Deduct only each person's own tax and NI on earnings from UC earned income

[MaxGhenis]

Summary

uc_earned_income subtracted benunit_tax from the benefit unit's gross earnings. That is income tax and NI summed over every member, on all their income. So a claimant's tax on their own dividends, property or pensions came off their earnings, and any tax beyond a member's own earnings came off their partner's. The taper then bit less and UC was overstated.

UC Regs 2013 reg 55(5)(b) and reg 57(2) step 3 deduct only tax and NI paid by the person "in respect of" their employment or trade. This PR deducts, per person:

• their own income tax on their earnings;
• their Class 1 employee, Class 2 and Class 4 NI;
• their own pension contributions.

The work allowance then comes off the combined earnings, and the taper applies.

The #1942 counterexample: a couple in 2026, a pensioner aged 70 with State Pension £16,000 and a partner aged 45 earning £13,000, council rent £20,000. UC goes from £5,297.16 to £4,919.86, the legal figure.

On the Enhanced FRS the change cuts 2026 UC by £0.549bn against current main, from 292k benefit units (256 records), with no gainers. 90% of that is tax on dividends, and 43 records with over £50,000 of dividends carry £0.440bn of it. About 42% (£0.231bn) is one member's tax no longer coming off their partner's earnings. Those units get UC at all only because of a separate capital-modelling problem (#1948).

#1943 (State Pension as UC unearned income), which this was stacked on, has merged. Fixes #1942.

Law (legislation.gov.uk, read 2026-09-30 from /data.xml)

UC Regs 2013 reg 55(5) (https://www.legislation.gov.uk/uksi/2013/376/regulation/55):

In calculating the amount of a person's employed earnings in respect of an assessment period, there are to be deducted from the amount of general earnings or benefits specified in paragraphs (2) to (4)— (a) any relievable pension contributions made by the person in that period; (b) any amounts paid by the person in that period in respect of the employment by way of income tax or primary Class 1 contributions under section 6(1) of the Contributions and Benefits Act; ...

UC Regs 2013 reg 57(2) step 3 (https://www.legislation.gov.uk/uksi/2013/376/regulation/57):

Deduct from the amount resulting from step 1 or (if applicable) step 2 any payment made by the person to HMRC in the assessment period by way of national insurance contributions or income tax in respect of any trade, profession or vocation carried on by the person. If the amount resulting from steps 1 to 3 is nil or a negative amount, the amount of the person's self-employed earnings in respect of the assessment period is nil ...

Reg 22(1)(b) then tapers "their combined earned income" above the work allowance. So:

• Per person. Each person's deductions come off their own earnings only. A partner with no earnings has nothing for their tax to come off.
• Only in respect of the employment or trade.
  • Deductible: employee Class 1 (reg 55(5)(b) names "primary Class 1"), and Class 2 and Class 4 on the trade.
  • Not deductible: voluntary Class 3; the High Income Child Benefit Charge; the pension annual allowance charge.
  • ADM H5004: "All retirement pension income should be taken fully into account for UC; there are no comparable deductions for Income Tax and NI in UC."

Which part of a person's income tax is "in respect of" their earnings

When one person has both earnings and other taxable income, the law does not say how to split their tax.

• Neither the regulations nor the ADM give an apportionment rule, and no tribunal decision on it was found.
• In practice DWP takes employees' figures from RTI (reg 61(2)). ADM H3170 paraphrases the deduction as income tax and Class 1 "deducted or paid in that assessment period", without "in respect of the employment".
• HMRC usually collects tax on a working pensioner's State Pension through their employer's PAYE ("Your employer will usually take any tax you owe off your earnings, including any tax you owe on your pension", GOV.UK). So the RTI figure can include it: it behaves like earnings as the top slice, or more under a K code. LITRG (a secondary source) says DWP deducts any tax paid in practice.
• Reg 61(3)(b) lets DWP depart from RTI where it "fails to reflect the definition of employed earnings in regulation 55", but no ADM paragraph tells decision makers to strip out tax on other income.
• ADM H5004 treats pension income as gross ("there are no comparable deductions for Income Tax and NI in UC"). So any tax a rule attributes to pensions is deducted nowhere.

