Apply the Pension Credit severe disability residence and carer conditions

[MaxGhenis]

Summary

The Pension Credit severe disability addition (severe_disability_minimum_guarantee_addition) now applies the statutory residence and carer conditions:

• Residence. No other person aged 18 or over may live with the claimant (and partner), unless their presence is ignored.
• Carers. A carer benefit removes or halves the addition only when it is paid for caring for a qualifying claimant or partner. Each carer's award is attributed to one person.

This builds on #1896, which added the couple rule (both partners qualify, or one qualifies and the other is blind). #1896 has merged and this PR targets main; its own change is git diff 3525263f...5008b53b (10 files).

Law encoded

Source: legislation.gov.uk current revised text (CLML, valid from 2026-07-16), fetched 2026-09-30.

State Pension Credit Regulations 2002 (SI 2002/1792):

Provision	What it says	Where
reg 6(4)	the addition applies if the claimant is treated as severely disabled under Sch I para 1	severe_disability_minimum_guarantee_addition
reg 6(5)(a)-(b)	£86.05 a week; £172.10 "if paragraph 1(1)(b) ... is satisfied ... and no one is entitled to and in receipt of" a carer benefit "in respect of caring for either partner"	same (two rates = 2 × the addition parameter)
Sch I para 1(1)(a)(i), (b)(i), (c)(i)	qualifying benefits: AA, DLA care middle/highest, PIP daily living, AFIP (and Scottish equivalents)	receives_severe_disability_addition_qualifying_benefit (existing relevant_benefits list)
Sch I para 1(1)(a)(ii), (b)(ii) first limb, (c)(iii)	"no person who has attained the age of 18 is normally residing with the claimant ... other than a person to whom paragraph 2 applies"	meets_severe_disability_addition_residence_condition, is_counted_resident_for_severe_disability_addition, new parameter severe_disability.other_resident_age_limit (18)
Sch I para 2(1), 2(2)(a), (b), (f)	ignore qualifying-benefit recipients, people certified blind or severely sight impaired, and qualifying young persons (reg 4A)	is_counted_resident_for_severe_disability_addition
Sch I para 1(1)(a)(iii), (b)(ii) second limb, (c)(iv)	no one "entitled to and in receipt of" Carer's Allowance or Carer Support Payment "in respect of caring for" the qualifying person (for (b): for one only, or neither)	num_severe_disability_addition_qualifiers_cared_for
Sch I para 1(1)(b), (c)	both partners qualify; failing that, one qualifies and "the other partner is certified as severely sight impaired or blind"	from #1896, unchanged

Carer attribution. The data do not say whom a carer cares for, so each award is attributed as follows:

• No one is their own carer. Each award covers one person (SSCBA 1992 s.70(7)), and each person has at most one carer (s.70(7ZA)).
• A carer cares for someone in their own benefit unit who receives a qualifying benefit, if there is one: a qualifying claimant or partner first, then another member such as a disabled child.
• Every other carer in the household, including any left over in their own unit, covers one qualifying claimant or partner of another benefit unit who is not yet cared for. Units are taken in descending order of their eldest member's age; on ties, the unit whose eldest member comes first in person order (the order people are listed in the input or dataset) goes first.
• Replace generic child and adult flags with each programme's legal definitions #1896's version counted every carer benefit in the benefit unit against its qualifiers and ignored carers in other benefit units. Replace generic child and adult flags with each programme's legal definitions #1896's 9b13503 stopped the recipient's own award counting.

Not modelled (documented on the variables):

• Hospital and backdating deeming: para 1(2)(a)-(c), and hence reg 6(5)(b)'s "otherwise than by virtue of paragraph 1(2)(b)".
• Other para 1 heads: para 1(3) (sight regained, 28 weeks) and polygamous marriages.
• Other para 2 persons: 2(2)(c) (sight regained), 2(2)(d)-(e) (charity carers), 2(3)-(4) (a new carer's first 12 weeks), 2(5) (commercial lodgers and landlords), 2(6)-(7) (joint occupiers), and para 3 ("resides with"). Each of these ignores a person, so leaving it out can only withhold the addition.
• Prisoners and religious orders: reg 6(3).
• UC carer element (para 1(4)). Reading universal_credit here would be circular, because uc_non_dep_deduction_exempt reads pension_credit. A UC carer living in the household already stops the addition through the residence condition, unless their presence is ignored.
• Carers outside the household, whom the data cannot link.
• Receipt is receives_carer_benefit, so it inherits carers_allowance, which does not apply the overlapping-benefit rule. In the calculator, a pensioner carer whose State Pension exceeds CA is treated as in receipt; a chip covers the fix. In dataset mode, CA comes from reported receipt.
• is_blind is not a policyengine-uk-data input. In microsimulation the blind-partner route (para 1(1)(c)) never applies, and a blind resident is counted rather than ignored (para 2(2)(b)).

