The independent adversarial review above is addressed at 3f5743165, three commits on top of 352310965, so the delta reviews on its own. The branch is not rebased. Item numbers are the review's.

Should fix

1. The diary columns are now at calendar-2024 prices, and the gate checks it. The review is right: a full year's CPI step (1.025) carried the April-to-March 2023-24 diary to about October 2024. The twelve divisions and bus fares now declare the diary-year window (from_window: fiscal_year, with the operation's fiscal_year_bridge). The engine's CPI index is calendar-year, so the diary year sits at three quarters of 2023 and one of 2024, and the step to calendar 2024 is 1.0186. consumption_basis reads each column's declared window and the bridge weights; a column on a full step fails it. Income predictors keep the full year's step, to the FRS 2024-25 incomes. On the head-r3 spine the ten diary divisions move −0.7% against main's stage.
2. "Closes UK domestic electricity and gas: record why ONS 04.5.1/04.5.2 is preferred over DESNZ × QEP, reconcile the two, and declare the price basis PE-UK should project from #1113" no longer orphans the exclusions. The open items are one follow-up issue, UK consumption-tax levels: open items after #1167 #1207, under the two deferral registers: the calibration-target items under WS-C deferred adjudications and the reconciliation layer #736 and the spine items under WS-E spine deferrals: known bugs and signed inconsistencies awaiting their increment #796. The three exclusions now track UK consumption-tax levels: open items after #1167 #1207. The two energy rows retire on its reconciliation of ONS spending with DESNZ volume at QEP prices, and the cars row on its sizing of business use of cars. The PR still closes UK domestic electricity and gas: record why ONS 04.5.1/04.5.2 is preferred over DESNZ × QEP, reconcile the two, and declare the price basis PE-UK should project from #1113.
3. The stale texts are restated, and the decision behind them is recorded.
   • gates.json: the energy_rake level is calendar 2024. The energy_rake and uk_aggregate_admin notes now say that since D1–D3 no bound row and no release-blocking check holds calibrated domestic energy, the VAT base or fuel duty; the stage gates hold the levels at prior weights, the bound ONS row holds road-fuel spend, and consumption_drift reports the rest.
   • The signed exclusions for obr.fuel_duties and COICOP 04.5 no longer say the rows that replaced them bind.
   • The contract's unmapped declarations, energy_pricing.py, the stage docstring and the stage notes are corrected too.
   • A release-blocking limit on the drift is a possible follow-up once R1 has measured it. It is not added here without that measurement.
4. The two totals keep frame-level bounds. Housing and transport are back in lcfs_consumption_support_bounds.json (−20,000 to 300,000 and 0 to 200,000), so the terminal uk_support gate holds the emitted totals; on the head-r3 spine both totals are inside them. The recomposed_totals and road_fuel_level notes now say what each can and cannot catch.
5. The fuel flag does not hold electric households that the redraw then fuels. The flag is a draw at the car stock's non-zero-emission share, keyed on the household identity. It is the stage's own designation of who buys fuel: an unflagged vehicle owner is the zero-emission household and has no fuel, before and after this PR. The third of flagged households at zero before the redraw was the two-week diary, not electric cars. What stands from the point is that the designation is random with respect to income and all-or-nothing per household. The stage notes now record that as a known residual.

Minor

6. Transport is recomposed around the priced bus fares as well as the levelled fuel. No household's transport total is below its petrol, diesel and bus fares on the head-r3 spine, and the end-to-end test checks all three.
7. The −0.8% was re-run drift. Main's stage re-run on the same spine reads £969.6bn against the spine's stored £979.0bn. Against that base the PR read +0.2% (£971.4bn) before this round, and reads −0.3% (£966.8bn) with the diary bridge. The receipts and the PR body now compare with the base run. Round 1's VAT-coverage note rested on the −0.8%, so it should be re-derived on R1.
8. The exclusion reasons separate their bases. Each gives the design-to-final move and the masked-arm comparison apart, says the four rows were masked together, and places the re-solve on the 2026-09-30 problem (1,090 rows).
9. The bound row's design-weight gap is more than the scope offset. ONS's own move from 2024 to 2025 is −5.2%; policyengine-uk 2.122.2 carries petrol and diesel −2.7%. With the scope offset the row reads about +2.3% on that version, so the calibration moves weight away from fuel buyers. The binding notes and receipts give that desk figure, and R1 records the measured one.
10. The parity claim is scoped to the re-pin. The new road-fuel row adds one ledger-only row to each national receipt; the changelog, the re-pin doc and the receipts say so.
11. The scripts are committed under docs/evidence/uk-consumption-taxes-1113/, with their aggregate outputs. Household dumps, weight vectors and the stage's exact donor minima stay out. M1's volume figures have no saved script; the README says so.
12. Questions and nits.
    • p537 is all fuel, light and power: £60.1bn on the donor, against £54.9bn for the electricity and gas derived from it. For 2,978 of the 4,202 diaries the derivation splits all of p537, so their other fuels leave with the drawn energy. For the 1,224 that report a combined payment, the rest (£5.2bn) stays in the housing remainder. The comment, notes and receipts claimed the remainder held all the other fuels; they are corrected.
    • Diagnostics no longer abort: the division capture records why it has no figures, and the litres audit falls back to the OBR cars share. Their row lookups pin the price basis and seasonal adjustment. A vendored resource that lags the pin still fails.
    • The target selectors pin price_basis, consumption_concept and seasonal_adjustment.
    • Northern Ireland's bus factor is left on England's index; the stage notes name it a declared proxy that carries the English fare-cap rise.
    • Arithmetic: 33.7% of flagged rows, 33.2% of their weight; 63% more furnishings per household at B3 (61% now). The 84.5% to 78.9% is not a copy error: UK: port the remaining uk-data fixes tracked in #1095, engine items included #1121's head reads 84.54% and 78.87% from its files, the 2026-09-30 build 84.52% and 78.93%.
    • The re-pin doc has its blank line.

