[MaxGhenis]

Fixes #2004. Also fixes the non-dependant half of #1934; its pensioner-gate half stays open.

Summary

Housing Benefit and the national Council Tax Reduction schemes now make non-dependant deductions the way the regulations set them. The law was read on legislation.gov.uk at the 7 April 2025 point in time. For the CTR scales, each year's version was also read.

Housing Benefit (SI 2006/213 reg 74; SI 2006/214 reg 55 is the same for these rules):

Rule	Before	Now
Banding, reg 74(2)	right=True, so £192.00 a week fell in the band below, and £0 paid nothing	Bands include their lower edge ("not less than"). An in-work non-dependant with no income pays the reg 74(1)(b) amount
Remunerative work, reg 74(1)-(2), reg 6(1) and (6)	Everyone banded on their own income	Only a non-dependant working 16+ hours, and not on IS, JSA(IB) or ESA(IR), is banded. Anyone else pays the (1)(b) amount (£20.40 a week in 2026-27)
Couples, reg 74(3)-(4)	Each member charged on their own income	One deduction for the couple, the higher, banded on the couple's joint income. Any other adult in the family (for example a young person on approved training) is a separate non-dependant on their own income
Non-dependant exemptions, reg 74(7), (8), (10)	None	Full-time students; State Pension Credit recipients; under-25s on Income Support, income-based JSA, or Universal Credit calculated on no earned income
Claimant exemption, reg 74(6)	None	No deduction for anyone if the claimant or partner is blind or receives AA, the DLA care component, the PIP daily living component or AFIP

Council Tax Reduction covers three schemes:

• England pensioner: SI 2012/2885 Sch 1 para 8.
• Wales: WSI 2013/3029 Sch 1 para 3 and Sch 6 para 5.
• Scotland: SSI 2012/319 reg 48; SSI 2012/303 reg 67, then SSI 2021/249 reg 90.

The changes:

• Scales. These used the Housing Benefit scale. Each scheme now has its own weekly scale for 2013-14 to 2026-27, with the amending instrument cited for every year.
  • 2026-27:
    • England: £5.20 to £15.95, on £279/£485/£605.
    • Wales: £6.70 to £20.05, on the same bands.
    • Scotland: £5.55 to £16.55, on £286/£497/£614.
    • The HB scale was £20.40 to £131.45.
  • Pensioner and working-age scales are identical in Wales and in Scotland in every year.
  • From 2027-28, the scales are uprated by gov.benefit_uprating_cpi, as Set DWP benefit rates to the announced amounts, not CPI projections #1925 did for the HB scale.
• Banding and couples. As for HB.
  • Wales's working-age scheme also deducts for both members of a couple with a Universal Credit award (Sch 6 para 5(3), "other than where there is an award of universal credit").
  • I read that award as the non-dependant couple's.
  • Sch 1 para 3(3), for pension-age applicants, has no such words. So only a working-age applicant's claim takes both.
• Exemptions, at any age (para 8(7)-(8) and equivalents):
  • Full-time students.
  • Non-dependants on IS, income-based JSA, income-related ESA or State Pension Credit.
    • Scotland's working-age scheme, from 1 April 2022, also exempts a non-dependant whose partner is on IS, income-based JSA or income-related ESA and who is not on Universal Credit.
    • The source: SSI 2021/249 reg 90(8)(a) exempts a "qualifying income-related benefit claimant". Reg 4(1) defines that as one "who is, or who has a partner who is," on one of those benefits.
    • The definition opens with "an applicant". Read that way, the limb could never apply to a non-dependant, so it is read for the non-dependant.
    • Before April 2022 (SSI 2012/303 reg 67(8)(a)), and in Scotland's pension-age scheme (SSI 2012/319 reg 48(8)(a)), only the person on the benefit is exempt.
  • An adult for whom someone else is entitled to child benefit (LGFA 1992 Sch 1 para 3, via para 8(8)(b) and equivalents).
    • Entitlement needs a claim (SSAA 1992 s.1(1)), so the family must claim child benefit (would_claim_child_benefit).
    • An election not to be paid keeps entitlement (s.13A), so the opt-out flag is not read.
  • Universal Credit calculated on no earned income, from the year each scheme added it:
    • England, 2015-16 (SI 2014/3312);
    • Wales, 2015-16 (WSI 2015/44);
    • Scotland, 2016-17 (SSI 2016/81).
• Applicant exemption (para 8(6) and equivalents). The national schemes now read the per-family council_tax_reduction_applicant_has_non_dep_exemption (Assess each Council Tax Reduction claim under its own family's scheme and exemption #2015). Before, only the English local schemes did.
• English working-age local schemes (Merton, Kingston, Newham, Westminster, Oxford) keep their own variables.
  • They share council_tax_reduction_individual_non_dep_deduction_eligible.
  • So the national exemptions are applied inside the national deduction, not there.
  • A branch-against-main differential (5 councils × 3 household shapes) left all 45 local-scheme value arrays unchanged.

New variables:

• housing_benefit_non_dep_deduction_exempt (Person)
• housing_benefit_non_dep_deductions_claimant_exempt (BenUnit)
• council_tax_reduction_non_dep_deduction_exempt (Person)

New parameters:

• HB remunerative_work_hours (16)
• HB income_related_benefit_age_limit (25)
• per CTR scheme, non_dep_deduction/{amount, remunerative_work_hours, exempt_universal_credit_without_earned_income}
• Scotland working_age_exempt_partner_of_income_related_benefit_claimant (from 2022-04-01)

HB age_threshold (18) now starts in 2006, when the 2006 Regulations were made, not 2019.

How this composes with main

Since this PR opened, main has changed who is charged and who claims (#2006, #2009, #2015, #2027, #2078, #2089). This PR keeps those structures and supplies the amounts and exemptions.

• Housing Benefit (housing_benefit_non_dep_deductions). For each family, the deduction is built in four steps:
  1. Each other family's amount: one per couple, the higher (reg 74(3)), and one for each other member.
  2. Only people who are non-dependants of the household head count (is_non_dependant_of_household_head, Give sharers, boarders and lodgers their own rent, tenure and LHA category #2006). Joint occupiers, boarders and lodgers are not non-dependants (reg 3(2)(d)-(e), 3(4)). A family that bears a share of the rent is a joint occupier, so it is left out even where the rent is nil. In that case is_non_dependant_of_household_head, which tests the family's own rent, counts it.
  3. The total is multiplied by the family's share_of_household_rent (reg 74(5), Give sharers, boarders and lodgers their own rent, tenure and LHA category #2006). So each family liable for the rent bears its share, and a boarder or lodger bears none.
  4. It is zero when that family's own claimant or partner is exempt (reg 74(6)).
• Council Tax Reduction (council_tax_reduction_non_dep_deductions). Each applicant family is assessed under its own scheme (council_tax_reduction_pensioner, Assess each Council Tax Reduction claim under its own family's scheme and exemption #2015).
  • Person level. council_tax_reduction_individual_non_dep_deduction and council_tax_reduction_non_dep_deduction_exempt report each person under the scheme of the household head's family. That family is the applicant under Give Council Tax Reduction to the household head's family, not the oldest adult's #2009.
  • Recompute for a jointly liable applicant on the other scheme. Its deductions are recomputed under its own scheme by the helpers in _national_non_dependants.py. Three rules depend on the scheme:
    • England deducts only from a pension-age applicant, since working-age applicants are in the local schemes.
    • In Wales, a pension-age applicant takes one deduction for a UC couple, and a working-age co-applicant takes one for each member.
    • Scotland's partner rule applies only to the working-age applicant.
  • Apportionment. The family's total is multiplied by council_tax_reduction_joint_liability_share (para 8(5)).
  • Exemption. The total is zero under the applicant family's own council_tax_reduction_applicant_has_non_dep_exemption.
• Benefit cap (Give Universal Credit and Housing Benefit their own benefit cap and LHA shared-accommodation rules #2089).
  • A non-dependant family's gross income includes its UC after the cap. That figure reads uc_benefit_cap, which replaced benefit_cap.
  • It cannot read uc_benefit_cap_reduction or universal_credit: both depend on Housing Benefit through benefit_cap_welfare_benefits. Reading either raises a CycleError (checked by the round-6 review).
  • So the helper applies the cap itself. It counts the claimant's or couple's capped benefits only (see "Gross income" below).
• Payee-only awards (Count only the claimant's and partner's legacy awards in means tests and passports #2027). Count only the claimant's and partner's legacy awards in means tests and passports #2027 added payee-only is_on_income_support, is_on_income_based_jsa and is_on_income_related_esa.
  • This PR still reads its own payee rule (is_award_payee) for the IS, JSA(IB) and ESA(IR) exemptions and HB reg 6(6).
  • A follow-up switches these to Count only the claimant's and partner's legacy awards in means tests and passports #2027's variables once this PR is on main.
  • Count only the claimant's and partner's legacy awards in means tests and passports #2027's analysis found that, in the Enhanced FRS, no member outside the couple reports these awards.