Pension-age HB and CTR (with #1938)

#1938 makes pension_age_severe_disability_premium read this variable, so pension-age HB and CTR follow it. HB(SPC) Regs 2006 (SI 2006/214) Sch 3 paras 6 and 12(1) have the same qualifying benefits, the same amounts, and nearly the same conditions. An independent audit (executed cases, law read verbatim) found where they differ:

1. An 18- or 19-year-old qualifying young person of another family. SPC para 2(2)(f) ignores any reg 4A qualifying young person. HB reg 3(2)(a) excludes from "non-dependant" only members of the claimant's own family, and para 6(6) does not ignore them. So HB withholds the premium where Pension Credit pays the addition. The shared definition overstates HB and CTR in that case. This belongs in pension_age_severe_disability_premium (Give pension-age HB and CTR the severe disability premium Pension Credit uses #1938) as an HB-specific residence test. Details are in the coordination comment on Give pension-age HB and CTR the severe disability premium Pension Credit uses #1938.
2. Blind claimant, partner on AA. HB 6(3) with 6(2)(a)(i) gives the premium only if the partner on AA is the claimant, whereas SPC (c) is symmetric. The shared result is lawful if the couple nominate that partner to claim, as HB(SPC) reg 63(1) lets them.
3. Unmodelled differences, none of which can change a household the model represents:
   • child disability payment in 6(6)(a);
   • hospital deeming, where 6(7) and 12(1)(b)(i) cover more cases;
   • 6(8) backdating and fraud-restriction deeming;
   • HB has no 12-week new-carer exception.

The documentation here now lists these differences, rather than saying HB applies the same test.

Merge order. #1938 alone has no residence test, so on its own it pays the HB/CTR premium to a pensioner living with an adult non-dependant; the law gives 0. It also conflicts with this PR (and #1896) in severe_disability_minimum_guarantee_addition.py and test_carer_support_payment.py. Resolved to this PR's formula and #1896's test file, every #1938 test passes on the merge: 3 Hypothesis properties, including its invariant that the pension-age HB/CTR applicable amount is at least the Pension Credit minimum guarantee, and 122 YAML cases. Suggested order: #1896, then this PR, then #1938 rebased onto it. #1896 now passes on its own: its 143cc3c corrected the CSP test that this PR had previously rewritten, and this PR no longer touches that file.

#1946 applies the same kind of conditions to the working-age legacy premium (HB Regs 2006 Sch 3 para 14), with its own qualifying-benefit list and non-dependant helper. It touches none of this PR's files. Its has_non_dependant_for_severe_disability_premium is one way to give #1938 the HB-specific residence test.

Invariants and tests

• Properties (test_severe_disability_addition_properties.py, Hypothesis, derandomised). Households are generated with a pension-age benefit unit and up to two other benefit units, each single or a couple and possibly with dependants, each person drawn with a disability benefit, blindness, a carer benefit and education.

  1. Differential: the number of rates equals that of severe_disability_addition_reference.py. A separate agent wrote that reference from the law text without seeing this implementation. Its carer rule was later rewritten as explicit person-to-person matching: brute force within a unit, bipartite feasibility across units. So the reference does not share the implementation's counting formula.
  2. Bounds: the addition is 0, 1 or 2 weekly rates × 52, and 2 only for a couple. The qualifying-benefit flag matches the para 1(1)(a)(i) list for every drawn benefit category.
  3. Residence (metamorphic): adding a counted adult resident removes every other unit's addition. Adding an ignored resident who does not qualify (blind, a qualifying young person, under 18) leaves every addition unchanged. Adding one who does qualify never lowers any addition.
  4. Carers (metamorphic): giving anyone a carer benefit never raises any addition.
  5. Symmetry: swapping the drawn attributes of a couple's partners leaves their addition unchanged.

