Count corporate SDLT once and let SDLT abolition remove all of it

[MaxGhenis]

Fixes #2159

Summary

Corporate SDLT was counted twice in household_tax, gov_tax and pre_budget_change_household_tax, and abolishing SDLT left one copy of it in all three. This PR counts it once and makes abolition remove all of it. It also adds a tax-head partition, specified independently of the model's own lists, with tests that every total counts each of its heads exactly once.

National effect on certified data: all three totals fall by £3.131bn a year from 2025-26 to 2030-31, and household net income rises by the same amount. Poverty rates don't change. Measured by real runs, below.

The bug

• expected_sdlt returned stamp_duty_land_tax + property_sale_rate + corporate_sdlt (variables/gov/hmrc/expected_sdlt.py on main).

• GOV_TAX_VARIABLES, HOUSEHOLD_TAX_VARIABLES and PRE_BUDGET_CHANGE_HOUSEHOLD_TAX_VARIABLES list both expected_sdlt and corporate_sdlt, so corporate SDLT entered each total twice.

• gov.hmrc.stamp_duty.abolish zeroed expected_sdlt only. corporate_sdlt, listed separately, survived abolition in full.

• property_sale_rate is a share (0.045, unit /1), and adding it gave every household £0.045 of SDLT, including households in Scotland and Wales. Add dataset schema, remove State variables and add HBAI net income #1155 turned the original * into +.

  The datasets don't need the scaling anyway. policyengine-uk-data draws purchasers directly: property_purchased is a seeded draw at 3.85% in datasets/frs.py. So stamp_duty_land_tax is already the expected liability.

The fix

• expected_sdlt returns stamp_duty_land_tax: the household's own purchases and leases only.
• corporate_sdlt returns 0 when gov.hmrc.stamp_duty.abolish is set.
• No household is sdlt_liable once SDLT is abolished, so the stamp_duty_land_tax liability is zero too. Before, it kept showing the pre-abolition liability. expected_sdlt keeps its own check, so abolition also zeroes it when the liability is entered directly.
• The totals' lists are unchanged, so each part is counted once, and abolition zeroes both.
• The high value council tax surcharge tests drop five property_sale_rate: 0 overrides that existed only to cancel the stray £0.045.

Tax-head partition

New file: policyengine_uk/data/fiscal_partition.yaml (plain YAML, schema_version: 1).

Heads. Each head is one row of the OBR's current receipts table (EFO November 2025, Table A.5), split by payer or nation where the model splits it, and measured by exactly one variable. Its fields:

• obr_line, or not_an_obr_current_receipt with a reason (for example, student loan repayments are a financial transaction);
• obr_table_3_1, the line of the March 2026 EFO's Table 3.1 that holds it, where that table's footnotes say so;
• report_group and payer;
• the totals it belongs to;
• sign, and zero_at_baseline for the reform-only heads;
• reads, which declares the reads that are eligibility rather than inclusion.

SDLT is two heads on one receipts row ("Property transaction taxes"): sdlt_household (expected_sdlt) and sdlt_corporate (corporate_sdlt).

Coverage. The 29 heads sit on 12 of Table A.5's 30 tax rows. uncovered_receipts lists the other 18 (corporation tax, inheritance tax, air passenger duty and so on), so every row is either measured or declared uncovered.

Exclusions. 73 variables that must never be summed alongside the heads, each with its kind:

• component: inside a head, e.g. stamp_duty_land_tax, the NI classes, the income tax parts and reductions.
• measure_within: e.g. salary_sacrifice_pension_ni_*, already charged through employment_income.
• composite: e.g. corporate_tax_incidence, which is corporate_sdlt + business_rates.
• duplicate: baseline_business_rates.
• baseline_measure: baseline_vat.
• assessment: a tax amount worked out for a means test or a ceiling, e.g. uc_income_tax_on_earnings, ni_class_4_maximum.
• change_measure: a wage change already inside the incomes a head taxes, e.g. employer_ni_fixed_employer_cost_change.

Overlap guards.

