[juaristi22]

What and why

Budget 2026 raises every company car appropriate percentage by one point in 2030-31: Table 6.1 row 86, +£310m in 2030-31 and +£345m in 2031-32 (OBR split: income tax 226.2, NICs 92.4). The dashboard can't score it (PolicyEngine/uk-autumn-budget-dashboard-2026#316), because neither the engine nor the data had company cars.

• The engine. Model the company car and car fuel benefits, with Class 1A NICs policyengine-uk#2750 shipped in 2.165.0. It taxes the company car and car fuel benefits, and the Class 1A NICs on them, from nine person inputs that default to no car.
• The data (this PR). It locks 2.165.0 and adds a last stage, company_cars, that fills six of those inputs.

Without the stage, a file on 2.165.0 still scores the rise at £0.

Status: draft, CI expected red

Three things are still to come:

1. The HMRC facts. They come from Add HMRC benefits in kind statistics, July 2026, and keep one OBR EFO number per line and year chronicle#399 (HMRC Benefits in kind statistics, July 2026: 544 facts, raw workbook in R2). Once it merges, this PR:

   • re-pins the feed (docs/uk-chronicle-feed-repin.md);
   • vendors uk/hmrc_benefits_in_kind_facts.json and adds its country_package.json entry.

   Until then, every test that builds the UK graph fails on the missing resource, and so do the stage-plan tests.

2. The regenerations. The H2 fixture and the coverage manifest are regenerated once the facts are vendored; until then their parity and preflight tests fail.

3. The build. A pair build on 2.165.0 against a control on the same base, with the results below. The machine order is María's.

Merge order

• Base. Main after Budget 2026 ISA data, PR B of #1418: ISA subscriptions per adult, the Lifetime ISA anchor, additive ISA income, policyengine-uk 2.164.2 #1421: it carries Budget 2026 ISA data, PR B of #1418: ISA subscriptions per adult, the Lifetime ISA anchor, additive ISA income, policyengine-uk 2.164.2 #1421's 2.164.2 lock, so this PR's lock step is 2.164.2 → 2.165.0.
• UK: take up #1095's actionable items on policyengine-uk 2.165.1 (draft) #1422 (draft) locks policyengine-uk 2.165.1. Whichever of the two lands second:
  • merges main and keeps the newer lock and stamps;
  • regenerates the shared surfaces: the gates.json digests, the coverage manifest and the H2 fixture.
• One overlap with UK: take up #1095's actionable items on policyengine-uk 2.165.1 (draft) #1422. It signs the nine company_car_* inputs at their default (UK_ENGINE_DEFAULT_INPUTS). Whichever lands second removes only the six this stage writes. Availability, private-use payments and the 6 April 2020 registration flag stay signed.

Changes

1. policyengine-uk 2.165.0.

   • The lock, the floors, the version mentions and the US worker identity's lock digest.
   • The uprating and bonus-sacrifice pins keep every value.
   • Concept coverage gains the nine company car inputs.
   • 2.164.3 also counts other_investment_income in market and HBAI income, which the spine fills (UK: data fixes for the Budget 2026 household measures #1418), so HBAI-based outputs move too.

2. The FRS job tab's company car answers. The root stage carries four columns per person from their car job with the lowest JOBTYPE (one car per person, as in the engine):

   • the CARVAL band, or "unknown";
   • FUELTYP, or "unknown";
   • EXPBEN03 (private fuel);
   • CARAMT where CARCON says the employee contributed.

   They are dropped at the release boundary.

3. The company_cars stage, the last in the roster, after isa_subscriptions, whose income cells read employment_income. Nothing reads the pay it rewrites. Every draw is keyed on identity.

   • Holders. FRS-channel people keep their car. An SPI-channel person keeps their twin's only where their drawn EPB is positive.
   • Fuel. From FUELTYP: a hybrid is an alternative fuel; a diesel meets RDE2. An unknown fuel draws at HMRC Table 2C's shares.
   • List price. Drawn within the CARVAL band from WAS round-8 company cars' list prices, grown by the engine's CPI index from 2021. The open and unknown bands draw by fuel class. A cell with fewer than 3 WAS households (the floor of 3) is drawn uniformly.
   • CO2. Zero for an electric car. For the others, HMRC Table 2A's shares among cars that emit CO2, hybrids first into 1 to 50 g/km, uniformly within each band; a car at 1 to 50 g/km takes a 50-mile electric range.
   • Defaults kept. Availability, payments for private use and the 2020 registration flag keep the engine's defaults.

4. SPI-channel pay. It is the tape's PAY + EPB + TAXTERM, so the engine's car and fuel benefits come out of employment_income, capped at EPB (María's ruling of 30 October):

   • private fuel goes first where its benefit alone exceeds EPB;
   • a car that still doesn't fit has its list price scaled down, with its contributions dropped;
   • below £5,000 the car is dropped.

   Pay plus the benefits is the tape's total on every such row.