Inputs and choices

• Full-time student: current_education != NOT_IN_EDUCATION or in_HE, the helper the local schemes already use.

  • In the Enhanced FRS, current_education comes from FTED/EDUCFT ("in full-time education"), so it is the full-time flag.
  • Note the age-based default: with no input, an 18- or 19-year-old is in tertiary education, and so exempt.

• Claimant or partner: is_claimant_or_partner (Replace generic child and adult flags with each programme's legal definitions #1896).

  • This is the SSCBA s.137 couple a benefit unit is formed around: the claimant and at most one partner, supplied by the dataset or inferred from the benefit unit.
  • That matches the HB Regs 2006 (regs 2(1), 19) and the CTR regulations (reg 2(1)), where a family's other members are the children and young persons.
  • It is used for:
    • the claimant exemption;
    • joint income, which counts only the couple;
    • the one-deduction-per-couple rule, which covers only the couple (any other member is deducted separately);
    • benefit-receipt exemptions, which count only for the claimant or partner of the award;
    • the cap basket's personal benefits.

• Gross income:

  • A person's own income. Their taxable income (total_income, floored at zero), plus:
    • statutory sick, maternity and paternity pay, which are earnings (HB reg 35(1)(i)) that the model keeps outside total_income;
    • Maternity Allowance.
  • A claimant or partner of the non-dependant's family is banded on the couple's own income plus the family's UC, CTC, WTC, child benefit, Income Support, income-based JSA, income-related ESA and Pension Credit.
  • Anyone else in the family is banded on their own income.
  • UC is taken after the benefit cap.
    • The cap (uc_benefit_cap) is applied to the capped benefits other than Housing Benefit, which a non-dependant family does not receive.
    • A person's own contributory JSA and ESA, incapacity benefit, SDA and Maternity Allowance count only for the claimant or partner. UC reg 79(1) caps the benefits "to which a single person or couple is entitled".
    • The award is reduced by the excess less the childcare costs element, or not at all where that element is greater (UC reg 81(1)-(2)).
  • Disregarded or not added:
    • the disability benefits disregarded by reg 74(9) are untaxed and not added;
    • other untaxed income the regulations count, such as maintenance and student support, is not added;
    • the income-related awards are read from the benefit unit's total.

• On a benefit: the IS, JSA(IB), ESA(IR) and SPC exemptions, and HB reg 6(6), apply only to the person the award is payable to (HB reg 2(3)-(3A), "payable to him").

  • That is the member who reports the award, or the benefit unit's head where none does.
  • UC counts for both joint claimants (reg 2(3B), entitlement).

• HB reg 6(6): the person an IS, JSA(IB) or ESA(IR) award is payable to is treated as not in remunerative work, so pays the lowest amount whatever their hours.

• UC entitlement: universal_credit_pre_benefit_cap > 0. Reading universal_credit would be circular (see "Benefit cap" above).

• "Earned income" in the UC limb (UC regs 52-62):

  • Employment income plus statutory sick, maternity and paternity pay, less relievable pension contributions (reg 55(4)-(5)). Statutory adoption, shared parental, parental bereavement and neonatal care pay have no input.
  • Self-employed earnings, with a loss counting as nil (reg 57(2), steps 3-4). Any relievable pension contributions not already deducted from employed earnings come off. The minimum income floor applies where the UC model applies it (reg 62). It doesn't apply in the start-up period (reg 62(5)).
  • Other paid work.

  Testing the person's own earnings, rather than a joint award's, is an interpretation:

  • The limb reads "the award is calculated on the basis that the non-dependant does not have any earned income".
  • A joint award is calculated on combined earnings (UC reg 22(1)(b)).
  • I found no DWP guidance that settles it. Charge non-dependant deductions for non-dependants within the claimant's benefit unit #2017 keeps the couple reading for the five English local schemes.

Not modelled

• Deferral and protection.
  • Pension-age HB defers an increase in a deduction for 26 weeks (SI 2006/214 reg 59(10)-(13)).
  • The CTR applicant exemption is protected while AA, DLA or PIP is suspended in hospital (para 8(6)).
• HB:
  • a working full-time student's summer vacation (reg 74(7)(d));
  • under-25s on assessment-phase income-related ESA (there is no ESA group input).
• Wales: ESA(IR) recipients in the work-related activity group are not exempt, but the model treats them as exempt.
• CTR: persons disregarded for council tax discounts (LGFA 1992 Sch 1), other than students and the para 3 adults (someone else is entitled to child benefit for them).
  • is_apprentice has no £195 earnings cap, so it isn't used.
• All schemes: youth training allowances, a normal home elsewhere, long hospital stays, prison and armed forces operations.
• Gross income:
  • maintenance and student support are not added, though inputs exist (a follow-up);
  • other untaxed income is not added, apart from statutory pay, Maternity Allowance and the family benefits listed above;
  • statutory adoption, shared parental, parental bereavement and neonatal care pay have no inputs;
  • a self-employed non-dependant's gross income uses profit, with any loss floored at zero, because turnover has no input. DWP HB guidance A5.551 counts total income with no deduction for business expenses.
• Benefit cap basket. The after-cap UC used here leaves out bereavement allowance and the widow's benefits, which have no inputs.
  • Main's benefit_cap_welfare_benefits, the total UC and HB actually cap, also leaves out Maternity Allowance, as its documentation says.
  • This PR's helper counts Maternity Allowance, a welfare benefit under Welfare Reform Act 2012 s.96(10)(i).
  • Aligning main's cap is a follow-up.
• Joint-claim JSA couples. Both members are on income-based JSA (HB reg 2(3)(c)). The model has no joint-claim input, so only the payee is treated as on it.
• Scottish disability assistance (ADP, PADP, CDP, Scottish adult DLA) for the claimant exemption. It has no separate input.
• Before 2015, the model reads parameters at 1 January, not in fiscal years. That is unchanged.

Review

Each round was an independent Subfleet review with a workspace that could run tests. In rounds 1-5, every fix came with a test that fails on the reviewed head. The reviews and responses are in the PR thread.

• Round 1 (59c7082) asked for changes. It found five P1s:

  1. is_adult used as claimant or partner;
  2. a dependent child's income counted as the couple's;
  3. HB reg 6(6);
  4. self-employed losses cancelling wages in the UC test;
  5. UC, tax credits and child benefit missing from gross income.

  It also found two omissions to disclose.

• Round 2 (6257277) asked for changes:

  • UC earned income per regs 55 and 62;
  • benefit awards counted only for their payee;
  • UC after the benefit cap in gross income;
  • the LGFA Sch 1 para 3 disregard;
  • explicit young partners, fixed by adopting Replace generic child and adult flags with each programme's legal definitions #1896's is_claimant_or_partner.

• Round 3 (452d6bb) asked for changes:

  • statutory pay, Maternity Allowance and the family's income-related benefits in gross income;
  • statutory paternity pay as earned income;
  • pension contributions off self-employed earnings;
  • Scotland's working-age partner rule from April 2022;
  • the cap's childcare protection;
  • the child benefit claim for LGFA para 3.

• Round 4 (7a5cbe2, GPT-6.1 Sol) asked for changes:

  • Maternity Allowance in the cap basket;
  • a self-employment or property loss no longer cancelling other gross income.

• Round 5, a delta review, approved d872f57. That was the last head before main was merged in.

• Round 6 reviewed the composition with main at 0b6c3db.

  • It found no P1. It confirmed:
    • every composition point, from hand-worked probes in England, Wales and Scotland;
    • no dependency cycle, with both alternatives that would cycle tested;
    • no change to any English local scheme;
    • a 96/96 randomised differential for jointly liable households (45 with mixed schemes);
    • that main's changed tests are justified (listed under Tests);
    • every earlier round's counterexample.
  • Its findings, and how each was addressed:
    • P2: main's sharers property test assumed one deduction per person and no claimant exemption. Fixed in 8e60349 (see Tests).
    • P2: this body described the head before the composition. It is rewritten.
    • P3: the mixed-scheme recompute had no test. council_tax_reduction_non_dep_deductions_mixed_schemes.yaml adds the review's England, Wales and Scotland cases.
    • P3: the cap basket counted every member's own benefits, not just the couple's. 2574473 masks them with is_claimant_or_partner. A YAML case gives a 17-year-old in the family her own Maternity Allowance; it fails on 8e60349 (HB £5,470.40, CTR £551.20 instead of £6,232.20 and £691.60).
    • P3: 2025-26 mid-band HB coverage was thin. There are now second- and middle-band cases.
  • CI's derandomized Hypothesis run of the rewritten sharers test then found one more case, on 89ba1a1. At nil rent, a sharer was both a non-dependant and a bearer of a share, so the household's HB deductions (£510.90) fell short of the pool borne in full (£1,021.80). 4a0935b keeps any family that bears a share out of the pool (reg 3(2)(d)). The replayed CI example and a new YAML case both fail on 89ba1a1 and pass now.