• SDLT is split between household and company payers.
• SDLT, LTT and LBTT are geographically exclusive, and so are council tax and NI domestic rates.
• VAT and employer NI appear as a level in gov_tax and as a change in household_tax, never in the same total.
• Council tax is counted gross, with Council Tax Reduction counted as spending. The OBR's row is net of reduction schemes, so comparisons need the reduction netted off.
• income_tax includes the High Income Child Benefit Charge, so benefits must be counted gross of it.
• Business rates, corporate SDLT and three reform-only heads (corporate tax change, carbon tax, land value tax) share the shareholding key; each is a different receipt.
• Private school VAT may overlap VAT on imputed consumption.

Independence. The heads come from the OBR table and each variable's formula, not from the model's lists; a differential test compares the two. A second specification was written blind by a GPT-6.1 Sol lane that was not allowed to read the model's lists or this file (~/reviews/uk-corporate-sdlt-2159/spec-out.md). Reconciling the two:

Blind specification	This partition
Same variable for income tax, both NI heads, VAT, private school VAT, fuel, alcohol and tobacco duties, VED, the licence fee, business rates, CGT, NI domestic rates, the surcharge, both SDLT heads and the four hypothetical taxes	Agreed
Council tax: the net variable council_tax_less_benefit, matching the OBR's net receipts	Kept gross council_tax, which is what the totals sum, with the reduction in spending. Recorded on the head and as a guard
LTT and LBTT: the unscaled liabilities, because expected_ltt and expected_lbtt discount twice	Kept the variables the totals sum, with #2246 noted on both heads. #2315 fixes it
Student loan repayments: outside both tax totals	Kept as the model counts them, flagged not_an_obr_current_receipt for separate reporting
VAT change, employer NI response incidence, the two unallocated tax changes: exclusions from a total of levels	Heads of the totals that use them, zero_at_baseline
18 receipt rows with no household measure	Adopted as uncovered_receipts
About 20 further exclusions (income tax reductions, Class 4 ceiling, Universal Credit tax assessments, wage-change measures) and three guards	Adopted
stamp_duty_land_tax still shows a liability after abolition	Fixed here: abolition now makes no household sdlt_liable

I checked the two OBR claims I adopted at the source: Table A.5's rows and footnotes, and the OBR's council tax page (receipts include Northern Irish domestic rates and are net of reduction schemes).

Shared with W3. W3 (per-PR national impact reports, #2335) reads this file directly for its tax heads and partition identity, and adds spending_heads to it. The schema was agreed with that workstream. Both landing orders were checked: on the combined state this PR's tests and #2335's 17 partition and guard tests pass.

Invariants

1. Partition, differential. For each total, the variables the model sums are exactly the partition's heads for it, each listed once. No exclusion is listed. Every tax row of Table A.5 is a head's row or declared uncovered, never both.
2. No overlap, traced. Within a total, no head's formula reaches another head of that total unless the partition declares the read. The declared reads are TV licence → over-75 discount → Pension Credit income → income tax and NI, and private school VAT → attendance imputation. Every declared read must still occur.
3. Coefficient one. Raising any head by 1,024, with every other input unchanged, raises each total that includes it by exactly 1,024 per member, and moves no other head of those totals.
4. Identity. On random households, in 2025-2030, at current law and with SDLT or council tax abolished, each total equals the sum of its heads, less those dropped_when names. Every sign: nonnegative head is ≥ 0. Every zero_at_baseline head is 0 at current law. At most one of SDLT, LTT and LBTT is non-zero per household.
5. SDLT once. expected_sdlt equals stamp_duty_land_tax. Removing corporate SDLT revenue lowers each total by exactly corporate_sdlt. Abolition lowers each total by exactly stamp_duty_land_tax + corporate_sdlt, and leaves stamp_duty_land_tax, expected_sdlt and corporate_sdlt all zero. With a dataset, weighted corporate_sdlt is the corporate revenue statistic.
6. Abolition removes SDLT and nothing else. A Welsh shareholder keeps business rates and loses corporate SDLT.