5. Targets. hmrc.spi.employment_income.amount_by_total_income_band and the local hmrc.employment_income.amount read employment_income_with_benefits_in_kind_annual, a new named term for employment income plus benefits_in_kind, which HMRC's employment income counts. The two count rows keep employment_income_annual.

6. The facts. hmrc_company_car_statistics.py reads Full eCPS input coverage: 158 required columns, 8 evidence-backed exclusions #399's 2024-25 rows by their source record ids, in persons and pounds:

   • a cell the feed lacks is refused, never read as zero;
   • a resource whose Table 2A or 2C total, or electric count, sits from Table 1 beyond each sheet's rounding is refused too.

7. Gates. uk_stage_company_cars_invariants is release-blocking and runs on the synthetic smoke too. It fails on:

   • a non-finite or negative value;
   • a holder without a price;
   • inputs on a person without a car;
   • an unknown fuel;
   • a CO2 figure missing for a car that emits it, or set for an electric car;
   • private fuel on an electric car;
   • contributions above the price;
   • an electric range outside 1 to 50 g/km;
   • an SPI benefit above EPB, or SPI pay off the tape's total.

8. Measured, not bound. No HMRC company car fact is a calibration target: the stage draws to Tables 2A and 2C, so binding them would be circular (María's ruling of 28 October on the ISA facts). The calibrated file is reported against Tables 1, 3A and 4.

Results

PENDING the pair build. Every figure is checked for disclosure before it goes in.

• Holders against HMRC's 920,000 (2024-25 provisional), and by income range against Table 3A.
• Car and fuel taxable value, income tax and Class 1A against Table 1 (£3,070m, £1,090m, £420m).
• The static 2030-31 cost of COMPANY_CAR_TAX_RISE_2030 against Table 6.1 row 86 (+£310m).
• Gate and fit movement: the re-pointed SPI rows, obr.income_tax, obr.ni_employer (Class 1A), and HBAI income (2.164.3).

Tests

• New:
  • test_uk_company_cars.py, engine-free and engine (22 tests);
  • test_uk_company_cars_stage_health.py;
  • the carrier tests in test_uk_frs_spine.py;
  • the named term in test_uk_measure_simulation.py.
• The engine tests on 2.165.0 show that:
  • the engine reads the stage's inputs;
  • SPI pay plus the engine's benefits is the tape's total to the penny;
  • the 2030-31 rise reaches the holders alone.
• The engine-free suite, run locally in chunks. Everything passes except the failures listed under Status, plus two local-only ones:
  • the uk-extra-absent test, since this environment installs the extra;
  • a live release-pointer test.
• After vendoring: the full suite, tools/run_integration_tests.sh and the H2 parity test.

Disclosure

The FRS and WAS are UK Data Service data (SN 9563, SN 7215). The stage's receipt carries weighted aggregates, rounded counts and support bands only. The HMRC figures quoted here are published.

Not in this PR

• Class 1A is not passed through to households (Pass Class 1A NICs through the employer NI incidence policyengine-uk#2752). The 2030-31 rise is scored statically, comparable with the OBR's costing.
• Vans, fuel provided for part of the year, optional remuneration arrangements, and the rest of the "not modelled" list in policyengine-uk#2750.
• Whether to anchor holders to Table 3A by income range, if the FRS count falls short of HMRC's. That is María's call after the results.
• The dashboard measure: policyengine.py certifies a build on ≥2.165.0, then the dashboard adopts the release and adds the measure.

🤖 Generated with Claude Code

[vahid-ahmadi]

Review of 21e3623 (draft, so an early look as a comment). First round; there are no earlier comments.

The stage writes what policyengine-uk 2.165.0 expects. I checked the nine inputs at the 2.165.0 tag. The six written ones have the right units, periods and enum names (GBP stocks uprated by CPI, g/km, miles, CarFuelType names, bool). The draws are keyed on identity, there is one car per person by lowest JOBTYPE, household weights pass through, and no company car fact is bound. The stage is the last spine stage, so calibration still runs after it: it adds no weights, but the solve then moves holders, which fits "measured, not bound". There are four shoulds and one nit.

Findings

1. should: on a frozen fleet, expect the static 2030-31 score well under £310m. The gap comes from the size of the base, not the EV share.

• The rise is one point in every band, zero-emission and the maximum included. So the extra taxable value is about 1% of total list price (less contributions), plus 1% of the fuel multiplier for each fuel-benefit holder. The frozen EV share barely moves the rise. It moves the baseline level instead (EVs go from 3% to 9% by 2029-30).
• The base the rise acts on is the build year's fleet, carried forward to 2030-31. Fuel type and CO2 are not uprated, list price and contributions rise by CPI, and the weights are not projected.
• Rough HMRC-implied base for 2024-25: Table 1 car taxable value £3,070m over Table 2A's band-weighted 2024-25 appropriate percentages gives about £27bn to £37bn of list price (depending on PHEV range and EV price premium). With CPI to 2030-31 and 40,000 fuel holders, that is about £330m to £440m of extra taxable value. At HMRC's 2024-25 ratios (income tax 35.5% of taxable value, Class 1A 15%), that is roughly £170m to £225m.
• The OBR's NICs line (£92.4m / 15%) implies about £616m of extra taxable value, roughly a £60bn base. Its income tax line gives the same 36.7% ratio. So even a file that hits HMRC's 920,000 holders and £3,070m exactly should come in around 30 to 45% short. Recipients rose from 840,000 to 920,000 in one year, and HMRC flags under-reporting of payrolled cars.
• Matching Table 1's taxable value does not pin this base: a file can hit £3,070m with a different EV mix and different prices. Suggest the receipt add weighted total list price (by electric and not), so Results can compare it with the HMRC-implied range and read the £310m gap as base size.