• Round 7, a delta review of this head, gates the merge.

Invariants

These hold for every input:

1. An exempt non-dependant's deduction is zero, for HB and for CTR.
2. An eligible, non-exempt non-dependant pays between the scale's lowest and highest amounts. The national CTR deduction is zero in English working-age households.
3. A non-dependant working under 16 hours pays the lowest amount, whatever their income.
4. A claimant or applicant exemption zeroes the family's deductions.
5. A family's deductions are computed in three steps:
   • sum, over the other families in the household, each family's higher member's amount (both members' for a Welsh working-age UC couple);
   • recompute that sum independently;
   • multiply it by the family's share of the rent (HB) or of the council tax liability (CTR).
6. Monotonicity: raising a non-dependant's income never lowers any deduction.
7. Exhaustively, in every year 2015-2030 (2019-2030 for HB), uprated years included: an income at a band's lower edge pays that band's amount, and the next lower representable income pays the band below's.
8. Differential: every CTR parameter value equals the enacted sums in tests/fixtures/council_tax_reduction_non_dependant_deductions.yaml.
9. Differential: the exemption flags equal a reading of the provisions computed straight from the inputs.
10. HB deductions over the families liable for a household's rent equal the pooled amount (one per couple, the higher, with sharers never in the pool) times the share borne by families without an exempt claimant. They never exceed the sum of individual deductions. This is main's sharers property test, restated for reg 74(3) and (6).

Tests

• YAML (all with hand-worked expectations from the law):
  • HB:
    • band edges in and out of remunerative work (2026-27), plus 2025-26 cases in four of the six bands;
    • couples, and a dependent child's income;
    • each exemption and its age boundary;
    • each claimant exemption, and that a child's DLA doesn't count;
    • aggregation.
  • The Set DWP benefit rates to the announced amounts, not CPI projections #1925 review's counterexample, end to end. A 40-year-old claimant pays £6,000 rent and lives with a zero-income 20-year-old full-time student. HB is £6,000; Set DWP benefit rates to the announced amounts, not CPI projections #1925 alone would have given £4,978.20.
  • CTR:
    • each scheme's band edges (2026-27, plus 2025-26 spot checks), and not in remunerative work;
    • every exemption at age 40, and the UC-limb start dates;
    • couples, including the Welsh UC rule;
    • the applicant exemption, and English working-age households;
    • jointly liable applicants under different schemes;
    • National council tax reduction: non-dependant deductions use Housing Benefit amounts and skip exemptions; pensioner gate ignores benefit exclusions #1934's example end to end: a guarantee credit pensioner whose son earns £30,000 working 37.5 hours. The deduction is £691.60 a year, not the HB scale's £6,232.20, and CTR is £1,308.40.
  • Each review round's counterexamples are in non_dependant_deduction_review_cases.yaml.
• Main's tests that change. Each change follows from the law; none hides a regression.
  • household_benefits_individual_non_dep_deduction.yaml.
    • Main's 2025-26 cases had no hours input. Under reg 74(2) they now pay the lowest amount, so they are rewritten as cases in remunerative work.
    • Main's "Other investment income does not move a worker's band" pinned a stopgap until this PR. A worker is banded on all gross income: £739.02 a week, the top band, £126.65 × 52.
  • housing_benefit_non_dep_deductions.yaml.
    • Main's Give sharers, boarders and lodgers their own rent, tenure and LHA category #2006 cases are unchanged.
    • The investment-income non-worker now pays reg 74(1)(b)'s £19.65 × 52 = £1,021.80. Main's 0 came from right=True, and its comment deferred the reg 74(1)(b) amount to this PR.
    • Three cases are added: one deduction per couple, the claimant exemption, and a sharer of a nil rent who is not a non-dependant.
  • housing_benefit_individual_non_dep_deduction_eligible.yaml. The 18-year-old "Eligible" case now sets current_education and would_claim_uc: false. Otherwise the age default makes them a student, and an income-free 18-year-old is entitled to UC without earned income. Both would exempt them.
  • council_tax_reduction_individual_non_dep_deduction.yaml. This is rewritten on England's 2025-26 pensioner scale (SI 2025/39).
    • A non-worker with £556 a week of investment income pays £5.00 × 52 = £260.
    • A worker's band counts it: £15.35 × 52 = £798.20.
    • Main's cases used the HB scale, and its comment deferred the remunerative-work test to this PR.
  • council_tax_reduction_pension_credit.yaml. The guarantee-credit case's son has no hours, so he pays England's £5.20 a week (it was the HB scale's £20.40). CTR is £1,729.60.
  • claimant_family_scheme.yaml. The 40-year-old adult child has no work, no award and no education, so pays para 8(1)(b)'s £5.20 × 52 = £270.40. CTR is £1,800 − £270.40 = £1,529.60. Main's £1,800 came from right=True giving zero income nothing.
  • test_rent_of_sharers_boarders_and_lodgers_properties.py (main's Give sharers, boarders and lodgers their own rent, tenure and LHA category #2006 test). Its non-dependant invariant assumed the household's HB deductions equal the sum of every non-dependant's individual amount. With reg 74(3) and (6) they are invariant 10 above, so the test now asserts that. A sharer's family is never in the HB pool, even at nil rent.
• test_non_dependant_deduction_properties.py:
  • Hypothesis populations for invariants 1-6 and 9, with explicit examples for the rare Welsh-couple and under-25 cases;
  • the exhaustive edge check (7);
  • the fixture differential (8).
• Scale verification. An independent check (Subfleet, GPT-6.1 Sol) compared all 294 CTR scale values with the amending instruments as made. It returned ALL MATCH across 42 scheme-years.
  • It also confirmed two facts: every instrument applies from 1 April, and Wales's and Scotland's pensioner and working-age scales are identical.
  • Three drafting points don't change any value:
    • the Welsh-language text of WSI 2016/50 has £195 where the English has £194 (the parameter uses the English text and says so);
    • SI 2012/3085 left "x 1/7" off two 2013-14 English band amounts;
    • SSI 2021/249 saves the 2021-22 working-age scale for 1-3 April 2022.
• Mutation testing (on 7a5cbe2, mutate3.py). There were 31 hand-made mutations, each reverting one rule. The YAML suites kill 30, and a round-4 case kills the 31st. The property tests kill 21; YAML cases kill the ten survivors.
• Runs at this head (4a0935b).

  • CI, merged with main 07985ed. All checks pass:

    • YAML: 4,038 passed.
    • Python, four shards: 7,658 passed, 61 skipped, 1 xfailed. Shard 1 timed out at the 60-minute job limit on the first attempt, before any test failed, and passed on a re-run.

  • The host's disk was paused for local test suites at this head. Every new YAML case passes in CI's YAML job. Some were also run by direct calculation:

    • against both the previous and the new code: the Maternity Allowance case, the nil-rent sharer case, and the replayed CI example;
    • against the new code: the Scottish mixed-scheme cases and the investment-income case.

    Charge non-dependant deductions for non-dependants within the claimant's benefit unit #2017's branch, which merges this head, also ran the targeted YAML suites (1,645 passed) and the three property suites (15 passed).

  • Round 6 ran the targeted suites at 0b6c3db: 1,108 + 150 YAML passed, and the property tests passed.

  • ruff format and ruff check on the changed Python files: clean.

Impact (Enhanced FRS 2024-25, certified uk-data 1.58.1)

CI's aggregate impact job for 4a0935b compared main (07985ed) with this PR merged into it. Both sides ran on the certified dataset: policyengine-uk-data 1.58.1, sha256 056e80fd8e91. Each fiscal year is a full microsimulation run on each side, and nothing is scaled. The job passed with no check broken. The model changes since round 6 (the cap basket and the joint-occupier mask) moved no figure here: spending, caseloads and incomes equal the CI run for 0b6c3db (base 76cfc9e) to the rounding shown.