Tests

File	Covers
tests/test_fiscal_partition.py	1–4 (schema, differential, traced guard, 29 perturbation cases, Hypothesis identity)
tests/test_sdlt_counted_once_properties.py	5 (Hypothesis, two households, 2024–2027)
tests/policy/baseline/gov/hmrc/stamp_duty/corporate_sdlt.yaml, expected_sdlt.yaml, sdlt_liable.yaml	Fixed cases for 5 and 6

Red on main, green here. I ran the final test_fiscal_partition.py (static, traced and SDLT perturbation cases) against main's model code (3f174d0):

• the traced guard fails for all three totals, because expected_sdlt reaches corporate_sdlt;
• raising corporate_sdlt by 1,024 moves gov_tax by 2,048.

On this branch (3ada49f) all 39 partition tests and the SDLT property test pass locally, and so do the 24 SDLT YAML cases. CI runs the full suite.

National impact (real runs)

Setup.

• Data: policyengine-uk-data 1.58.0, enhanced_frs_2024_25.h5, sha256 5ba399ea5ab9a6c7d4e9aaf6c623080beb34d9ad5f330f6244dbad72b8d960b6, the certified release.
• Code: base is main at 6d89338 and head is this branch's model code. Both are policyengine-uk 2.126.8.
• Runtime: both use policyengine-core 3.33.0, the certified bundle's core. The repo lock pins 3.32.27.
• Each figure is a weighted total from Microsimulation runs: no scaling and no interpolation. 36 runs in all: 2 versions × 3 scenarios × 6 years.

Current law, £bn a year

	2025-26	2026-27	2027-28	2028-29	2029-30	2030-31
gov_tax, base	978.967	1,021.444	1,060.540	1,096.037	1,133.645	1,172.964
gov_tax, head	975.836	1,018.312	1,057.409	1,092.905	1,130.514	1,169.833
Change in gov_tax, household_tax, pre_budget_change_household_tax and gov_balance	−3.131	−3.131	−3.131	−3.131	−3.131	−3.131
Change in household_net_income	+3.131	+3.131	+3.131	+3.131	+3.131	+3.131
expected_sdlt, base → head	19.448 → 16.316	20.440 → 17.308	21.414 → 18.283	22.393 → 19.262	23.395 → 20.264	24.493 → 21.362

The −£3.131bn is £3.130bn of corporate SDLT plus about £1.4m of stray £0.045s across roughly 31m households. corporate_sdlt (£3.130bn) and stamp_duty_land_tax are unchanged.

Accounting checks on the data

Check (every year 2025-26 to 2030-31)	Base	Head
gov_tax change from zeroing corporate SDLT revenue, ÷ corporate_sdlt	2.0000	1.0000
Corporate SDLT left after abolition	£3.130bn	£0
Households with ≥2 of SDLT, LTT, LBTT non-zero	86–96	0
Non-finite outputs	none	none

SDLT abolition costing. The measured effect barely moves. Abolishing SDLT changes gov_tax by −£20.440bn on base and −£20.438bn on head in 2026-27, and likewise in every year. Base's baseline held corporate SDLT twice and its abolished scenario once, so the difference was one copy either way. What changes is the level of both scenarios.

Distribution. Head − base, in household net income per household, weighted mean £, 2026-27:

D1	D2	D3	D4	D5	D6	D7	D8	D9	D10	All
45.94	60.27	53.38	86.63	81.15	82.47	93.36	114.09	164.62	223.17	99.84

Corporate SDLT is allocated by shareholding, so the removed copy is worth more to higher deciles. The other years are in the evidence folder.

Poverty. BHC and AHC, absolute and relative: unchanged to four decimal places in every year. HBAI income does not read these variables; hbai_excluded_income lists corporate_tax_incidence, but no current HBAI income formula reads it.

Things this does not change

Follow-ups.