Interaction with #1198 and #1145. Agreed, and the PR body now carries the merge-order steps.

One new observation. The housing total now contains the energy columns, and a small number of households carry extreme electricity and gas values. Values of that size predate this PR, and nothing bounds the energy columns themselves. #1207 carries it, with the furnishings over-draw and the other open items.

Verification at 3f5743165: 1,487 tests pass across the LCFS stage, stage-health, donor-uprating, energy and bus pricing, source-stage, coverage, gate, country-spec, feed, register, contract and microcosm-data contract suites, with test_h2_uk_spine_parity and test_uk_graph. The licensed round-trip of the support bounds passes too. load_country_spec("uk") loads, the coverage manifest --check passes and the mirrored gate digests are current. The H2 fixture is regenerated; only lcfs_consumption's contract hash moves. All seven lcfs gates pass on the head-r3 spine.

[juaristi22]

Re-review (Claude Code) at 3f5743165

Static review of 352310965..3f5743165 against the previous review and the round-2 reply. I ran nothing. The CI run for this head was still queued when I looked.

Verdict. Every item from the previous review is fixed or answered convincingly, and the diary bridge is right: 1.55763 / (0.75 · 1.51964 + 0.25 · 1.55763) = 1.018633 on the engine's calendar-year CPI. Two disclosure-control items went to the author separately. The rest below is minor.

Previous items

#	Status	Note
1 Diary basis	Fixed	The 12 divisions and bus declare from_window: fiscal_year with the 0.75/0.25 bridge. consumption_basis fails a missing window, a missing column or wrong weights (stage_health.py:2702-2725). Income predictors keep the full step, which is consistent.
2 #1207	Fixed	#1207 is open, the three exclusions track it, and the retire conditions are reworded.
3 Stale text	Mostly fixed	gates.json 333/888, the signed exclusions, energy_pricing.py and the stage notes are fixed. Leftovers are in item 7 below.
4 Support bounds	Fixed	Housing and transport carry main's bounds again, and uk_support applies them.
5 EV flag	Answered	Accepted: the recipient has no fuel type, so the flag is the only zero-emission identity. One residual isn't recorded: the rate is the car-stock ZEV share applied per household. Battery-only households are a smaller share of vehicle-owning households, because many EV households also run a combustion car, so the flag likely over-designates no-fuel households.
6 Bus in transport	Fixed	UK_LCFS_RECOMPOSED_PARENTS nets and restores bus, and the end-to-end test checks fuel + bus.
7 Consumption figure	Fixed	Base 969.58, then 971.44 (+0.19%) before this round and 966.76 (−0.29%) now. The ten divisions move −0.68%, and the bridge alone is −0.62%.
8 Exclusion bases	Fixed	The figures match m4_report.json.
9 Bound-row gap	Fixed	pe-uk 2.122.2 carries both fuels at −2.691% for 2025: 36.887 × 0.97309 = 35.894 against 35.099, which is +2.26%. So the calibration moves weight away from fuel buyers.
10 Parity scope	Fixed
11 Scripts	Fixed	receipts.json hashes match. M1 has no script, as the README says.
12 Nits	Mostly fixed	The p537 split matches the derivation masks, and the selectors pin all three dimensions. Minor item 1 below has a new fail-open in the diagnostics.

New

• Two disclosure-control items have gone to the author separately.
• Hardening: the sanitiser docs/evidence/uk-consumption-taxes-1113/scripts/sanitize_outputs.py:28 drops only donor_floor. stage_dev_run.py:145 also copies support_clip entries, and those carry donor_min and donor_max. The block is empty today only because the named columns don't exist. Add both keys to DROPPED, or drop the block.

Minor

1. The diagnostics now fail open on bad data, not only on a missing year. DiagnosticUnavailableError also covers ambiguous or duplicate rows (lcfs_consumption.py:616, :753), so a data-integrity error is recorded as "unavailable". And a resource vendored at the pin but missing the year quietly drops the litres audit to the uniform OBR share; only the receipt notes it.
2. The recomposed totals carry an uprating residue.
   • Parents: take the CPI bridge (×1.0186).
   • The parts netted from them: stay at their own factors. Drawn energy isn't uprated. Petrol and diesel take the litre-based spending proxies (diesel ×0.898).
   • Residue: the remainders keep about 0.019 × drawn energy (roughly £1bn in housing) and the fuel factor gaps (roughly £1.7bn in transport) on the committed totals. So the totals aren't strictly on one basis.
   • Fix: rescale each drawn part by parent factor ÷ part factor before netting, or record the residue.
3. UK consumption-tax levels: open items after #1167 #1207's R1 acceptance contradicts its own desk figure. It expects levelled road fuel to move under 2% from design to final, but the bound row reads +2.3% at design, so fitting it moves road fuel about 2.2%.
4. Nits.
   • consumption_basis checks the declared window and weights, but not that the receipt's factor equals after / (0.75 · before + 0.25 · following), though both are recorded.
   • Stale text remains:
     • the uncarried_spend docstring (lcfs_consumption.py:1305-1307, "stay inside the housing remainder");
     • binding notes saying 04.5 is "bound per fuel" and that obr.fuel_duties_cars is "Bound at FY2024-25";
     • donor_uprating.py:3.
   • Per-division movement: the bridge is a uniform −0.62%, but divisions move from −1.4% to +2.2% from re-run noise. That matters when reading division_capture.
   • Local paths: the scripts and output spine fields carry local absolute paths.

No tests were weakened and no counts were edited to match. The count changes follow the code.