This PR follows the statutory words. It takes earnings as the lowest slice of the person's non-savings income, after the allowances they actually have (uc_income_tax_on_earnings):

• Savings and dividends sit above earnings and other non-savings income. ITA 2007 s.16 makes them "the highest part" of income, apart from the items s.1012 ranks above them, such as some termination payments. It does so for income tax purposes; UC is not one, so for UC this is an analogy.
• Property income sits above earnings. This is how the engine already stacks it. From 2027-28, Finance Act 2026 s.6 adds ITA s.16A, which places property income "immediately before the savings income", and s.25(3A), which sets allowances against other income first.
• Other non-savings income sits above earnings. That covers pensions, State Pension and taxable benefits. ITA 2007 fixes no order among these.
• The legacy notional-tax rules also tax earnings alone. HB Regs 2006 reg 39, IS reg 39 and JSA reg 102 take earnings less the claimant's personal allowance, at the basic rate. UC keeps that approach only for wholly notional earnings (ADM H3271-H3272).

The rule makes one invariant exact: unearned income never changes a person's deductions unless it changes their allowances (the personal allowance taper above £100,000, or Marriage Allowance). Tax reductions such as the married couple's allowance come off the tax on earnings first, as allowances do, so they keep the invariant. The property tests check this, including for working pensioners with a married couple's allowance. Tax refunds are an exception in law: reg 55(4A) counts a whole repayment as employed earnings, even tax on other income. The model does not model refunds.

For working pensioners whose State Pension is coded against their job, this deducts less than DWP probably does in practice.

The alternative is earnings as the top slice of non-savings, non-property income (rule C), which is closer to RTI when HMRC codes out a State Pension. In this model rule C never deducts less tax than this rule: its earnings slice is at least as wide and sits higher, and the marginal rates never fall. The runs agree, with no unit getting less UC under rule C. As a sensitivity, a real run of rule C cuts 2026 UC by £0.530bn instead of £0.549bn. The issue's own sketch, "tax on all income minus tax without earnings", would not fix a single claimant whose other income is within the personal allowance. For example, with £20,000 of earnings and £5,000 of property income it deducts the whole £2,286, the same as the bug.

Change

• uc_income_tax_on_earnings (new, Person):
  • taxes the bottom slice of earned_taxable_income that earnings occupy, reading the incomes outside that base from gov.hmrc.income_tax.earned_taxable_income_exclusions, so a reform that changes the list is followed;
  • uses the rUK or Scottish schedule, as earned_income_tax does;
  • takes tax reductions (gov.hmrc.income_tax.income_tax_subtractions) off that tax first, floored at nil, so it never exceeds the person's tax and never includes the High Income Child Benefit Charge or the pension annual allowance charge.
• uc_national_insurance_on_earnings (new, Person): ni_class_1_employee + ni_class_2 + ni_class_4.
• uc_individual_earned_income (new, Person): uc_mif_capped_earned_income less the person's own relievable pension contributions (none after 75: Finance Act 2004 s.188(3)(a)), tax and NI, floored at nil.
• uc_earned_income: the sum of uc_individual_earned_income over the unit, less uc_work_allowance, floored at nil.
• The UC docs page describes the deductions.

benunit_tax is no longer used by UC. It stays as a variable.

Household results (hand-computed, in uc_earnings_deductions.yaml)

Case (2026)	Deducted tax	UC before	UC after
Mixed-age couple from #1942	partner's £86 + £34.40 NI, not the pensioner's £686	£5,297.16	£4,919.86
Single parent, earnings £20,000 + property £5,000, council rent £9,600	£1,486 on earnings, not £800 on property	£11,748.50	£11,308.50
Self-employed £18,000 (start-up), pension £4,000, dividends £3,000, council rent £12,000	£1,086 + £325.80 Class 4, not £800 + £268.75	£4,563.10	£3,975.29

The YAML also covers:

• a Scottish taxpayer, where earnings sit below pension income in the Scottish bands;
• a Marriage Allowance recipient, whose £252 reduction lowers their tax on earnings;
• Class 3 NI, not deducted;
• a non-earner's pension contributions, which no longer come off the partner;
• the High Income Child Benefit Charge, not deducted;
• 2027 property rates of 22%, not deducted;
• a married couple's allowance, which comes off the tax on earnings first, including with pension income and with the pension annual allowance charge;
• pension contributions after 75, not deducted;
• a reform exempting pensions from income tax.

The "before" figures are from the pre-fix formula, run as a scenario on this branch. Both rows reflect main's #1950 (property income and dividends are not UC unearned income) and #1881 (the trading allowance is not given without gross receipts).