• LTT and LBTT are discounted twice (expected_sdlt adds the property sale rate; expected_lbtt and expected_ltt discount purchases twice #2246). On 1.58.0 in 2026-27, LBTT is £2.31bn but expected_lbtt is £0.104bn. Land transaction tax is £0.070bn but expected_ltt is £0.003bn. The partition records this as a data_note on both heads. The fix should follow this PR's convention, the purchase draw being the expectation.
• Corporate SDLT is held at 2020 revenue. £0.705bn residential plus £2.425bn non-residential, from gov.hmrc.stamp_duty.statistics, never uprated. Update corporate SDLT and business rates statistics and project them with OBR receipts #2322 updates it.
• Alcohol, tobacco and vehicle excise duties are £0 for every household on 1.58.0 in every year. Recorded as data_notes; W3's report flags inert heads.
• Private school VAT is zero at current law. gov.contrib.labour.private_school_vat is 0, so the head is reform-only. A guard records the possible overlap with VAT on imputed consumption. Charge VAT on private school fees at current law #2319 addresses it.
• 76 households have negative VAT on 1.58.0 (−£72m weighted in 2026-27), from negative imputed consumption or full-rate share. Recorded on the vat head; policyengine-uk-data#565.
• uk-data test. policyengine-uk-data's tests/test_property_purchased.py compares weighted expected_sdlt with total SDLT receipts, using wide bounds (£5–50bn). expected_sdlt is now household-only (£17.3bn in 2026-27), still inside the bounds. A tighter benchmark should add corporate_sdlt. That change goes in the batched data release.

Merge

Lands on the hub's gates. The earlier "hold until 2 November" was set before cos d1152, in which Max ruled no model freeze for the 28 October Budget. Evidence and scripts: ~/reviews/uk-corporate-sdlt-2159/ (impact.py, compare.py, impact.md, the runs).

axiom: n/a: this corrects how the model's tax totals add up its own variables (corporate SDLT incidence, a modelling convention), not a statutory rule rulespec-uk encodes

🤖 Generated with Claude Code

Reproducer: policyengine_uk/tests/policy/baseline/gov/hmrc/council_tax/high_value_council_tax_surcharge.yaml, policyengine_uk/tests/policy/baseline/gov/hmrc/stamp_duty/corporate_sdlt.yaml, policyengine_uk/tests/policy/baseline/gov/hmrc/stamp_duty/expected_sdlt.yaml, policyengine_uk/tests/policy/baseline/gov/hmrc/stamp_duty/sdlt_liable.yaml::No household is liable once Stamp Duty is abolished, policyengine_uk/tests/test_fiscal_partition.py::test_partition_schema, policyengine_uk/tests/test_fiscal_partition.py::test_partition_heads_are_the_variables_each_total_sums

[MaxGhenis]

Local targeted runs at 685c1e7 (policyengine-uk installed editable, Python 3.13), through the shared heavy-test lock. CI runs the full suite.

• policyengine-core test on tests/policy/baseline/gov/hmrc/stamp_duty/corporate_sdlt.yaml, expected_sdlt.yaml and tests/policy/baseline/gov/hmrc/council_tax/high_value_council_tax_surcharge.yaml: 22 passed.
• pytest policyengine_uk/tests/test_sdlt_counted_once_properties.py: 1 passed (15 Hypothesis examples, derandomised) in 97 s.

An independent review is running; findings will follow here.

[MaxGhenis]

Independent review (GPT-6.1 Sol, 9 Oct) asked for changes; all addressed at 2d606f7ca:

• Effect section was wrong (major). Abolition does not save £3.13bn more: the old abolished scenario still carried corporate SDLT, so abolition saved S + C + r before and S + C now. The body now gives the per-household algebra and labels it as deduced from the formulas, not measured on production data.
• Docstring overstated the data (minor). expected_sdlt now says the FRS-based datasets impute an annual purchaser indicator with a seeded draw (policyengine-uk-data datasets/frs.py:1874-1879).
• uk-data consumer (minor). tests/test_property_purchased.py:74 compares weighted expected_sdlt with total receipts; noted as a follow-up for the next batched data release.
• Added Fixes #2159 and the axiom: line; the HVCTS note now says five overrides were removed.

Mutation check, run locally: with expected_sdlt adding corporate_sdlt back, corporate_sdlt.yaml failed 3 of 5 cases and the property test failed at its first assertion (test_sdlt_counted_once_properties.py:144). Restored; the working tree is clean.

[MaxGhenis]

Stacked on this PR: #2319 (fixes #2318) makes VAT on private school fees current law from 1 January 2025 (Finance Act 2025 ss.47 and 49), and updates this file's private school entries to match:

• private_school_vat is no longer zero at baseline and moves to gov_tax only;
• a new private_school_vat_change head in household_tax;
• the exclusion baseline_private_school_vat;
• the vat guard becomes overlapping_base;
• the identity property draws pupils.

