Narrow the legacy award readers property: an adult non-dependant ends the severe disability premium

[MaxGhenis]

Fixes #2293

Summary

test_legacy_award_readers_properties.py::test_other_members_awards_never_change_the_readers claimed that adding a benefit-unit member who is not the claimant, the partner or a dependant never changes a reader of the legacy income-related awards. That is false: an added adult aged 18 or over is a non-dependant, and a non-dependant ends the family's severe disability premium. The test passed on main only because its 15 derandomised examples did not draw the case, so any edit to the test could have surfaced it in an unrelated PR.

This PR changes the test only. The model's behaviour for the severe disability premium is right and is unchanged. The review of this PR found a second way an added member changes a reader, which is a model defect: the benefit cap counts the member's own awards. That is #2298; this PR holds the cap at zero in the one property until it is fixed (see "The benefit cap" below).

The counterexample

On main (9475c95), 2025-26: a claimant aged 18, not in education, disabled (pip_dl £2,000), with a carer benefit and £200 of reported Income Support, in England with no savings. The added member is an 18-year-old with no awards.

Variable	Without the added member	With
severe_disability_premium	£4,729.28	£0
income_support	£13,771.55	£9,042.27
council_tax_reduction_applicable_income (a family reader)	£25,074.71	£20,345.43

Each falls by the premium. No other family reader moves.

The law

IS Regs 1987 Sch. 2 para. 13(2)(a)(ii): a single claimant is severely disabled only if "he has no non-dependants aged 18 or over normally residing with him or with whom he is normally residing". The parameter gov.dwp.disability_premia.severe_non_dependant_age cites that paragraph and its Housing Benefit counterpart, HB Regs 2006 Sch. 3 para. 14(2)(a)(ii).

Change

The property now builds each family three ways (without the other member, with them reporting no award, with them as drawn) and checks two things that are true:

1. What the member reports never changes a reader. The family readers, the member readers and the premium are equal between "reporting nothing" and "as drawn". This is the claim in the test's name.
2. Their presence changes a reader only by ending the premium. Where the premium is the same with and without the member, every reader is the same. Where it differs, it has fallen to zero from a positive amount, and the member is at least severe_non_dependant_age (18). Even then, the family's income-related ESA and income-based JSA awards (claimant_or_partner_esa_income, claimant_or_partner_jsa_income) and whether a member is on one of them (is_on_income_related_esa, is_on_income_based_jsa) are the same: they read reports and capital only. Whether a member is on Income Support can change, when the lost premium takes the award to zero.
3. Income Support falls by at most the premium. It never rises, and its fall is never more than the premium lost. It can be less: an award is never negative, so a family whose Income Support is already below the premium loses only what it had, and an ineligible family loses nothing.

The counterexample is pinned three ways as @examples: the 18-year-old with no award, the same with £3,000 of Income Support, and a 17-year-old with £3,000 of income-related ESA. A fourth @example pins the review's benefit cap family (below).

A new fixed test, test_adult_non_dependant_ends_the_severe_disability_premium, states the rule on that family: a 17-year-old leaves everything unchanged; an 18-year-old ends the premium, income_support and council_tax_reduction_applicable_income fall by exactly the premium, and every other family reader is the same in all three.

Of the issue's three options this is the first and third together: the family's own award is held fixed for the awards claim, and the premium rule gets its own statement. The second (a generator that cannot draw a non-dependant) would have dropped most adult members from the property.

The benefit cap

benefit_cap_reduction sums jsa_income and esa_income, which are the whole benefit unit's awards, so an added member's own income-based JSA or income-related ESA raises the family's cap reduction, and Universal Credit and the council tax reduction income fall. On main, a couple on £3,000 of JSA each with two children and a lodger: an added member aged 20 reporting £3,000 of JSA takes the cap reduction from £0 to £1,047.52 and Universal Credit from £15,118.92 to £14,071.40. In law the cap applies to the welfare benefits "to which a single person or couple is entitled" (WRA 2012 s.96(1)). #2298 has the details and the fix.

The cap is not one of the readers this test is about, so the property holds benefit_cap_reduction at zero, with a comment naming #2298, and pins that family as an @example. When #2298 is fixed, removing the hold makes the example check the fix. One other property compares families with and without a member, the Merton non-dependant one; it fixes Universal Credit at zero, which closes this channel, so it needs no hold. The fixed severely disabled family is exempt from the cap.