Enhanced FRS impact (real runs)

Setup:

• Dataset: a copy of enhanced_frs_2024_25.h5 (sha256 e433e532…), one private copy per run.
• One Microsimulation per package state, all from clean trees:
  • main bfc5fea, which includes Count State Pension as Universal Credit unearned income #1943, Replace generic child and adult flags with each programme's legal definitions #1896, Stop counting interest, dividends and rent as Universal Credit unearned income #1950 and Stop stacking the trading allowance on expenses already netted from profit #1881;
  • this PR at 295ca1e (the commits since then change only tests and merge main 7668506);
  • rule C on this PR (scratch 0379cf3);
  • the old deductions applied per person (scratch 5a7bfaa), for the split below.
• Scripts (dataset_impact.py, compare_impact.py, decompose.py, poverty_flips.py, capital_high_dividend.py) and outputs are kept outside the repo.
• Earlier runs on older bases gave −£0.505bn (main 44240bd), −£0.485bn (main 7b9fc37) and −£0.487bn (main 3c48247). Main has since gained Stop counting interest, dividends and rent as Universal Credit unearned income #1950 (interest, dividends and rent are not UC unearned income), Stop stacking the trading allowance on expenses already netted from profit #1881, Compute the State Pension triple lock from its statutory inputs #1939, Apply the benefit cap to mixed-age couples on Universal Credit, and keep them off pensioner Council Tax Reduction #1944 and others. Over the same span the UC units on main with over £50,000 of dividends went from 30.9k getting £0.519bn of UC to 48.9k getting £0.700bn, and £0.058bn of the £0.062bn rise in the loss is in units with over £10,000 of dividends.

This PR against main:

Year	UC on main	UC change	Losers	Records	Lose all UC	Mean loss	Median loss (weighted)	Rule C instead
2025	£74.82bn	−£0.503bn	287k	225	18.7k	£1,755	£139	−£0.486bn
2026	£78.91bn	−£0.549bn	292k	256	25.3k	£1,884	£216	−£0.530bn
2027	£80.72bn	−£0.576bn	293k	259	29.7k	£1,962	£221	−£0.555bn
2028	£82.28bn	−£0.602bn	298k	263	29.5k	£2,017	£241	−£0.580bn
2029	£83.41bn	−£0.603bn	297k	277	29.5k	£2,029	£242	−£0.580bn
2030	£80.35bn	−£0.624bn	316k	279	38.3k	£1,977	£200	−£0.603bn

Checks across every year:

• There are no UC gainers.
• UC before and after the cap never rises.
• Earned income never falls, up to float32 rounding of about £0.02 on six-figure earnings.
• CTR, Pension Credit and Housing Benefit do not change.
• The benefit cap reduction falls by £8-19m a year.
• No unit gets less UC under rule C than under this PR.

Where the 2026 loss comes from. The fix does two things, and a real run separates them:

Step	UC change	Losing units	Records
1. Apply the old deductions per person, so no member's deductions exceed their own earnings	−£0.231bn	78k	50
2. Then deduct only the tax and NI on earnings	−£0.318bn	234k	225
Both	−£0.549bn	292k	256

Step 1 is almost all couples (£0.224bn of it, 46 records), and £0.222bn of it is in units with over £10,000 of dividends. The pattern is a member whose tax, mostly on dividends, is larger than their own earnings, with the rest netted off their partner's earnings. A member with no earnings at all, as in the #1942 couple, accounts for under £0.0001bn.

By the kind of tax no longer deducted (decompose.py: each losing unit's loss split pro rata over the deductions it no longer gets):

Tax no longer deducted from earnings	UC change
Tax on dividends	−£0.494bn
Tax on pensions, State Pension and taxable benefits	−£0.020bn
High Income Child Benefit Charge	−£0.019bn
Deductions the per-person floor stops (e.g. a non-earner's pension contributions)	−£0.009bn
Property income tax	−£0.007bn

• The aggregate rests on a few records.
  • 84 records with over £10,000 of dividends carry £0.506bn (81k units).
  • 43 records with over £50,000 carry £0.440bn.
  • The top 10 records carry 75% of the loss.
• The rest is small. Units with at most £10,000 of dividends lose £0.043bn between 211k units (172 records).
• The high-dividend units are the separate problem. On main, 48.9k UC units (59 records) with over £50,000 of dividends get £0.700bn of UC. Since Stop counting interest, dividends and rent as Universal Credit unearned income #1950, dividends are not UC unearned income; capital counts only through its assumed yield (reg 72), and these units' £6.03bn of dividends sits on £0.264bn of modelled capital, under the £16,000 limit. This PR cuts their UC to £0.260bn, because their dividend tax no longer wipes out their earnings. Whether they should be eligible at all is UC-eligible units with six-figure dividends: dividends treated as yield of a few thousand pounds of imputed capital #1948.
• The UC earned income deducts the whole benefit unit's tax, including tax on pensions and other unearned income #1942 couple itself barely registers in aggregate. Only 17k of the 1.23m mixed-age units with State Pension get UC, and their UC falls by £0.002bn.

