Fix benefit-cap statutory-pay earnings exemption by MaxGhenis · Pull Request #2399 · PolicyEngine/policyengine-uk

MaxGhenis · 2026-10-13T17:14:11Z

Raising statutory sick pay could end the benefit-cap earnings exemption because it used a separate calculation that omitted statutory pay while deducting its tax and NI. The exemption now sums the assessed claimants' existing actual UC net earnings before the minimum income floor. This addresses the benefit-cap portion of #2370 and fixes item 1 of #1986 apart from its earnings threshold.

The release family is pinned in a plain regression and an explicit Hypothesis example. The monotonic property is retained, and statutory pay is compared with equivalent wages after the cap as well as before it. Policy cases cover all three modelled statutory payments, pension tax, relievable contributions, the MIF exclusion, and basic income with and without its means-test toggle (the award is set above the couple cap, so both cases bite). The dividend-tax earnings-exception regression now runs as a normal test, with its obsolete strict xfail and outcome-allowlist entry removed. The private-pension case identifies the lowest-slice tax allocation as the model's convention, as documented in uc_income_tax_on_earnings, rather than wording prescribed by reg. 55(5)(b).

Law: UC Regulations 2013 reg. 82(1)(a) and (4) test actual claimant/couple earned income and exclude earnings imputed by the minimum income floor. Reg. 55(4)–(5) includes statutory pay and permits employment-specific deductions. The existing £10,152 annual threshold is retained in this release fix. #1820 uses the same expression, benunit.sum(uc_individual_earned_income_before_mif * is_uc_assessed_claimant), plus a dated National Living Wage threshold parameter; its version of is_benefit_cap_exempt_earnings.py supersedes this file when it merges.

Scope: reusing UC's earned-income variable changes the old measure in each of these ways, including current main's #1965:

Statutory sick, maternity and paternity pay count as earnings.
Only income tax and NI on earnings are deducted; tax on unearned income is excluded.
The person's own relievable pension contributions are deducted.
taxable_miscellaneous_income counts as earnings; its classification is the open question in UC counts royalties, trust income and overseas pensions in miscellaneous_income as earned income #1961.
A trading loss is floored at nil instead of reducing employed earnings.
Each person's net earned income is floored at nil before summing, so one partner's excess deductions cannot reduce the other's earnings.
Voluntary Class 3 NI is not deducted.
Reg. 77 company income counts through the UC earned-income variable since Model Universal Credit reg 77 for owner-managers of a company #1965.
Basic income counts through the UC earned-income variable when gov.contrib.ubi_center.basic_income.interactions.include_in_means_tests is enabled. Main's formula never counted it, so this is new behaviour under that reform toggle (default off); see Open questions below.

History: the pinned family already loses its earnings exemption at the 2.127.0 version commit e01f2f5. #2119 did not change the cap earnings formula. Commit 70851ab fixed the UC earnings variable through #2380 but left the cap's duplicated calculation behind.

CI gap: both source and installed-wheel modes collect the property with locked Hypothesis 6.169.0. Hypothesis's function_digest hashes cleaned function source and signature, not the module name or file path. However, internal/conjecture/providers.py::_get_local_constants mines literals from imported local modules, and internal/constants_ast.py::is_local_module_file explicitly excludes /site-packages/. Source model modules therefore supply a different constant pool from installed-wheel modules, changing integer draws despite the same derandomized seed. Source #2394 collects and passes the property; release 38366204407 fails it. Separately, #2119's original green PR run tested a synthetic merge based on e01f2f5, which lacked the property. The explicit example covers the release family in both modes regardless of generated inputs.