Invariants

1. A member outside the claimant's family reports any Income Support, income-based JSA or income-related ESA: no reader changes, for the family or its existing members.
2. Adding such a member changes a reader only if it ends the family's severe disability premium, and never changes the family's ESA or JSA awards or whether a member is on one of them.
3. Only a member aged 18 or over ends it, and it ends in full. (IS Sch. 2 para. 13(3) ignores a non-dependant on a qualifying disability benefit, a blind one and some carers; the test's added members are none of those.)
4. Income Support falls by at most the premium lost, and never rises.

Tests

The new invariant was stressed before it replaced the old one. The old property passed for months by not drawing its counterexample, so the narrowed one was run on 600 random examples (four seeds of 150, not derandomised, the test's own strategies): all passed. That run did not draw the benefit cap family the review found, so it was evidence, not proof. After the review it was stressed again at 4380010, with the cap held and the new bounds, in a log that records the commit and directory (stress_2294_r1.log): 600 random examples, all passed.

Fail on the old claim. The pinned examples are the issue's counterexample: with the old assertion (readers equal with and without the member) the first one fails on council_tax_reduction_applicable_income, £25,074.71 against £20,345.43, as the issue's repro shows on main.

Run locally at 4380010 (the later commit changes the docstring and a comment only):

• pytest policyengine_uk/tests/test_legacy_award_readers_properties.py: 12 passed (the four pinned examples run first in the changed property).
• ruff format --check and ruff check on the file: clean.

The full suites run in CI.

Microsimulation impact

None: the PR changes one test file and no model code, so no simulation result can differ.

Note for #2279

#2279 edits other_members in the same file (it sets the 16th-birthday date fact false). The two changes touch different functions and merge cleanly; the pinned examples here are aged 17 and 18, so they do not depend on that fact.

axiom: n/a: test-only change, no policy is added, updated or fixed

🤖 Generated with Claude Code

[MaxGhenis]

Hand-off to the UK hub (from the QYP follow-ups session, local_1b6fef51-8f78-48d7-addf-36982ce71831).

• Head: 757e4061749f57e744a6ce391bf36bf5a0d7dfe5. Marked ready for review. Test-only: one test file and the changelog fragment.
• Independent review: GPT-6.1 Sol, three rounds. Round 3 APPROVE at 757e40617, no findings: ~/reviews/pe-uk-qyp-followups-2026-10-10/reviews/pr2294_review_r3.md. Rounds 1 and 2 requested changes (the benefit cap channel, the Income Support bound, then a wording overclaim), all fixed.
• Passed: Lint and the four install jobs at this head. The test code at this head is identical to 4380010bf (the last commit changes the docstring and a comment only; the reviewer confirmed the executable code is the same), where CI Test passed, the whole file passed locally (12) and 600 random examples of the changed property passed (stress_2294_r1.log, which records the commit).
• Pending: CI Test at this head.
• Depends on: nothing. It merges cleanly with Count 16-year-olds outside education as UC and Pension Credit qualifying young persons until the 1 September after their 16th birthday #2279.
• Decisions: none.
• Follow-ups: Benefit cap counts another member's own income-based JSA and income-related ESA #2298 (the benefit cap counts other members' own JSA and ESA, a model defect; when it is fixed, drop the CAP_HELD hold in this test so its @example checks the fix). Benefit cap counts another member's own income-based JSA and income-related ESA #2298 is out of this session's scope; route it.
• Merge method: merge commit.

[MaxGhenis]

UK hub: landed on the gates at head 757e4061749f57e744a6ce391bf36bf5a0d7dfe5 (merge commit).

• Checks: gh pr checks exit 0, every check passing, including Test at this head.
• Mergeable: MERGEABLE; not a draft; no CHANGES_REQUESTED review.
• Independent review: GPT-6.1 Sol round 3, APPROVE at this exact head with no findings (~/reviews/pe-uk-qyp-followups-2026-10-10/reviews/pr2294_review_r3.md). Rounds 1 and 2 requested changes, which were addressed.
• Impact: none on model outputs. Test-only (one test file plus the changelog fragment). It fixes the over-strong property reported in Property test claims an added non-dependant never changes the legacy award readers; the severe disability premium is a counterexample #2293: an adult non-dependant ends the severe disability premium.
• Follow-up: Benefit cap counts another member's own income-based JSA and income-related ESA #2298 (the benefit cap counts other members' own JSA and ESA). When that is fixed, drop this test's CAP_HELD hold so its @example checks the fix.