#2319 is rebased on 5f08c244c. Once this merges, #2319 retargets to main.

[vahid-ahmadi]

Thanks Max. Reviewed at 3ada49f. The fix is right and the partition is a really useful guard. Nothing blocking from me; a few small points below.

What I checked

• The bug is real on main. expected_sdlt adds corporate_sdlt (expected_sdlt.py:15-18), and gov_tax.py, household_tax.py and pre_budget_change_household_tax.py all list both, so corporate SDLT goes in twice and abolition leaves one copy behind.
• The fix. expected_sdlt is now just stamp_duty_land_tax, corporate_sdlt is zero under gov.hmrc.stamp_duty.abolish, and sdlt_liable turns off with it. The only readers of sdlt_liable are stamp_duty_land_tax (defined_for) and the tests, so making it abolition-aware doesn't reach anything else. corporate_tax_incidence and hbai_excluded_income now lose corporate SDLT under abolition too, which is what you'd want.
• Tests. I used a fresh uv sync environment in a worktree (policyengine-core 3.32.27), not the shared venv. At the head, the two new Python files pass (40), and so do the stamp duty and council tax YAML directories (36). I copied the new tests onto main (0bc08f2), and they fail there: 6 of the 40 Python tests (the traced guard for all three totals, sdlt_corporate entering once, the sum identity, and the SDLT property test) and 13 of 24 YAML cases.
• National effect. I re-ran it on the certified Microcosm UK 2024-25 release (sha256 aa31bdf6…) rather than the Enhanced FRS, comparing the PR's merge base (3f174d07) with the head in the same environment:

	2025-26	2029-30
gov_tax, base → head	911.695 → 908.564	1,060.817 → 1,057.686
Change in gov_tax, household_tax, pre_budget_change_household_tax	−3.131	−3.131
Change in household_net_income	+3.132	+3.131
Saving from abolishing SDLT, base → head	18.286 → 18.285	21.873 → 21.872

So the −£3.131bn reproduces on a second dataset. Abolition now removes exactly stamp_duty_land_tax + corporate_sdlt (£15.155bn + £3.130bn in 2025-26), where base left the £3.130bn corporate copy in.

• Other users. policyengine.py, the API and the 2026 Budget dashboard don't read expected_sdlt or corporate_sdlt directly; they get the corrected totals. policyengine-uk-data's test_sdlt_total_reasonable reads expected_sdlt against a £5–50bn band, which the new value still clears.

Findings

1. Scorecard evidence goes stale (should, other repo). policyengine-scorecard's docs/uk_replay/evidence/classification_sources.json keeps a copy of the old expected_sdlt.py source, and its budget_2018:tax:016 row lists expected_sdlt as a mechanism variable. Neither breaks, but the snapshot will no longer match, so it's worth refreshing when this lands.
2. SDLT and LTT/LBTT are now on different bases until Stop discounting LTT and LBTT twice #2315 (should). expected_sdlt is the unscaled liability, while expected_ltt.py:13-15 and expected_lbtt.py:13-15 still multiply by property_sale_rate. That was already inconsistent on main, and the partition records it under expected_sdlt adds the property sale rate; expected_lbtt and expected_ltt discount purchases twice #2246, so it's fine here. But it means Welsh and Scottish property transaction taxes are about 4.5% of what they should be in the same totals, so I'd merge Stop discounting LTT and LBTT twice #2315 straight after this.
3. property_sale_rate has only those two readers left (nit). Once Stop discounting LTT and LBTT twice #2315 lands it may be unused. Its description ("This percentage of properties are sold every year") and the TODO in stamp_duty/property_sale_rate.yaml could say it's only the LTT/LBTT scaling, or the parameter could go.
4. The level is still high (outside this PR). On certified Microcosm 2025-26, household SDLT is £15.16bn plus £3.13bn corporate, £18.3bn in all. policyengine-uk-data's own test cites £13.9bn of official SDLT for 2024-25. The double count was part of the gap and this PR removes it. The rest looks like the purchase rate or the price distribution in the data, which deserves its own issue.

Happy to approve once you've seen these. None of them needs a change in this PR.

🤖 Generated with Claude Code