• Mutation check. Deliberately broken versions of the code were run against the tests, and each was caught:

  • no residence test: caught by the differential and residence properties;
  • a claimant's own carer benefit counted: caught by the differential;
  • qualifying young persons counted as residents: caught by the residence property.
  • The independent review ran nine more, and the existing tests caught seven. The other two, age > instead of >= and reversed tie order, are now caught by new YAML cases.

• YAML. severe_disability_residence_and_carers.yaml has 32 cases, covering:

  • each residence exception and the age-18 boundary;
  • qualifying young persons in and outside the benefit unit, and a third adult in the couple's unit;
  • each carer-attribution case: one award across benefit units, a carer of their own disabled child, a qualifying carer with no one else in their unit, a surplus carer in either placement, and the age-tie order;
  • a floor that keeps the addition from going negative with malformed inputs.

  Replace generic child and adult flags with each programme's legal definitions #1896's own SDA cases pass unchanged.

• Multi-unit fuzzing. Households of 3-5 benefit units, with interleaved members, age ties, and claimant flags both supplied and derived: 0 mismatches against the reference over about 30,000 benefit units.

• Suites on this branch:

  • CI run 36787353663 on 214a2fa: all six jobs pass, with 1,507 YAML cases passed and pytest 2,068 passed, 1 skipped. The later head 228de58 adds two YAML cases and documentation only; its run is 36805257533.

• Independent review. An Opus reviewer requested changes at bde8f7d (eight findings, all fixed; see the PR comments). It then approved at 214a2fa. Its approval covers:

  • a proof that the within-unit carer count is the best possible pairing;
  • an exhaustive check of 1,700 unit configurations;
  • 0 mismatches in about 42k fuzzed benefit units against the reference and its own matching check;
  • reproduced impact figures.

  Its last test gap and wording nits are fixed in 228de58.

• Independent audits. Opus agents read the law verbatim and executed every claim; each material finding was re-verified by a second agent.

  • A 498-household grid of single and couple cases × qualifiers × blind flags × 13 carer placements × residence gave 0 mismatches against an independent oracle.
  • The audits found the cross-unit carer double count and the overstated HB docs, both fixed here.
  • They also found issues outside this PR's scope; see Follow-ups.

Impact (real microsimulation runs)

Enhanced FRS 2024-25 (policyengine-uk-data 1.57.4, SHA-256 03fe15e4…68d4). Each figure compares real Microsimulation runs of main 3525263f (the main this head last merged, which includes #1896) and this head 5008b53b; nothing is scaled.

	2025	2026	2027	2028
Pension Credit, £m	−59.5	−63.2	−64.6	−65.6
Benefit units no longer receiving Pension Credit, thousands	5.3	5.2	5.0	5.4
Households with lower net income, thousands	16.5	16.6	16.4	16.5

• Nobody gains.
• Housing Benefit falls by a much smaller amount that rests on fewer than 10 sample records, and Council Tax Reduction changes by less than £0.1m, so neither is given. For the same reason household net income is not given; it falls by slightly more than Pension Credit.
• In 2026, HBAI absolute poverty after housing costs rises by 0.077 percentage points (54k people). The other HBAI poverty measures either move for fewer than 10 sample households or move both ways as the median shifts, by a negligible net amount, so they are not reported.

For 2025 and 2026, the earlier head's run (143cc3c4 → 214a2fad, measured on 30 September) gave the same Pension Credit figures to £0.1m; it did not cover 2027 or 2028. Analyses made on that head (a carer-only counterfactual, a comparison with #1938 merged, and a breakdown of why CTR did not move) were not re-run on this head and are not repeated here; on this head CTR changes by less than £0.1m in every year.