2. should: the SPI pay-identity check in the release-blocking gate can't fail.
company_cars.py:983-987 sets new_pay = pay - carve.benefit, then measures |new_pay + carve.benefit - pay|, so max_abs_pay_plus_benefit_error_gbp is zero by construction. That means uk_stage_company_cars_invariants (stage_health.py:4954-4960) never tests the claim in its notes. The real risk is the engine's benefit on the written inputs differing from the carved amount. The price scaling assumes the defaults for availability and private-use payments, and that contributions were dropped, so a later engine change would break it without the gate noticing. Suggest one more materialize of company_car_benefit + car_fuel_benefit on the output frame, comparing each SPI-channel row with carved.

3. should: one CI failure is not on the body's expected-red list.
engine-uk fails on tests/engine/uk/test_uk_annual_income_measures.py:156: set(receipt["computed_measures"]) == set(UK_ANNUAL_INCOME_MEASURES). The test never computes the new employment_income_with_benefits_in_kind_annual. This is independent of chronicle#399. Computing it there, and asserting it equals employment_income_annual when nobody has a car, would fix it. The job stops at this first failure ("1 failed" and nothing passed), so CI has not run the new engine tests in test_uk_company_cars.py yet.

4. should: the lock step pulls in more than the body lists.
2.164.2 → 2.165.0 includes 2.164.3 (other_investment_income in market_income, household_market_income and hbai_household_net_income) and 2.164.4 (also in capital_income). The body mentions only HBAI outputs. These engine values are stage predictors:

• was_wealth materializes household_net_income and capital_income (was_wealth.py:63-89);
• was_lisa reads household_net_income (was_lisa.py:164);
• lcfs_consumption and etb_services use hbai_household_net_income (lcfs_consumption.py:650, etb_services.py:62).

So the wealth, LISA, consumption and services draws can move for households with other investment income. Suggest adding these to Results' "gate and fit movement" list for the pair build. 2.165.0 also adds Class 1A to ni_employer, which the body already covers.

5. nit: unknown-fuel draws never produce a hybrid.
company_cars.py:588 draws only ELECTRIC, PETROL or diesel, and :920 flags hybrids from the FRS code alone. So "petrol/diesel including hybrids" from Table 2C never feeds the hybrids-first split into 1 to 50 g/km. The effect is small, but worth a sentence in the doc.

Checked and fine: the three defaults (availability 1, no private-use payments, the 2020 flag) only matter in 2020-22 or slightly raise each car's taxable value for part-year holders; the WAS donor (cars not vans, household weight, floor of 3); CARVAL/FUELTYP mapping and the open band; CO2 bands within Table 2A's bounds; the SPI carve-out arithmetic (fuel first, exact price scaling, £5,000 floor); the re-pointed SPI targets, since taxable_employment_income now includes benefits_in_kind, so total-income bands agree.

Verdict

Sound design. Items 2 and 3 are small code fixes. Item 1 is about setting expectations before the pair build, and item 4 is about what that build should report. Not ready until the #399 vendoring, the regenerations and the build, as the body says.

What I ran

Fresh clone at 21e3623 and uv sync --all-packages --locked (no UK engine extra), deleted afterwards; one process at a time.

• Engine-free pytest: test_uk_company_cars.py and test_uk_company_cars_stage_health.py gave 24 passed. With test_uk_frs_spine.py, test_uk_measure_simulation.py, test_uk_source_stages.py, test_uk_terminal_gates.py, the ISA, local-target, private-school and stochastic tests, test_engine_core_pins.py, test_country_spec.py and microcosm-data's test_contract.py, 848 passed and 4 failed. All 4 are KeyError: 'hmrc_benefits_in_kind_facts.json', as expected.
• I read the 2.165.0 variables, parameters and the company_car_tax_rise_2030 scenario at the tag, and the 2.164.2...2.165.0 compare.
• CI: engine-free fails on the missing facts and integration-uk on the same resource (expected), engine-uk as in item 3; lint, wheels and engine-us pass.

Not verified: the engine tests (no UK engine locally, no build or Modal), any FRS or WAS figure, and the score itself. Item 1's figures are rough arithmetic from published HMRC and OBR numbers.

🤖 Generated with Claude Code