Validation:

Source, in the locked environment (Python 3.13.9, core 3.33.2, Hypothesis 6.169.0, Ruff 0.17.0):
- both full property files passed, 14 tests; the dividend earnings-exception regression now passes as a normal test;
- the benefit-cap and UC policy folders and their sibling YAML files passed, 589 policy tests;
- at the final head c0f501b5e, the earnings-exception YAML file passed 8 cases, including the basic-income case and its toggle-off control (run by the hub; log kept with the hub's review files).
From the installed wheel, the release's mode (uv build with .github/build-constraints.txt; installed with --no-deps into a scratch venv synced with --frozen --extra dev; run from an empty directory with the CI Hypothesis profile and strict xfail handling; installed files checked byte-identical to 05610e4d7; the later commits change only the earnings-exception YAML file, one comment in a test and the CI allowlist, so these results carry over):
- test_uc_statutory_pay_properties.py and test_uc_income_from_capital_properties.py: 14 passed;
- test_uc_company_owner_properties.py, test_uc_claimant_income_properties.py, test_benefit_cap_and_ctr_pension_age_properties.py and test_benefit_cap_specified_benefits_properties.py: 42 passed.
A wider source run of eight other search-matched Python files was interrupted by a host freeze at 36 of 122. CI runs the full suite.
ruff format --check and ruff check pass.
CI's release jobs at c0f501b5e (run 38479445636) ran the whole suite from the wheel: 5,537 passed, 69 skipped, 1 xfailed. Release test / Result still failed on one more release blocker, which main has had since 10/12. test_variable_code_sharing.py::test_a_simulation_leaves_coverage_no_new_code_objects skips in wheel mode ("coverage is not measuring with sys.monitoring") and had no allowlist entry. The statutory-pay failure hid it on main's release runs 38306431347 and 38366204407. The final commit, 8b997692f, adds a wheel-only allowlist entry limited to that skip reason; the test still runs and passes in source mode. Replaying run 38479445636's outcomes through check_test_shards.py with the entry gives: every collected test ran in exactly one shard and passed, apart from allowlisted outcomes.

Impact (measured by the UK hub, 2026-10-13): a real Microsimulation pair for 2026 on enhanced_frs_2024_25.h5 from uk-data 1.58.0 (sha256 5ba399ea5ab9…), with stored weights. The branch is 05610e4d7, which includes main a98f8f0 and so #1965. The later commits c0f501b5e and 8b997692f change only tests and the CI allowlist. The baseline is the same code with main's version of is_benefit_cap_exempt_earnings swapped in, which is this PR's only model change. A first run at e6564306b, before the main merge, gave the same figures to the pound.

2026	Baseline	This PR	Change
Universal Credit	£83,569.0m	£83,577.9m	+£8.9m
Benefit cap reductions	£1,156.0m	£1,147.1m	−£8.9m
Housing Benefit	£8,456.2m	£8,456.2m	£0

About 1.7k households gain (2 survey records). No household's Universal Credit or Housing Benefit falls.
The exemption flag moves for many more households than the cash does, because most of them are not capped. About 400k households gain an earnings-exempt benefit unit and about 9k lose their only one, with no change in their awards. The run does not break these down by cause.

Open questions:

Basic income: enabling gov.contrib.ubi_center.basic_income.interactions.include_in_means_tests makes basic income count toward this earnings exception through the existing UC earned-income variable. The default is false. Main's formula never counted basic income, so this PR introduces it under the toggle, and two YAML cases pin it. Pin benefit cap exemption rules with regression tests and legal references (#1818) #1820 uses the same expression and should settle its treatment alongside the remaining threshold part of Benefit cap, targeted childcare and local CTR use the wrong UC earned income figure #1986 item 1.
Miscellaneous income classification remains UC counts royalties, trust income and overseas pensions in miscellaneous_income as earned income #1961. The release fix follows the existing UC variable's treatment.
Reg. 77 company income: reg. 77(1) applies "for the purposes of this Part" (Part 6), while reg. 82 sits in Part 7. Counting it here follows reg. 82(4)'s reliance on the Part 6 figure and Model Universal Credit reg 77 for owner-managers of a company #1965's own note that the cap test should count company earnings. The dataset has no owner-manager inputs yet, so the impact is nil today.
Relieve trading losses by each programme's rule #2085 (trading losses) rewrites this file to the same expression. Whichever lands second takes the other's version, and Relieve trading losses by each programme's rule #2085's impact against main then excludes this change.