Follow-ups (not in this PR)

• The 2023-24 to 2025-26 addition rates are CPI projections (£75.69, £81.22, £83.25) rather than the statutory £76.40, £81.50 and £82.90. Set DWP benefit rates to the announced amounts, not CPI projections #1925 sets them.
• Chips filed:
  • apply the overlapping-benefit rule to Carer's Allowance;
  • the HB non-dependant deduction exemptions (HB(SPC) reg 55(6)-(9)), which National council tax reduction: non-dependant deductions use Housing Benefit amounts and skip exemptions; pensioner gate ignores benefit exclusions #1934 covers for CTR only;
  • Scottish qualifying benefits (PADP, ADP daily living, Scottish adult DLA) as inputs and data.
• CTR Guarantee Credit passport: Disregard guarantee credit recipients' income and capital in pensioner council tax reduction #1909 (merged 1 October).
• Reg 4A(5) in Replace generic child and adult flags with each programme's legal definitions #1896's qualifying-young-person test (income support in the young person's own right). Flagged on Replace generic child and adult flags with each programme's legal definitions #1896.

axiom: uk/regulation/uksi/2002/1792/schedule/I TheAxiomFoundation/rulespec-uk#360 queued; uk/regulation/uksi/2006/214/schedule/3 TheAxiomFoundation/rulespec-uk#383 queued

🤖 Generated with Claude Code

[MaxGhenis]

Pension-age CTR schemes: checked against the law text (legislation.gov.uk, fetched 2026-09-30). This matters once #1938 makes pension-age CTR read this variable.

Scheme	Premium conditions	Non-dependant definition	Single rate
England, SI 2012/2885	Sch 2 para 6 (valid 2026-02-13)	reg 9(2)(a)	£86.05 (Sch 2 Part 4)
Scotland, SSI 2012/319	Sch 1 para 7 (valid 2026-04-01)	reg 3(2)(a)	£86.05
Wales, WSI 2013/3035	Sch 2 para 6 (valid 2026-01-26)	not read	£86.05

All three follow HB(SPC) Sch 3 para 6, not SPC Sch I:

• Same qualifying benefits, with small differences in the Scottish and Welsh lists:
  • Scotland's partner limb, para 7(2)(b)(ii), also lists child disability payment care.
  • Wales lists no Scottish benefits and no carer support payment.
• Same residence test. No non-dependant aged 18 or over, ignoring recipients of those benefits and blind people. England and Scotland exclude only members of the applicant's own family from "non-dependant", so an 18- or 19-year-old young person of another family counts. SPC Sch I para 2(2)(f) ignores that person. This is the same divergence as HB; see the Give pension-age HB and CTR the severe disability premium Pension Credit uses #1938 comment.
• Same blind-partner rule. A blind partner who does not qualify is treated as not a partner, so the applicant must be the one who qualifies. This is lawful here if the couple nominate that partner as applicant.
• Same carer limbs. Scotland's para 7(2)(a)(iii) says "in receipt of" without "entitled to and".
• Same amounts.

So the CTR premium agrees with this PR's Pension Credit addition wherever HB does. The HB-specific residence test proposed for #1938 would serve CTR too.

The law texts are saved with the evidence for this PR.

[MaxGhenis]

Independent review (at bde8f7d): REQUEST_CHANGES. Response, all addressed in 214a2fa. The branch is rebased onto #1896's 143cc3c.

The review found no law-to-code error for any Enhanced FRS household. It found:

• 0 mismatches against the reference over about 49k fuzzed benefit units;
• 0 mismatches against a separate bipartite-matching check over about 12k units;
• the reported pair-B impact reproduced exactly.

The findings and what changed:

