Test that AFCS payments never confer severe disability; document AFCS vs AFIP (stacked on #1946)

[MaxGhenis]

Fixes #2114

Follows #1946 (merged 2026-10-03). This PR carries no formula change: it is tests and documentation only.

Summary

#1946 fixes is_severely_disabled_for_benefits paragraph 5. It reads armed_forces_independence_payment instead of any afcs payment, as Child Tax Credit Regulations 2002 reg 8(5) and Working Tax Credit Regulations 2002 reg 17(4) require. This PR adds what #1946 does not have, as listed in the review on this PR:

• Documentation.
  • afcs_reported: what FRS benefit code 8 covers.
  • armed_forces_independence_payment: its legal basis, and that it is not part of afcs.
• YAML cases, appended to Give the legacy severe disability premium its statutory conditions (stacked on #1896) #1946's is_severely_disabled_for_benefits.yaml. They check that AFCS payments confer nothing, alone or next to an award that is not severe, through every amount that reads the flag.
• Hypothesis properties for the AFCS rule across those amounts.

The first version of this PR made the formula change itself on main; it was consolidated into #1946 instead, which has now merged.

What afcs holds

afcs adds afcs_reported, which policyengine-uk-data builds from FRS benefit code 8:

• The FRS 2024-25 data dictionary (UKDA study 9563) labels code 8 "Armed Forces Compensation Scheme".
• The questionnaire (card H3) asks about the "Armed Forces Compensation Scheme (formerly War Disablement Pension), including Guaranteed Income Payments".
• The code frame has no armed forces independence payment (AFIP) code.

AFIP is one of six injury benefits in the AFCS Order 2011 art 15(1): lump sum, supplementary award, guaranteed income payment, AFIP, fast payment and medical expenses. CTC reg 8(5) and WTC reg 17(4) name only AFIP. UC Regs 2013 reg 24(2)(b) names no armed forces payment.

Tests

Appended to tests/policy/baseline/finance/benefit/family/is_severely_disabled_for_benefits.yaml: 7 cases, computed by hand from the law and the published rates.

• DLA and PIP. An AFCS payment alongside DLA care at the middle rate, and alongside PIP daily living at the standard rate: not severe. Either award still makes a child disabled under CTC reg 8(2), but neither is a severe disability paragraph.
• CTC. A child with an AFCS payment adds no severely disabled child element.
• WTC.
  • An AFCS payment adds no severe disability element.
  • AFIP at the 2024-25 rate adds it. That rate is £184.30 a week (AFCS Order 2011 art 24A(3), as amended by SI 2024/347 art 4), so £9,583.60 a year, and the element is £1,705 (WTC Regs Sch 2).
• UC. A 17-year-old qualifying young person with an AFCS payment gets no higher disabled child addition. On the enhanced PIP daily living rate they get it: £495.87 × 12 = £5,950.44 in 2025-26.

The file passes in full on this head: 21 of 21 (#1946's 14 cases and these 7).

New tests/test_severe_disability_afcs_properties.py. Each example is a population of 10 to 20 random families: adults, and dependants aged 0 to 18 who are in non-advanced education from 16. Every dependant is a child or qualifying young person for UC and CTC. Each person gets random DLA care, PIP daily living and Attendance Allowance categories, AFCS payments and AFIP.

1. Differential. The flag equals the tax credit condition recomputed from the inputs: CTC reg 8(3)-(5) plus WTC reg 17(2)'s higher-rate Attendance Allowance.
2. Metamorphic. Setting every AFCS payment to zero, or to another random amount, changes neither the flag nor the amounts that read it: the UC higher disabled child addition, the CTC severely disabled child element and the WTC severe disability element. The legacy severe disability premium reads its own qualifying-benefit list after Give the legacy severe disability premium its statutory conditions (stacked on #1896) #1946, not this flag. It gets the same AFCS check as an independent case.
3. Monotone. Giving everyone AFIP turns the flag on and never off.

Results on this head:

• All 3 properties pass.
• Against the formula on main before Give the legacy severe disability premium its statutory conditions (stacked on #1896) #1946, all 3 fail.
• The examples are not vacuous. A separate probe drew 10 examples from the same population strategy (333 people): 33 people have an AFCS payment and no severe-qualifying benefit, and each of the four amounts is positive for 68 to 111 people or benefit units.

Effect

There are no formula or parameter changes, so model output is unchanged.

For context on #1946's AFCS change, here are real Microsimulation runs on private copies of the Enhanced FRS 2024-25 (2024-25 to 2030-31, weights unchanged):

• Formula alone: no change in any year. The dataset stores is_severely_disabled_for_benefits, and build_from_multi_year_dataset sets every stored column as an input for every year, so the stored flag wins. Main 7b9fc379 against a branch with only the AFIP formula, on the same copy, gave every reported aggregate identical: benefits, premiums, net income and poverty.
• Stored flag also corrected for code 8 only (one column changed; the rebuild matches the stored flag on all 113,403 records without a code 8 payment):
  • The flag drops by 130k people in 2025-26 (from 4.14m to 4.01m). No one under 20 reports a code 8 payment.
  • Housing Benefit is the only payable benefit that moves: less than £0.1m in every year (−£27k in 2024-25, −£36k in 2030-31). Fewer than 10 survey records are affected, so there is no further breakdown.
  • The model's severe_disability_premium is computed for every benefit unit, whether or not it claims a legacy benefit. Its notional total falls by £1.1bn in 2025-26. That is not spending: only the HB, IS and CTR applicable amounts read it, and only HB moves.
  • Runs on main bfc5fea1 gave the same figures. On that baseline, policyengine-uk-data#494's full stored-flag definition lowers net income by £7.9m in 2025-26 (CTR £7.0m, HB £0.9m), so code 8 is under 0.5% of it.