Change	2025-26	2026-27	2027-28	2028-29	2029-30	2030-31	2031-32
Housing Benefit (£m)	+17.3	+18.0	+18.5	+19.0	+19.5	+20.0	+20.4
Council Tax Reduction (£m)	+6.5	+6.9	+7.3	+7.8	+9.2	+9.7	+10.4
Household net income (£m)	+23.8	+24.9	+25.8	+26.8	+28.7	+29.7	+30.8
Housing Benefit caseload	+239	+240	+240	+241	+242	+244	+245
CTR caseload (k)	+6.9	+7.0	+7.2	+7.3	+6.9	+7.0	+7.3
Absolute poverty, BHC (people)	0	+1.3k	+1.3k	+234	+235	+236	+237
Absolute poverty, AHC (people)	-2.5k	-2.5k	-2.5k	-2.5k	-2.5k	-2.5k	-2.6k
Relative poverty, BHC (people)	-1.1k	+191	+28.3k	+192	+1.9k	+194	+1.3k
Relative poverty, AHC (people)	-2.3k	-2.5k	+245	-2.5k	-2.2k	-2.5k	-2.6k

The flagged move: +25.3k pensioners in relative poverty before housing costs in 2027-28. This comes from the poverty line moving, not from anyone becoming poorer.

• The line. A separate run on the same certified data (relative_line.py; main 76cfc9e against 0b6c3db, whose relative BHC figures equal this head's) compared each person's equivalised HBAI income with the line.
  • The 2027-28 median rises by £13.62, from £41,297.37 to £41,310.99.
  • So the line (60% of the median) rises by £8.18 a year, from £24,778.42 to £24,786.60.
• Who crosses it.
  • 29.4k people fall below the line and 1.1k rise above it.
  • Of the 29.4k, 28.1k have the same income as on main, to the penny. That includes 24.9k of the 25.5k pensioners among them.
  • Only 1.3k fall below because their own income fell.
• The other years. The median doesn't move in 2026-27 or 2030-31. There, 1.3k people enter relative BHC poverty, all on lower incomes, and 1.1k leave it.
• After housing costs. The relative AHC figures move with the base for the same reason. Against main 76cfc9e, the 2030-31 AHC median rose £2.45 and 19.4k people entered relative AHC poverty, all with unchanged incomes; CI showed +16.9k that year. Against 07985ed, 43 commits later, that year shows −2.5k.

The rule-by-rule decomposition in earlier rounds was for d872f57, on an older main and dataset, so it isn't repeated here. For that head, 2026-27:

• inclusive bands alone: −£7.4m;
• remunerative work and gross income: +£0.4m;
• the couple rule: £0.0m;
• the non-dependant exemptions: +£4.2m;
• the claimant exemption: +£10.2m.

Merge order and overlaps

• Merged into main, and composed here:
  • Replace generic child and adult flags with each programme's legal definitions #1896 (is_claimant_or_partner);
  • Set DWP benefit rates to the announced amounts, not CPI projections #1925 (DWP rates, including the 2025-26 HB scale);
  • Give sharers, boarders and lodgers their own rent, tenure and LHA category #2006 (sharers, boarders and lodgers: head-only non-dependants, the rent share and the joint-liability share);
  • Give Council Tax Reduction to the household head's family, not the oldest adult's #2009 (the CTR claimant is the household head's family);
  • Assess each Council Tax Reduction claim under its own family's scheme and exemption #2015 (each CTR claim under its own family's scheme and exemption);
  • Count only the claimant's and partner's legacy awards in means tests and passports #2027 (payee-only legacy awards);
  • Means-test the Council Tax Reduction applicant and partner, not the benefit unit's claimant #2078 (the CTR applicant's means test);
  • Give Universal Credit and Housing Benefit their own benefit cap and LHA shared-accommodation rules #2089 (the UC and HB benefit caps).
• Charge non-dependant deductions for non-dependants within the claimant's benefit unit #2017 (in-unit non-dependants) is stacked on this branch. Merging this PR retargets it to main.
• Give Scotland and Wales working-age council tax reduction their own rules (UC routes, passport, child amounts) #1966 (Scotland and Wales working-age CTR) deferred non-dependant deductions to National council tax reduction: non-dependant deductions use Housing Benefit amounts and skip exemptions; pensioner gate ignores benefit exclusions #1934. It reads council_tax_reduction_non_dep_deductions, so it picks this up.
• After merge: the follow-up that switches the IS, JSA(IB) and ESA(IR) exemptions and reg 6(6) to Count only the claimant's and partner's legacy awards in means tests and passports #2027's is_on_* variables.

Axiom

axiom: uk/regulation/uksi/2006/213/74, uksi/2006/214/55, uksi/2012/2885/schedule/1/paragraph/8, wsi/2013/3029/schedule/1/paragraph/3 and schedule/6/paragraph/5, ssi/2012/319/48, ssi/2021/249/90 queued TheAxiomFoundation/rulespec-uk#407

• HB. rulespec-uk's HB pipeline (uk/policies/housing_benefit_composed_entitlement_pipeline.yaml) has the reg 74 bands for 2026-27 only, with inclusive >= edges. But it:

  • bands every non-dependant;
  • multiplies one amount by a non-dependant count;
  • has no couple rule or exemptions.

  The 2025-26 amounts are in rulespec-uk#382.

• CTR. Every CTR spine takes non-dependant deductions as a supplied input.

• The issue. All seven provisions are in axiom-corpus.

  • rulespec-uk#407 holds the encode brief: module paths, verbatim law, required outputs, and companion tests from the same 2026-27 law as this PR's YAML tests. It carries the pe-parity label.
  • An addendum there records this head's composition: apportionment, the per-applicant scheme and the couple-only cap basket.

Reproducer: policyengine_uk/tests/policy/baseline/finance/benefit/family/housing_benefit/non_dep_deduction/housing_benefit_non_dep_deduction_chain.yaml::A zero-income full-time student does not reduce the claimant's HB, policyengine_uk/tests/policy/baseline/finance/benefit/family/housing_benefit/non_dep_deduction/housing_benefit_non_dep_deduction_chain.yaml::A zero-income non-student non-dependant pays the lowest deduction, policyengine_uk/tests/policy/baseline/finance/benefit/family/housing_benefit/non_dep_deduction/household_benefits_individual_non_dep_deduction.yaml::Exactly 192 a week is in the band starting at 192, policyengine_uk/tests/policy/baseline/gov/local_authorities/council_tax_reduction/council_tax_reduction_non_dep_deductions.yaml::A Pension Credit pensioner living with a working son (issue 1934)

🤖 Generated with Claude Code

[MaxGhenis]

Composition note: #2015 (stacked on #2009 → #2006) deletes the two household flags this PR reads. council_tax_reduction_household_has_pensioner becomes council_tax_reduction_pensioner and council_tax_reduction_household_has_non_dep_exemption becomes council_tax_reduction_applicant_has_non_dep_exemption; both are BenUnit, so each jointly liable claimant uses its own family. #2015 does not add the national applicant exemption; that stays here. Whichever lands second should do three things:

1. Read the applicant's own exemption flag for applicant_exempt.
2. Do not substitute the family flag into each_member_deducted: that is evaluated on the non-dependant family, and Sch 6 para 5(3) turns on the applicant's scheme. Instead build two household pools of non-dependant families' deductions, the higher member's and the sum for a non-dependant couple on UC. Each claimant takes the sum pool when it is in Wales and its own council_tax_reduction_pensioner is false, and the higher-member pool otherwise.
3. Pick England's pensioner scale in the person-level variables by country only. Each claim's own scheme already decides whether its award uses it. The two tests that pin "no national deduction in English working-age households" then check that the working-age award is unaffected instead.

See the "National schemes and the exemption" section of #2015. (Corrected per this PR's author.)

[MaxGhenis]

Response to review r1 of #2007 (59c7082)

The review asked for changes and was right on all five P1s. Each is fixed below with the law, the change, and a test that fails on 59c7082 and passes now.