1. Negative addition from supplied inputs. Three members flagged as claimant or partner, all carers, gave −£4,474.60.
   • Fixed: the couple branch is now max_(2 - cared_for, 0).
   • New YAML case: three flagged members, all on AA and CA, give 0.
2. No test at the age-18 boundary. New YAML case: an 18-year-old not in education counts as a resident. The age > mutant is now caught.
3. The overlap between the residence test and carer attribution. The body now says:
   • the residence condition alone accounts for the whole −£63.2m;
   • carer attribution alone would lower Pension Credit by £16.6m for 4.5k benefit units (a real run with the residence condition forced true, on the final code);
   • all of those units also fail the residence test.
4. Replace generic child and adult flags with each programme's legal definitions #1896 was failing a test that this PR fixed. Replace generic child and adult flags with each programme's legal definitions #1896's own 143cc3c now corrects that test, so Replace generic child and adult flags with each programme's legal definitions #1896 passes by itself. This PR takes Replace generic child and adult flags with each programme's legal definitions #1896's version of test_carer_support_payment.py and no longer touches it.
5. A carer whose own disabled child qualifies. Within their unit, a carer now cares first for a qualifying claimant or partner, then for another member who receives a qualifying benefit, such as a child on DLA care at the middle rate. New YAML case: the grandmother keeps £4,474.60.
6. A surplus carer in a unit disappeared. Each person has one carer (s.70(7ZA)), so carers left over in a unit now join the household pool. New YAML case: the result is the same whichever unit the second carer is in.
7. Tests.
   • Ties: the tie rule (the unit whose eldest member is listed first) is now in the variable documentation and has a YAML case. The tie_last mutant is now caught.
   • Oracle independence: the reference's carer rule is now explicit person-to-person matching: brute force within a unit, and the bipartite feasibility check this review used across units. It no longer reuses the implementation's counting.
   • Generator: every benefit unit may now be a couple with dependants.
   • Fuzzing: the review's multi-unit fuzzer, run against the new reference with claimant flags both supplied and derived, gives 0 mismatches over about 30k benefit units.
8. Wording.
   • The Give pension-age HB and CTR the severe disability premium Pension Credit uses #1938 +£1.5m sentence now says this PR's formula replaces that change.
   • Test counts come from CI.
   • The CSP test name issue is gone, because the file is now Replace generic child and adult flags with each programme's legal definitions #1896's.

Impact is unchanged to £0.1m on the final code. Against 143cc3c in 2026:

• Pension Credit −£63.2m;
• CTR £0.0m (Housing Benefit and household net income are not given: the HB change rests on fewer than 10 survey records);
• 16.6k benefit units lose and none gain.

[MaxGhenis]

Independent re-review at 214a2fa: APPROVE. It was done by an Opus agent that did not write this code.

• Within-unit carer attribution is a maximum matching. The proof is by Hall's theorem: each carer is excluded only from themselves. Each claimant or partner is covered first, with no loss of placements, and the count of leftover carers is exact. The exhaustive check over 1,700 unit configurations found 0 mismatches.
• The reference's matching-based carer rule is independent of the implementation and correct: 0 mismatches against the reviewer's brute force over 7,140 configurations.
• Fuzzing: 0 mismatches over 17,966 units (bipartite-matching check) and 23,912 units (reference).
• Round-1 findings: all eight are resolved. Eight of nine deliberate code breakages are caught by the tests.
• Impact: pair B (2026) and the carers-only counterfactual (Pension Credit −£16.58m, 4.55k units, all of which also fail the residence test) reproduce exactly from the saved runs.

The remaining non-blocking items are fixed in 228de58:

• a YAML case pins self-exclusion for a qualifying dependant carer;
• the tie order is now documented as person order;
• the surplus-carer case is tested in both placements;
• the reference docstring is fixed;
• the body now cites the CI counts.

Disclosed modelling choice: a carer with both a qualifying claimant or partner and a disabled child in their unit is attributed to the claimant or partner first.

[vahid-ahmadi]

Reviewed: approve with nits.

The residence and carer limbs now match SPC Regs Sch I para 1(1)(a)-(c) and para 2. The para 2(2)(a), (b) and (f) exceptions, the 18+ threshold and the reg 6(5) single and double rates are encoded correctly. Self-exclusion and the one-carer-per-award rule (s.70(7), (7ZA)) are also right.

1. Nit: num_severe_disability_addition_qualifiers_cared_for.py:27-32. The docstring says a UC carer in the household "already stops the addition through the residence condition". That fails when the UC carer's presence is ignored under para 2(2)(a) or (b), for example a resident on PIP daily living who gets the UC carer element for caring for the claimant. That person is neither counted nor attributed, so the addition is paid where para 1(1)(a)(iii) bars it. Please say so in the docstring, and optionally add a YAML case pinning the gap.
2. Nit: relevant_benefits.yaml has no Scottish qualifying benefits (pension age disability payment, ADP daily living, Scottish adult DLA), though every limb of para 1(1) and 2(2)(a) lists them. In Scotland a resident on ADP is counted when the law ignores them, and a claimant on PADP does not qualify. The chip is filed; please link it from the variable docs.
3. Nit: num_severe_disability_addition_qualifiers_cared_for.py:72-76 and :80-81 work out the eldest-member representative twice. Compute age and representative once. (Use one carer attribution for the legacy severe disability premium and the Pension Credit addition (stacked on #1946 and #1951) #1977 deletes this file, so this only matters if Use one carer attribution for the legacy severe disability premium and the Pension Credit addition (stacked on #1946 and #1951) #1977 stalls.)
4. Nit: is_blind is not a data input. The docstring could add that this can only withhold the addition (blind residents are counted, and the blind-partner route never fires); it never overpays.