Not changed here

• UC reg 24(2)(b) has no AFIP limb and adds blindness: UC higher disabled child addition: add the blindness limb and stop keying it on the tax credit severe flag #1970.
• AFIP has no FRS code, so it is zero in dataset runs: a known under-count.
• War pensions constant attendance allowance. For the severe disability premium, the IS and HB Regulations' "attendance allowance" includes "any payment based on need for attendance which is paid as part of a war disablement pension" (IS Regs 1987 reg 2(1); HB Regs 2006 reg 2(1)). Code 8 cannot separate it.

Axiom

axiom: n/a: tests and documentation only; the provision changes in #1946. For #1946's change, CTC Regs 2002 reg 8 and the WTC reg 17(2)-(6) qualifying benefits are queued as an encoding addendum on TheAxiomFoundation/rulespec-uk#361 (TheAxiomFoundation/rulespec-uk#361 (comment)). uk:regulations/uksi/2013/376/24 is already correct: higher_rate_disabled_child_addition_applies has no armed forces input.

🤖 Generated with Claude Code

[vahid-ahmadi]

Review pass (Claude Code, on Vahid's behalf), on closing this in favour of #1946.

#1946 already contains this PR's fix. At 9dd45148, is_severely_disabled_for_benefits reads armed_forces_independence_payment > 0 for paragraph 5 instead of afcs > 0, the same change as here at 296934cb. #1946 also adds Attendance Allowance at the higher rate, from WTC Regs 2002 reg 17(2).

What #1946 doesn't have yet, and should carry over before this closes:

• tests/test_severe_disability_afcs_properties.py: the metamorphic property that changing any AFCS payment leaves the UC higher disabled child addition, the CTC severely disabled child element, the WTC severe disability element and the severe disability premium unchanged.
  • Its differential invariant recomputes reg 8(3)–(5) without Attendance Allowance. That holds today because the generator never sets aa_category, but on Give the legacy severe disability premium its statutory conditions (stacked on #1896) #1946 the docstring should say so, or the recomputation should add the reg 17(2) limb.
• The YAML cases that pair an AFCS payment with DLA care at the middle rate or PIP daily living at the standard rate, and the UC higher disabled child addition cases.
• The afcs_reported and armed_forces_independence_payment documentation and references: FRS benefit code 8 is AFCS plus war disablement pensions, and AFIP is not part of afcs.

Timing. #1946 is a draft stacked on #1896, so closing this delays the AFIP fix until both land. If #1896 isn't close to merging, I'd merge this first and rebase #1946 over it; otherwise consolidating is fine.

[MaxGhenis]

Agreed: consolidated rather than raced. #1896 has merged and #1946 is next in the train. I kept the formula change out of this PR so #1946 stays on its tested head.

This PR is now stacked on #1946 (abef7e9a, on #1946's 5ab03ee0). It carries only what you listed:

• Properties. test_severe_disability_afcs_properties.py, with the differential widened to Give the legacy severe disability premium its statutory conditions (stacked on #1896) #1946's full definition. The generator now draws aa_category, and the reference adds WTC reg 17(2)'s higher-rate Attendance Allowance.
• YAML cases. AFCS alongside DLA care middle and alongside PIP daily living standard; the CTC child element; the WTC element with AFCS and with AFIP; the UC higher-rate cases. They are appended to Give the legacy severe disability premium its statutory conditions (stacked on #1896) #1946's YAML file and leave out its own cases.
• Documentation for afcs_reported and armed_forces_independence_payment.

On this head the YAML file passes 16 of 16, the 3 properties pass, and the code-health tests pass. Against pre-#1946 main, all 3 properties fail. Once #1946 merges I'll retarget this to main.

🤖 Generated with Claude Code

[MaxGhenis]

Merged at the reviewed head 2fd8ea19c with a merge commit (--match-head-commit). These were the gates:

• CI: gh pr checks exit 0 on this head (Test 29m49s; Lint and the smoke imports pass). GitHub reported the PR as MERGEABLE.
• Independent reviews (Subfleet, review task):
  • Review 2 (20261002-095538-afcs-sev-review2) returned REQUEST_CHANGES with 12 findings; all were addressed in 0b718f844.
  • Review 4 (20261003-161533-afcs-sev-review4, hard tier, at 2fd8ea19c) marked 10 resolved and 2 partly resolved. The partly resolved items and its new findings concerned the PR body and the two outward comments only, not the code.
  • Review 5 (20261003-180537-afcs-sev-review5) approved those text fixes.
  • Reviews 1 and 3 were lost without output.
• Scope: tests and documentation only, with no formula or parameter change, so model output is unchanged. The AFCS-to-AFIP formula fix itself landed in Give the legacy severe disability premium its statutory conditions (stacked on #1896) #1946.
• Outward follow-ups posted:
  • the encoding addendum on Encode TCA 2002 ss.3 and 8, CTC Regs regs 2, 3 and 5, and WTC Regs regs 4, 9, 11, 12, 14 and 17 (claims, child and QYP, WTC elements) TheAxiomFoundation/rulespec-uk#361;
  • the code 8 decomposition on Store the tax credit severe disability flag, not a mix of premium tests policyengine-uk-data#494.

🤖 Generated with Claude Code