#	Finding	Disposition	Change	Test
1	is_adult is not "claimant or partner" (an 18-year-old dependant's PIP exempted the claimant; a 17-year-old claimant's PIP did not)	Fixed	New helper is_claimant_or_partner: the benefit unit's head, or any member who is not an SSCBA s.142 child or qualifying young person (is_child_or_qualifying_young_person_for_child_benefit, already on main). That is how HB Regs 2006 regs 2(1) and 19 and the English and Welsh CTR regs (reg 2(1)) define a family's members other than claimant and partner. It is used in the HB claimant exemption and in council_tax_reduction_household_has_non_dep_exemption (which the local schemes also read)	non_dependant_deduction_review_cases.yaml: "A dependent 18-year-old's PIP does not exempt the claimant" (HB £6,232.20, CTR £691.60); "A 17-year-old claimant's PIP exempts their family"
2	A dependent child's income was added to the couple's	Fixed	Joint income (reg 74(4), para 8(4)) sums only the claimant and partner of the non-dependant's family. Other members are banded on their own income. One deduction per couple (reg 74(3)) covers only the couple; each other member is deducted separately	"A parent non-dependant is banded on their own income…" (HB £1,060.80, CTR £270.40); household_benefits_individual_non_dep_deduction.yaml "A dependent young person in the family is banded on their own income"
3	HB reg 6(6): a person on IS, JSA(IB) or ESA(IR) is not in remunerative work	Fixed	The HB individual deduction treats the claimant or partner of an IS, JSA(IB) or ESA(IR) award as not in remunerative work. CTR exempts those recipients outright, so it needs no equivalent	"On income-related ESA, 16 hours is not remunerative work" (£1,060.80)
4	A self-employment loss cancelled wages in the UC no-earned-income test	Fixed	has_earned_income: employment income plus self-employed earnings floored at nil (UC reg 57(2) step 3)	"A self-employed loss does not cancel wages…" (both exemptions false; HB £1,060.80, CTR £270.40)
5	Taxable total_income omits UC, tax credits and child benefit	Fixed	Gross income for a claimant or partner adds the family's UC (before the benefit cap, which would be circular through HB), CTC, WTC and child benefit to the couple's taxable income. The disability benefits disregarded by reg 74(9) are still excluded, since they are untaxed and not added	"Universal Credit counts towards gross income" (£290 a week, £3,346.20)

Also from the review:

• Benefit receipt (IS, JSA(IB), ESA(IR), SPC, UC) now exempts only the claimant or partner of the award, in both exemption variables.
• Two omissions are now listed as not modelled, because no input exists:
  • pension-age HB's 26-week deferral of a deduction increase (SI 2006/214 reg 59(10)-(13));
  • the CTR applicant exemption's protection while AA, DLA or PIP is suspended in hospital (para 8(6)).
• The property tests' generator now includes dependent young people with their own income (in full-time education, or on approved training and so not exempt) and self-employed losses. The oracles use the claimant-or-partner rule, reg 6(6) and the floored earnings.
• The Welsh "award of universal credit" reading stays as stated in the PR description, an interpretation.

Not changed: the review's UNVERIFIED items on impact (exact zeros, decomposition, poverty causation). The impact is rerun on the new head after the host is free (the Axiom build has priority until about 10:00Z). The PR description's impact section is replaced with the rerun.

[MaxGhenis]

Response to review r2 of #2007 (6257277)

The review asked for changes again and was right on all five P1s. Findings 2-5 are fixed in the next commit, each with a YAML test that fails on 6257277 (non_dependant_deduction_review_cases.yaml, "Round 2" section). Finding 1 is fixed by adopting #1896's is_claimant_or_partner when #1896 lands; it is about to merge, and its owner asked #2007 to use it.

#	Finding	Disposition	Change	Test
1	Claimant or partner via "head or not a child benefit QYP" misclassifies explicit young partners (HB reg 2(1) "partner"), and in the reverse direction a third adult; this reaches the English local schemes through the shared exemption flag	Fixed with #1896	The helper is replaced by #1896's is_claimant_or_partner, which caps the set at the head and one partner (SSCBA s.137 couple) and can be supplied as an input for explicit couples. The r1 original student case also resolves, since #1896 presumes a 19-year-old 26 years younger than the parent is their child	The r2 explicit-couple cases are added after the rebase, along with the r1 cases
2	"No earned income" missed SMP, relievable pension contributions and the minimum income floor (UC regs 55 and 62)	Fixed	has_earned_income: employment income plus SSP and SMP, less relievable pension contributions (reg 55(4)-(5)); self-employed earnings, a loss nil (reg 57(2)), with the minimum income floor where the UC model applies it (reg 62); and other paid work	SMP £8,000 → not exempt (£1,060.80/£270.40); wages all paid into a pension → exempt (£0/£0); MIF applies → not exempt (£1,060.80/£270.40)
3	ESA(IR) receipt was extended to the working partner (HB reg 2(3A): "payable to him")	Fixed	is_award_payee: IS, JSA(IB), ESA(IR) and SPC count only for the member who reports the award, or the head where none does; UC still counts for both joint claimants (HB reg 2(3B): entitlement). Used for the exemptions and for HB reg 6(6)	Recipient and partner earning £210 a week at 16 hours → HB £2,436.20, CTR £270.40
4	Pre-cap UC overstated gross income (UC reg 81)	Fixed	Gross income now uses UC after the benefit cap. The cap is applied to the capped benefits other than Housing Benefit, which a non-rent-liable family does not receive; this avoids the HB → UC → cap → HB cycle	Pre-cap £30,000, cap £20,000 → HB £5,470.40 (365 band), CTR £551.20
5	CTR's LGFA 1992 Sch 1 para 3 disregard (an adult for whom another person is entitled to child benefit) was omitted although inputs exist	Fixed	The CTR exemption adds an adult qualifying young person (child benefit, s.142) who is not the claimant or partner of their family	19-year-old on approved training with child benefit → CTR exempt, parent alone £270.40; HB still charges both (£3,497.00)

On the property tests, the review said the oracles repeat the implementation's classification. They now take the payee as the head (the generator reports no awards) and read the child benefit QYP flag from the model. After the #1896 rebase they read is_claimant_or_partner from the model, which is not code under test. The generator now draws dependent young people and losses. SMP, pension and MIF cases are pinned in YAML rather than generated.

[juaristi22]

Codex review — changes required

Reviewed commit e92b5d107d48fe6d6cd62927059e092a75d585c2. PARTIAL; 1 confirmed blocking finding(s). Independent code/test and policy/source reviews were consolidated for the changed behavior and affected dependencies.

code-1 — [P1] Replace the QYP-complement helper with actual claimant/partner roles

The current helper _non_dependants.py:28–36 still defines the claimant/partner set as the head or anybody who is not a Child Benefit child/qualifying young person. That is neither a relationship test nor a two-person limit. It can exclude an actual young partner in education and include a third adult. The helper controls the new disability exemptions, joint gross income and one-deduction-per-couple aggregation, so this is a material remaining integration blocker. The author's latest response also explicitly defers this finding to adoption of #1896; the captured head has not adopted it.

Executed illustration: a reported continuing HB award for an intended couple aged70/19, both parent-flagged, with the19-year-old in qualifying education and receiving standard PIP daily living, plus a separate40-year-old working non-dependant earning£30,000. The QYP complement excludes the young partner and sets the claimant exemption false. With £10,000 council rent and HB assessable income isolated at zero, the deduction is£6,232.20 and final HB£3,767.80. A control supplying the legally appropriate claimant/partner disability exemption gives deduction£0 and final HB£10,000. This snapshot lacks the canonical explicit claimant/partner input; the diagnostic's relationship is stated as the intended premise, while the code defect is the inability of this helper to honor that premise.

Adopt the canonical relationship/role variable after the prerequisite lands, then add explicit young-partner and third-adult regression cases through final awards. Do not repeat the other four r2 findings: the current tree now includes statutory sick/maternity pay, employed pension deductions and MIF; the benefit-payee test; UC after the benefit cap; and CTR's adult Child Benefit QYP exemption. Confidence: high for the current unresolved helper/integration defect.

Independent source review corroborates the actual-couple and claimant/partner conditions in HB regulation 74(3)–(6), HB interpretation and England CTR paragraph 8. The overlapping code and policy finding is counted once.

The current tree addresses round-two findings 2–5; this review does not repeat those stale blockers. Current 2026 scale values were corroborated. The remaining confirmed blocker is adoption of the canonical claimant/partner roles and coverage of their downstream HB/CTR behavior.

Outstanding evidence/decision:

• The prerequisite claimant/partner implementation is absent, so the eventual composed behavior cannot yet be tested.
• Historical CTR rates/dates for 2013–2025 were not all corroborated. Complete UC net-earnings treatment, including self-employed pension contributions, remains source-unresolved and is not asserted as a second confirmed defect.

Validation: 123 focused tests passed, including round-two regression tests. An executed controlled diagnostic isolates the missing partner disability exemption's final-HB effect. The intended relationship is an explicit scenario premise, and supplied benefits/HB assessable income isolate this path; the run does not establish any actual data record's demographics. Six official source originals were inspected and their cached bytes verified.