Verified: I read SPC Regs 2002 reg 6(5), Sch I paras 1-2 and reg 4A on legislation.gov.uk. Locally, the whole pension_credit YAML directory passes (113) and so do the 4 Hypothesis properties. I also ran five household probes (single pensioner on AA with an adult son: plain; on PIP daily living; on PIP daily living plus CA; aged 17; aged 17 plus CA), and each matched the law. There are no PR checks because the base is #1896's branch, but the workflow_dispatch run 36805257533 on 228de58 passed all six jobs.

[MaxGhenis]

Reporter-loss check (redacted on 3 October: the detailed breakdown that was here rested on fewer than 10 survey records in several cells, so it has been removed from this public comment).

The Pension Credit reporter-gap work (which produced #2018) found that this PR raises the number of GB Pension Credit reporters with no modelled entitlement, on both survey and calibrated weights. This comment checked whether those losses come from the regulations or from exemptions the model misses.

The checks covered the residence test, the carer attribution compared with the FRS care table (who cares for whom), and each exemption in SPC Regs 2002 Sch I para 2 that the FRS can observe. Some exemptions cannot be observed in the FRS: the para 3(3) date, carers engaged by a charity, and new carers.

Conclusion. Under those checks, the extra non-entitled reporters are what the regulations produce, and this PR's logic needs no change. Most of them stay entitled once Pension Credit uses the FRS's own capital (#2018 with policyengine-uk-data#513). Scottish disability benefits (PADP, Scottish adult DLA, ADP daily living) are not modelled, and FRS ADP codes are unmapped (policyengine-uk-data#500).

[MaxGhenis]

I revised the reporter-loss comment above after an independent review. The conclusion is unchanged: this PR's logic needs no change. The changes:

• Corrected figures. The losses are now unrounded differences. They rest on fewer than 10 survey records, so they are not repeated here.
• Capital transition. A direct test now checks the affected records' entitlement under FRS capital.
• Carer test. The carer check now tests the changed allocations directly against the FRS care table. All agree in the affected entitled groups, and almost all agree across GB.
• Scottish benefits. PADP, Scottish adult DLA and ADP daily living are not modelled, and FRS ADP is unmapped (policyengine-uk-data#500). This doesn't change the affected entitled groups.
• Exemption table. It now gives each provision's full conditions and treats the FRS fields as candidates, not established exemptions.
• Rate caveat. The model's 2025-26 addition is a CPI projection (£83.25 against the published £82.90; Set DWP benefit rates to the announced amounts, not CPI projections #1925 fixes it).
• DWP comparison. It is now descriptive only.

[MaxGhenis]

Merged under Max's PE-UK merge-on-gates rule (decision d740): CI green on the merged head, an independent review approving that exact head, impact stated from a real microsimulation run, stack order respected.

• Independent review of 5008b53b: subfleet 20261002-230114-mt-review-1951e found the own diff byte-identical to the previously approved one (Opus re-review at 214a2fad; vahid-ahmadi "approve with nits" at 228de587) with the three merges of Replace generic child and adult flags with each programme's legal definitions #1896 and main clean, and all local tests passing; its findings, and those of two follow-up checks, were on the description only. Final sign-off: subfleet 20261003-142810-mt-signoff-1951 APPROVE at 5008b53b.
• CI: all checks pass on 5008b53b.
• Impact: real runs on the Enhanced FRS 2024-25 (1.57.4), main 3525263f against 5008b53b, as stated in the description (Pension Credit −£59.5m to −£65.6m a year, 2025-28).
• Stack: Use one carer attribution for the legacy severe disability premium and the Pension Credit addition (stacked on #1946 and #1951) #1977 (shared carer attribution) builds on this PR and Give the legacy severe disability premium its statutory conditions (stacked on #1896) #1946.