Poverty. AHC child poverty rises 0.30pp in 2026-27 and 2027-28, 0.02pp in 2028-29 and 2030-31, and not at all in 2025-26 and 2029-30 (all people: at most 0.11pp). Each year it is at most one survey household crossing the AHC line. BHC child poverty rises 0.11pp from 2027-28 (all people: at most 0.03pp), one survey household, two from 2029-30. All of this is sampling noise, not a finding.

Invariants (tests)

policyengine_uk/tests/test_uc_earnings_deductions_properties.py is a Hypothesis test (derandomized, 10 examples each). It runs over singles, couples and mixed-age couples, including working pensioners with a married couple's allowance, with:

• employment, self-employment (with and without the minimum income floor), State Pension, private pension, property, savings and dividend income;
• children, tenure, rent;
• London, the North East, Wales and Scotland;
• the years 2020, 2026 and 2027.

1. Unearned income never changes earned income. Adding any kind of unearned income to any adult leaves every uc_individual_earned_income, and uc_earned_income, unchanged.
2. Monotone. UC before and after the benefit cap is non-increasing in each kind of unearned income for each adult.
3. Differential against the tax engine. uc_income_tax_on_earnings equals the income_tax the same person pays with their earnings as their only income; uc_national_insurance_on_earnings equals their NI. Neither exceeds the person's own tax or NI.
4. Differential against the old formula. Earned income is never lower, and UC never higher, than under the pre-fix formula, run as a scenario with Marriage Allowance claimed. The fix only removes deductions.

Invariants 1-3 keep adjusted net income under the £100,000 taper and switch Marriage Allowance off, because a change in allowances legitimately changes the tax on earnings. A married couple's allowance is covered: one family shape has a working partner aged 93-96 with the allowance, whose reduction comes off the tax on earnings first.

test_uc_state_pension_properties.py (from #1943):

• UC earned income deducts the whole benefit unit's tax, including tax on pensions and other unearned income #1942's strict xfail test_tax_on_state_pension_does_not_reduce_partners_earned_income now passes, and its marker is removed.
• Pound-for-pound and the property-income equivalence are widened to all families (with_earnings=False is gone), with Marriage Allowance switched off.
• New strict xfail test_state_pension_does_not_change_partners_marriage_allowance pins Marriage Allowance is booked on the recipient as the transferor's unused allowance, not the fixed s.55B transfer #1947.
  • The model books Marriage Allowance on the recipient as min(partner's unused allowance, £1,260).
  • ITA 2007 s.55B gives a fixed £252 reduction and cuts the transferor's own allowance.
  • So in the model the pensioner's State Pension raises the partner's tax on earnings. With £12,000 of State Pension the model gives UC £6,085.16 against £6,009.26 in law.

Main's own property files had pinned #1942 as strict xfails; this PR removes them and widens the invariants:

• test_uc_unearned_benefits_properties.py (Count contributory ESA, Maternity Allowance, industrial injuries benefit and capped Carer Support Payment as UC unearned income #1958): the award invariants now cover every family, with Marriage Allowance off.
• test_uc_income_from_capital_properties.py (Stop counting interest, dividends and rent as Universal Credit unearned income #1950): "interest, dividends and rent leave the award unchanged" now covers working families too, with Marriage Allowance off, comparing the benefit units whose personal allowances the income leaves unchanged. Above £100,000 of adjusted net income the allowance tapers (ITA 2007 s.35(2)), which raises the tax on earnings: the exception stated above.
  • The award after the benefit cap is compared only where the cap exemption is also unchanged. The model's earnings exception to the cap nets tax on all of a person's income off their earnings (Benefit cap, targeted childcare and local CTR use the wrong UC earned income figure #1986), so dividends can switch the cap on.
  • New strict xfail test_tax_on_dividends_does_not_remove_the_benefit_cap_earnings_exception pins that case: a single London claimant earning £11,000 with £10,000 of dividends loses £2,081.80 to the cap, although reg 82(1)(a) rests the exception on earned income. Use each rule's own Universal Credit earned income in the benefit cap, targeted childcare and local CTR #1999 fixes it.