Focused tests used an existing cached Python 3.13 environment (Core 3.32.9, NumPy 2.1.3, pandas 2.3.1, microdf-python 1.2.1, Hypothesis 6.168.2, pytest 8.4.2) with this PR’s isolated source snapshot. No dependencies were installed; this was not a freshly synced lock environment or a full-suite/population run. This is a review comment, not a formal GitHub review vote.

Live check before posting (2026-10-02T10:41:08.415704+00:00): same commit, UNSTABLE; 0 failed checks; 1 pending checks.

[vahid-ahmadi]

Review pass (Claude Code, on Vahid's behalf) at e92b5d10

Short version: the rules match the law I read, and the tests pass locally. Two things before merge: the claimant-or-partner helper still needs #1896's role variable (the Codex review's P1, which I agree with and won't repeat), and one new should-fix on the student exemption's age default in the household calculator.

1. Should-fix: with no education input, every 18- or 19-year-old non-dependant counts as a full-time student and is exempt. housing_benefit_non_dep_deduction_exempt and the CTR equivalent use is_full_time_student_non_dep, which reads current_education. That variable's formula defaults ages 18–19 to TERTIARY when nothing is supplied (variables/household/demographic/current_education.py:36). The PR body notes this, but the effect in the household calculator is large. Probe at this head, 2026, a 45-year-old HB claimant (council rent £10,000) with a 19-year-old non-dependant earning £30,000 at 37.5 hours a week:

Non-dependant	current_education	Exempt	HB deduction
age 18, no input	TERTIARY (default)	yes	£0
age 19, no input	TERTIARY (default)	yes	£0
age 20, no input	NOT_IN_EDUCATION	no	£6,232.20
age 19, NOT_IN_EDUCATION supplied	NOT_IN_EDUCATION	no	£6,232.20

On main this non-dependant is charged, so the PR introduces the gap for HB and the national CTR schemes (the English local schemes already had it through the same helper). Datasets are unaffected, since the Enhanced FRS stores current_education. Suggested fix: base the student exemption on a stored or supplied education value only, not the age default. Alternatively, don't treat the default as full-time study for someone in remunerative work (16+ hours). Either way, add the 19-year-old case above as a test.

2. The claimant-or-partner helper (Codex P1). _non_dependants.py:28-36 is still "head or not a child-benefit QYP". The author has said it will switch to #1896's is_claimant_or_partner, and #1896 is still open, so this PR should wait for that rebase and add the explicit young-partner and third-adult cases the r2 response lists.

Checked and correct:

• Law: I read HB Regs 2006 reg 74 and HB (SPC) Regs reg 55 on legislation.gov.uk.
  • The bands are "not less than" at the lower edge, and the 2026-27 amounts are £20.40 to £131.45 on £192/£279/£365/£485/£605.
  • Couples get one deduction, the higher, on joint income (reg 74(3)-(4)).
  • The claimant exemption covers blindness, AA, the DLA care component, the PIP daily living component and AFIP (reg 74(6)).
  • The student exemption is reg 74(7)(c)-(e); the under-25 IS/JSA(IB)/ESA(IR)/UC-no-earned-income exemption is reg 74(8); SPC is reg 74(10) (reg 55(9) for pension age).
  • England CTR Sch 1 para 8 has £5.20/£10.60/£13.30/£15.95 on £279/£485/£605, with its own para 8(7)-(8) exemptions, including the LGFA Sch 1 disregard and the UC limb.
• Remunerative work: banding applies only at 16+ hours and not on an income-related benefit (reg 6(6)); anyone else pays the (1)(b) amount.
• has_earned_income counts miscellaneous_income, the same as the UC model's own uc_mif_capped_earned_income, so it is consistent within the model.
• Tests at this head: the HB and local-authority YAML suites pass 205/205, and test_non_dependant_deduction_properties.py passes 6/6. CI's Test job was still running when I checked; the other six checks pass.

[MaxGhenis]

Composition note from #2078 (stacked on #2009), which makes the CTR means test assess the applicant and partner rather than the benefit unit's claimant and partner.

#2078 changes council_tax_reduction_household_has_non_dep_exemption to read is_council_tax_reduction_applicant_or_partner (new Person variable) in place of is_claimant_or_partner, following "the applicant or his partner" in SI 2012/2885 Sch 1 para 8(6). The two differ where the household head is liable but is not the benefit unit's claimant or partner. Example: a blind grandmother head sharing a benefit unit with a young couple, who were not exempt before.

When this PR's national exemption (council_tax_reduction_non_dep_deduction_exempt) composes with #2078, it should use the same identity for the applicant or partner. #2015 replaces the household flag with a per-family one and gets the same note.

[MaxGhenis]

Response to round 6

Round 6 (an independent Subfleet review of 0b6c3db, the composition with main) asked for changes and found no P1. Its report is below this response. Each finding is addressed at 4a0935b.

P2-1: main's sharers property test failed.

• 8e60349 replaced the assertion with the pooled amount (one per couple, the higher) times the share borne by families whose claimant isn't exempt, plus an upper bound. The module docstring now states that invariant.
• CI's derandomized run of that test then found a case on 89ba1a1. At nil rent, a sharer was counted as a non-dependant (is_non_dependant_of_household_head tests the family's own rent, which is nil) while also bearing a share. The household's deductions were £510.90 against £1,021.80.
• 4a0935b keeps any family that bears a share of the rent out of the HB pool: it is a joint occupier, not a non-dependant (HB Regs 2006 reg 3(2)(d)).
  • The test's pool leaves out sharer families.
  • A YAML case pins it: "A sharer of a nil rent is not a non-dependant", [0, 0].
  • Replaying the CI example fails on 89ba1a1 and passes on 4a0935b. The YAML case gives [1, 0] on the old formula.
  • No aggregate moves (CI impact below).

P2-2: the body described the pre-composition head. The body is rewritten for 4a0935b:

• the composition with Give sharers, boarders and lodgers their own rent, tenure and LHA category #2006, Give Council Tax Reduction to the household head's family, not the oldest adult's #2009, Assess each Council Tax Reduction claim under its own family's scheme and exemption #2015, Count only the claimant's and partner's legacy awards in means tests and passports #2027 and Give Universal Credit and Housing Benefit their own benefit cap and LHA shared-accommodation rules #2089, including the per-applicant scheme recompute;
• every main test that changes, and why;
• the "Not modelled" list without the stale apportionment and benefit_cap_reduction lines;
• impact from CI's certified-data run at this head;
• the 2027-28 relative-poverty flag, explained by the line comparison.

P3-3: no test of the mixed-scheme recompute. council_tax_reduction_non_dep_deductions_mixed_schemes.yaml adds the review's probes A2, B3, B4, E1 and E2.

• My first version had wrong person-level expectations in E1 and E2. The 70-year-old applicant is on Pension Credit, so their own exemption flag is true; as the review noted, that flag matters only for a non-dependant.
• 89ba1a1 corrects the expectations. Both cases now also assert who is eligible as a non-dependant.

P3-4: the cap basket counted every member. 2574473 counts a person's own contributory JSA and ESA, incapacity benefit, SDA and Maternity Allowance only for the claimant or partner (UC Regs 2013 reg 79(1)).

• The new case gives a 17-year-old her own Maternity Allowance. It gives HB £5,470.40 and CTR £551.20 on 8e60349, and £6,232.20 and £691.60 now (run by direct calculation on both).
• Main's own cap total still leaves out Maternity Allowance. The UK hub has taken that as a separate follow-up.

P3-5: thin 2025-26 coverage. There are now 2025-26 second- and middle-band cases (£45.15 and £101.35). 9ae8f16 also replaces main's "Other investment income does not move a worker's band" with the lawful band, £126.65 a week.

CI at 4a0935b:

• every check passes;
• YAML: 4,038 passed;
• Python: 7,658 passed, 61 skipped, 1 xfailed. Shard 1 timed out at the 60-minute job limit before any failure and passed on a re-run. That shard runs 47-60 minutes on every PR.

Aggregate impact on certified uk-data 1.58.1 (base main 07985ed). Spending, caseloads, incomes, absolute poverty and relative BHC poverty are identical to the run for 0b6c3db. Only relative AHC poverty differs, because the base moved.

A delta review (round 7) of 4a0935b gates the merge.

---

Round 6 review (independent Subfleet review of 0b6c3db)

VERDICT: REQUEST CHANGES

Review of #2007 at 0b6c3db (composing merge bd5dd8d plus the #2078 merge), against main 76cfc9e.

The model code is right on every composition point I tested. No P1 was found. Two P2s block merge at this exact head, and neither needs a model change:

1. At 0b6c3db one of main's tests fails, so CI is red. The test-only follow-up 8e60349 fixes it; I reviewed it and it passes locally.
2. The PR body describes the pre-composition head d872f57 and misstates what this head does and what it changes.

Clearing them needs CI green on 8e60349 plus a refreshed body. A delta check of those two suffices unless the model code changes.

Findings

P2-1: main's sharers property test fails at the reviewed head (fixed by test-only 8e60349)

• File: policyengine_uk/tests/test_rent_of_sharers_boarders_and_lodgers_properties.py:285 (main's Give sharers, boarders and lodgers their own rent, tenure and LHA category #2006 assertion np.allclose(per_household(sim, hb), hb_total, ...)).
• Why it fails: the household's HB deductions no longer equal the sum of every non-dependant's individual amount. Apply Housing Benefit and national Council Tax Reduction non-dependant deductions as legislated #2007 applies two rules the assertion predates:
  • HB Regs 2006 reg 74(3): "Only one deduction shall be made under this regulation in respect of a couple … the higher amount shall be deducted".
  • Reg 74(6): "No deduction shall be made in respect of any non-dependants occupying a claimant's dwelling if the claimant or his partner is … blind …".
• Failing input: the derandomised Hypothesis example, in review-artifacts/sharers/at-0b6c3db29.log and in CI.
  • Actual per-household HB deductions: [£1,021.80, £510.90, 0, …].
  • Main's assertion expects [£1,021.80, £2,043.60, 0, …].
  • The model's value is the lawful one: one deduction per couple, apportioned over the liable families, with none borne by an exempt claimant.
• CI: run 39506301397 ("Code changes", 0b6c3db) concluded failure. That test is the only failure in 7,533 tests: 7,470 passed, 61 skipped, 1 xfailed.
• Resolution: 8e60349 is test-only (+16/−1, this one file). It replaces the assertion with "pool of per-family couple-max amounts × Σ rent share of non-exempt liable families", plus an upper bound.
  • That oracle matches the code: housing_benefit_non_dep_deductions.py:39,48.
  • Run against 0b6c3db's model code, it passes: 1 passed in 341.76 s.
• P3 left in 8e60349: the module docstring's invariant 4 (lines 21–22) still says HB deductions "sum to the household's non-dependants' deductions".

P2-2: the PR body does not describe this head

Source: the live body (gh pr view 2007, head 8e60349), saved as review-artifacts/pr-body-live.md.

Composition says the opposite of the code.

• "Not modelled: … joint-occupier apportionment (reg 74(5), para 8(5)). Apportionment is Give sharers, boarders and lodgers their own rent, tenure and LHA category #2006."
• The code now multiplies Apply Housing Benefit and national Council Tax Reduction non-dependant deductions as legislated #2007's deductions by share_of_household_rent (HB) and by council_tax_reduction_joint_liability_share (CTR).
• Each CTR applicant is recomputed under its own scheme, which the body does not describe. That recompute is the main design change in bd5dd8d.

Removed variables are named.

• "council_tax_reduction_household_has_non_dep_exemption existed … It now applies to the national schemes too." That variable is gone; the code reads main's BenUnit council_tax_reduction_applicant_has_non_dep_exemption.
• "Main's benefit_cap_reduction also leaves out Maternity Allowance." benefit_cap_reduction was removed by Give Universal Credit and Housing Benefit their own benefit cap and LHA shared-accommodation rules #2089.

Impact is for another head.

• The body says: "Real runs of main (9eccc54) and this branch (d872f57)", with HB +£7.5m and CTR +£3.1m in 2026-27.
• The CI aggregate impact at this head (certified uk-data 1.58.1, base 76cfc9e) shows:
  • HB: +£18.0m in 2026-27 (+£17.3m to +£20.4m, 2025-32).
  • CTR: +£6.9m in 2026-27 (+£6.5m to +£10.4m).
  • Household net income: +£23.8m to +£30.8m.
  • CTR caseload: +6.9k to +7.3k.
  • A flagged +25.3k pensioners in relative BHC poverty in 2027-28, which the body does not mention.

Changed main tests aren't listed. "Two existing tests change" omits:

• claimant_family_scheme.yaml (£1,800 → £1,529.60);
• the HB investment-income case (0 → £1,021.80);
• the rewritten CTR investment-income cases;
• the two removed 2014 cases.

Merge-order section is stale. It calls #2006 a draft; #2006 is merged.

Failing input: N/A (disclosure). Expected:

• the body describes the composition (per-family amount × Give sharers, boarders and lodgers their own rent, tenure and LHA category #2006 share; per-applicant scheme recompute; per-family claimant/applicant exemption);
• it lists main's tests that changed and why (all four are justified; see item 2 below);
• it reports impact for this head (impact8/9 or the CI report), with the 2027-28 relative-poverty flag explained. Line comparison: the 2027 BHC median moves +£13.62, and 28.1k of the 29.4k entrants have unchanged income, so it is a poverty-line artifact.

P3 (non-blocking)

P3-3: the new mixed-scheme recompute path has no test in the PR.

• The "other" branch is at council_tax_reduction_non_dep_deductions.py:51-83.
• No YAML case has a jointly liable household whose applicants are on different national schemes.
• The property generator has one claimant family and one non-dependant family per household (test_non_dependant_deduction_properties.py:402), so it cannot reach this path.
• My probes and the 96-household differential below pass. I suggest adding probes A2, B3/B4 and E1/E2 as YAML cases.

P3-4: the after-cap UC basket counts every member, not just the couple.

• _non_dependants.py:40,96 uses add(benunit, …), which sums contributory benefits and MA for every member.
• UC reg 79(1) counts "the welfare benefits to which a single person or couple is entitled", and main's benefit_cap_welfare_benefits counts only the couple.
• Probe G4: a 19-year-old QYP with £2,000 of own contributory ESA. The helper gives £15,821; reg 79(1) gives £17,821, which is what main's universal_credit gives.
• This is immaterial in practice. UC reg 5(5) says "A person who is receiving universal credit, an employment and support allowance or a jobseeker's allowance is not a qualifying young person", so only a QYP on MA could reach it. Masking the person-level benefits with is_claimant_or_partner would align the basket with main.
• The MA divergence itself (probe G3: helper £15,821 against main's £17,821) is legally right (WRA 2012 s.96(10)(i)) and disclosed.

P3-5: 2025-26 mid-band HB coverage is thinner. Main's 2025 "Middle amount" (101.35) and second-band (45.15) cases were replaced:

• by 2026-27 exact-edge cases;
• by 2025 top and lowest spot checks.

The replacement is justified: main's cases had no hours input, so under reg 74(2) they now pay the lowest amount.

1. Composition with main

HB. housing_benefit_non_dep_deductions is computed as follows:

• the per-family amount (deduction_per_family, one deduction per couple, the higher);
• times is_non_dependant_of_household_head and share_of_household_rent;
• zero under the per-family claimant exemption.

This composes correctly with #2006's reg 3 and 74(5) treatment:

• Joint occupiers, boarders and lodgers are rent-liable, so they are never non-dependants.
• Each liable family bears its share.
• Subtracting the family's own amount is harmless, because a liable family's members are masked to 0.

Probes:

• C (boarder, lodger): pays_rent_to_householder = [0, 1, 0]; is_non_dependant_of_household_head = [0, 0, 1]; HB [£1,060.80, 0, 0]; CTR [£270.40, 0, 0]. That is reg 74(1)(b) at £20.40 × 52 and para 8(1)(b) at £5.20 × 52.
• D1 (rent sharing): a head couple plus a single joint tenant, with a non-dependant couple on a joint £400 a week. Shares are [2/3, 1/3, 0], so HB is [£3,646.93, £1,823.47, 0] (£105.20 × 52 apportioned).
• D2 (sharer blind): [£3,646.93, 0, 0].

CTR. For each applicant B, council_tax_reduction_non_dep_deductions uses:

• the person variable when B's own council_tax_reduction_pensioner equals the head family's;
• otherwise individual_non_dep_deduction(…, ~head_pensioner).

The result is then multiplied by council_tax_reduction_joint_liability_share and zeroed by the per-family council_tax_reduction_applicant_has_non_dep_exemption. This is right for joint applicants on different schemes.