Results at 8c88c7c (main 7668506 merged in): the UC YAML folder passes locally (240 tests) and CI runs the full suite. The four UC property files give 21 passed and 2 xfailed (the Marriage Allowance and benefit cap pins).

Not changed here (follow-ups)

• Marriage Allowance is booked on the recipient as the transferor's unused allowance, not the fixed s.55B transfer #1947: Marriage Allowance is booked on the wrong partner and scaled to the transferor's unused allowance.
• UC-eligible units with six-figure dividends: dividends treated as yield of a few thousand pounds of imputed capital #1948: UC-eligible units with six-figure dividends.
• Benefit cap, targeted childcare and local CTR use the wrong UC earned income figure #1986: the benefit-cap earnings exemption, the targeted childcare income test and four local CTR schemes each use a different UC earned income figure. All three predate this PR.
• UC gross earnings: trade losses offset employment, expenses and statutory pay mishandled, payroll giving missing #1997: gross UC earnings let a trade loss offset employment income, ignore employment expenses (reg 55(3)(a)) and statutory pay (reg 55(4)), and have no payroll giving deduction (reg 55(5)(c)). All four predate this PR.
• Dependants' earnings count in the unit's earnings today. Count only claimants' income in Universal Credit #1978 limits UC income to claimants.
• Minimum income floor. Reg 62(4) makes the floor a net figure after notional tax and NI. The model compares the gross floor with gross earnings and then deducts actual tax on actual earnings, as before this PR.
• Reg 55 has no express nil floor. The per-person floor here stops one person's deductions exceeding their own earnings. Reg 57 floors self-employed earnings at nil explicitly.

axiom: TheAxiomFoundation/rulespec-uk#387 queued (regs 52, 53, 55 and 57 are not in the pinned corpus, so no signed encode is possible yet; the EFRS oracle bridge feeds uc_earned_income into reg 22, so it cannot test this)

Merge

This merges on gates under Max's 2026-10-02 rule that PolicyEngine UK bug fixes merge on gates whatever their effect on aggregates: CI green on the head, an independent review approving that head, and the impact above from real runs. #1943 has merged, so stack order allows it.

🤖 Generated with Claude Code

[MaxGhenis]

Coordination note from #1950 (UC capital income). #1950 adds policyengine_uk/tests/test_uc_income_from_capital_properties.py. Its test_tax_on_dividends_does_not_raise_a_working_familys_award is a strict xfail(raises=AssertionError) pinned to #1942, which this PR fixes.

Whichever of #1949 and #1950 merges second should:

• remove that marker;
• drop with_earnings=False from test_interest_dividends_and_rent_leave_the_award_unchanged;
• re-run the combined Enhanced FRS impact, since the two PRs interact on dividend-holding units.

#1950's merge is waiting on a decision from Max (over the £1bn/yr line).

🤖 Generated with Claude Code

[MaxGhenis]

#1950 merged (770bc06), so #1949 lands second. On rebase:

• remove the strict xfail on test_tax_on_dividends_does_not_raise_a_working_familys_award in policyengine_uk/tests/test_uc_income_from_capital_properties.py;
• drop with_earnings=False from test_interest_dividends_and_rent_leave_the_award_unchanged;
• re-run the combined Enhanced FRS impact.

#2107 gives that file's generated adults explicit claimant roles, so rebase on it if it has merged.

🤖 Generated with Claude Code

[MaxGhenis]

UK hub merged main into this branch at 09ae2eac4. The owner session was cut off by a network drop.

• Parents: 8c88c7cd9 (approved by 20261003-144221-uc1949-delta2) and main 27dd5f7a (after Give sharers, boarders and lodgers their own rent, tenure and LHA category #2006).
• Conflicts: two test files, test_uc_state_pension_properties.py and test_uc_income_from_capital_properties.py. Both sides added is_claimant_or_partner to the generated adults; this branch's version is kept, including its would_claim_marriage_allowance guard.
• Own diff: identical outside those two files, with the same 12 paths (checked with git diff against each base).
• Tests: both property files pass locally, 10 passed and 2 xfailed.
• Next: CI runs on the new Linux workflow (about 25 min). The merge goes ahead on green under the merge-only delta check above.

[MaxGhenis]

Merged by the UK hub under the PE-UK merge-on-gates rule:

• CI: every check passed on the merged head (fast Linux workflow); gh pr checks exits 0.
• Independent review approved the pre-merge head; the main merge is merge-only (own diff identical: see the hub's earlier comment).
• Impact: unchanged from the description (no own-code change in the merge).