Probe	Actual = expected (annual 2026)
A1 England: pension-age head + working-age sharer, non-dependant £300 a week	CTR [£275.60, 0, 0] (£10.60 × 52 / 2); HB [£1,673.10, £1,673.10, 0]
A2 England: working-age head + pension-age sharer (recompute path)	CTR [0, £275.60, 0]; the person variable reports the head's scheme (0)
A3 England: two pensioner applicants, head blind	CTR [0, £275.60, 0]; HB [0, £1,673.10, 0]
B1 Wales working-age applicant, UC non-dependant couple (joint £357.69)	£1,388.40 (both members, Sch 6 para 5(3))
B2 Wales pension-age applicant, same couple	£694.20 (one deduction, Sch 1 para 3(3))
B3 Wales jointly liable: pension head + working-age sharer	[£347.10, 0, £694.20]
B4 Wales: working-age head + pension sharer	[£694.20, 0, £347.10]
B5 Wales working-age, UC couple with one earner	£348.40 (non-earner exempt, 5(8)(c))
E1 Scotland: working-age head + pension sharer; non-dependant on ESA(IR) with a working partner	[0, £284.70, 0]: partner rule (SSI 2021/249 reg 90(8)(a)) only for the working-age applicant
E2 Scotland: pension head + working-age sharer	[£284.70, 0, 0]

Welsh and Scottish working-age-only rules. Wales Sch 6 para 5(3) ("other than where there is an award of universal credit") is absent from Sch 1 para 3(3). The PR applies the UC-couple rule only when the applicant's own family is working-age, and the Scottish partner rule only in the working-age scheme. Both are correct.

Leakage into the English local schemes. There is none.

• council_tax_reduction_non_dep_deductions is read only by simulated_council_tax_reduction_benunit. There it feeds national_ctr, which is gated by the applicant's own national_scheme.
• The PR changes no local-scheme file, _legacy.py or the shared eligibility variable.
• Branch vs main differential, 5 councils × 3 household shapes: all 45 local-scheme value arrays are identical, and so are eligibility, applicant exemption, joint-liability share and pensioner status.

Cap.

• Cycle claim verified. Patching the helper to read uc_benefit_cap_reduction raises CycleError: housing_benefit → … → household_benefits_individual_non_dep_deduction → uc_benefit_cap_reduction → benefit_cap_welfare_benefits → housing_benefit_pre_benefit_cap. Reading universal_credit cycles the same way.
• Against UC regs 79-81 as main models them:
  • G1, a binding cap: helper £17,821 = universal_credit £17,821. By hand: 20,000 + child benefit 4,199 − cap 22,020 gives an excess of 2,179.
  • G2, childcare element larger than the excess: £20,000 on both (reg 81(2)).
  • Only the basket differs (P3-4).

Randomised differential against an independent oracle.

• 96 jointly liable households over two seeds; 45 have mixed schemes and 7 are Welsh working-age UC couples.
• They vary country, applicant ages, blindness, single or couple non-dependants, earnings, hours and UC.
• Result: 96/96 pass for HB, CTR, pensioner status, and zero for the non-dependant's own family.

2. Main's tests that this PR changed

All four changes are justified by the law; none hides a regression.

council_tax_reduction_individual_non_dep_deduction.yaml (rewritten).

• SI 2012/2885 Sch 1 para 8(1)-(2) bands only a non-dependant "in remunerative work", on "normal gross weekly income".
• A non-worker on £556 a week of investment income pays the lowest amount: £5.00 × 52 = £260 (2025).
• A worker's band counts the investment income: £739.02 a week falls in the £577 band, £15.35 × 52 = £798.20.
• Main's own comments marked these cases as stopgaps "until Apply Housing Benefit and national Council Tax Reduction non-dependant deductions as legislated #2007".

claimant_family_scheme.yaml (£1,529.60).

• The 40-year-old adult child has no work, no award and no education, so is a non-dependant of the pension-age head's claim and pays para 8(1)(b)'s £5.20 × 52 = £270.40. £1,800 − £270.40 = £1,529.60.
• Main's £1,800 came from amount.calc(0, right=True) returning 0 for zero income, which is the bug this PR fixes.

housing_benefit_non_dep_deductions.yaml.

• Main's Give sharers, boarders and lodgers their own rent, tenure and LHA category #2006 cases are kept unchanged.
• The investment-income non-worker now pays reg 74(1)(b)'s 2025 amount, £19.65 × 52 = £1,021.80; main's 0 was the same right=True artifact.
• Two cases are added: one per couple, and the claimant exemption.

The removed 2014 cases in council_tax_reduction_non_dep_deductions.yaml.

• Run on this head, they fail with ParameterNotFoundError: 'gov.dwp.JSA.income.capital.tariff_income[threshold]' … 2014-01-01. That is a main parameter Apply Housing Benefit and national Council Tax Reduction non-dependant deductions as legislated #2007 doesn't own, so removal is justified.
• The UC limb's false branch is still covered by the "Scotland: no Universal Credit exemption in 2015-16" case.

3. Earlier rounds still hold

• Round-4 counterexamples: 6/6 checks.
  • MA cap: UC £10,890.20, HB £6,232.20, CTR £691.60.
  • Self-employment loss: HB £3,346.20, CTR £551.20.
• Round-4 fresh inputs: 14/14.
• Rounds 2/3:
  • round-2 reruns: 21/21 across 10 households;
  • fresh probes: 4/4;
  • Scotland controls: 8/8.
• Review-cases YAML: 0 failures, inside the YAML run below.

To compose with main's API, I changed the old probes in two ways: benefit_cap → uc_benefit_cap (#2089) as an input and diagnostic. Expected values are unchanged.

4. Tests

Command	Result
export UV_CACHE_DIR="$PWD/.cache/uv" UV_PYTHON_INSTALL_DIR="$PWD/.cache/python"; uv sync --extra dev	OK, 7 min 45 s
uv run policyengine-core test …/housing_benefit …/universal_credit …/gov/local_authorities -c policyengine_uk	1,108 passed, 0 failed (1,454.5 s), EXIT 0
uv run --frozen policyengine-core test on LHA_claimant_composition.yaml, gov/dwp/benefit_cap.yaml, gov/dwp/benefit_cap/ and benefit_cap_by_scheme.yaml	150 passed, EXIT 0
uv run pytest -q -p no:cacheprovider -n 4 policyengine_uk/tests/test_non_dependant_deduction_properties.py	6 passed (928 s), EXIT 0
Main's test_rent_of_sharers_boarders_and_lodgers_properties.py -k test_non_dependants at 0b6c3db	1 failed (P2-1)
Same test from 8e60349, against 0b6c3db's model code	1 passed
Removed 2014 cases	2 failed with ParameterNotFoundError (main's JSA parameter)
review-artifacts/r6_composition_probes.py	31/33 pass. The 2 misses were my expectation error: the 70-year-old applicants are on Pension Credit, so their own person-level exempt flag is true. That flag only matters for non-dependants, and the non-dependant entries match.
review-artifacts/r6_cycle_check.py	shipped: no cycle; both alternatives: CycleError
review-artifacts/r6_cap_differential.py	G1/G2 agree; G3 disclosed MA divergence; G4 the P3-4 divergence
review-artifacts/r6_mixed_scheme_differential.py (two seeds)	48/48 + 48/48
review-artifacts/r6_local_leak_differential.py, branch vs a git archive origin/main export	45/45 local-scheme arrays identical
Prior-round reruns (review-artifacts/prior/)	6/6, 14/14, 4/4, 21/21, 8/8
CI run 39506301397 (0b6c3db)	1 failed (test_non_dependants), 7,470 passed
gh pr checks 2007 (8e60349, read 2026-10-21 02:04Z)	YAML and Python 2/4 pass; Python 1/4, 3/4 and 4/4 pending

The first probe run was invalid: my builder omitted household members. It is kept as r6_composition_probes_INVALID_no_household_members.json, and the corrected rerun is the one reported.

5. Impact

• UNVERIFIED: compare8.md, explain8.txt and decomp8_*.json had not landed. base8.log was still queued and base9.log ended in a KeyError in impact.py:132.
• From the CI aggregate report on certified data at this head (all signs as the design predicts):
  • HB up. The exemptions, the couple rule and the claimant exemption outweigh the new lowest amount for zero-income non-workers.
  • CTR up. Each scheme's own scale is far below the HB scale main used. The caseload rises +6.9k as smaller deductions bring families into entitlement.
  • Absolute poverty about 0. The flagged relative-BHC pensioner move is the median-line artifact described in P2-2.
• UNVERIFIED: how far the 2026-27 HB change (+£18.0m now, against +£7.5m in the body) comes from the composition rather than from main or dataset drift.
• No individual survey records were printed.

Not verified

• CI on 8e60349 (Python 1/4, 3/4 and 4/4 pending).
• The impact8 decomposition.
• The full model suite run locally (CI at 0b6c3db covers it: one failure).
• The CTR scale values for years other than 2026-27, which earlier rounds verified.

All scripts, logs, statutory texts and the review are committed on the detached workspace (review-artifacts/, including review-2007-r6.md). Nothing was pushed, and I made no GitHub comment or PR edit.