Reproducer: policyengine_uk/tests/test_uc_statutory_pay_properties.py::test_statutory_sick_pay_preserves_benefit_cap_earnings_exemption

axiom: TheAxiomFoundation/rulespec-uk#387 queued

That existing pe-parity issue queues reg. 55 statutory pay and earnings deductions; the pinned corpus lacks reg. 55, so the signed encoder is blocked. The existing reg. 82 module reads actual earnings excluding the MIF; cap-composition companion cases are added to the same queue. No billed encoder run or in-job review was started.

🤖 Generated with Claude Code

MaxGhenis · 2026-10-13T17:18:23Z

Handoff at e656430 on fix/release-statutory-pay-cap-exemption; pushed once.

The cap now uses actual UC net earnings before the MIF. The release family is pinned in a plain regression and @example; six YAML cases cover the permitted earnings/deductions. make format, locked Ruff format check and lint pass.

Local model validation is incomplete. The sandbox rejected nice's setpriority; the source probe was interrupted (130). Disk subsequently fell to 44.847 GB, below the job's 45 GB stop floor. No further tests ran. The wheel check is left to release CI below the 60 GB guard, and national 2026 impact is left to the hub below the 50 GB microsimulation guard (compressor 11.839 GB). A real enhanced-FRS file and sequential comparison helper are ready in .job/.

Outstanding: pinned regression, full property file, remaining targeted UC/cap Python and YAML tests, controlled constant-pool comparison, microsim impact and the hub's independent review. No in-job review or merge was performed. The PR body includes cause, law, source/wheel local-constant mining and the additional stale-base gap in #2119's original check. Axiom composition cases are queued in rulespec-uk#387.

.job/HANDOFF.md and .job/handoffs.jsonl are written. Git metadata had to stay workspace-local because the sandbox protects the shared .git; .job/result.bundle and the pushed PR head preserve the commit. The handoff documents the required GIT_DIR/GIT_WORK_TREE exports.

MaxGhenis · 2026-10-13T17:59:23Z

UK hub independent review (in-session Opus 5.5; the Codex and Claude lanes were closed), head e6564306b. A round-2 fix job is running for these findings.

VERDICT: REQUEST_CHANGES
Head reviewed: e656430

(UK hub in-session independent review, Opus 5.5, 2026-10-13 ~18:00Z. Used because every Codex lane was held and the Claude lane was closed.)

I ran no tests, simulations or builds. Everything below comes from reading code, hand arithmetic, legislation.gov.uk, and the logs of the PR's own CI run 38437021879, which finished during this review.

Findings

[blocker] A strict xfail pinned to the bug this PR fixes now passes, so CI is red and the release would stay red.
- Where: policyengine_uk/tests/test_uc_income_from_capital_properties.py:326-332, test_tax_on_dividends_does_not_remove_the_benefit_cap_earnings_exception, marked xfail(strict=True) for Benefit cap, targeted childcare and local CTR use the wrong UC earned income figure #1986.
- CI at this head: Test / Python 2/4 failed with [XPASS(strict)] PolicyEngine/policyengine-uk#1986 (1 failed, 1475 passed, 16 skipped). Test / Result failed. Shards 1, 3, 4 and YAML passed.
- Why it flips: the case is earnings £11,000 plus dividends £10,000. The old measure was 11,000 − 7,930 × 10.75% = 10,147.52, under 10,152. The new measure is 11,000, so the exemption holds with and without dividends.
- The release job runs the same test from the wheel, so merging as is swaps one release failure for another.
- Fix needed:
  - remove the marker;
  - remove the allowlist entry at .github/test_outcome_allowlist.toml:28-32;
  - update the stale docstring (lines 35-39) and comment (line 256) in that test file.
- Pin benefit cap exemption rules with regression tests and legal references (#1818) #1820 already makes the marker and docstring edit, but not the allowlist one.
[should-fix] The PR description misstates scope in three places.
- "The formulas differ in three ways" is incomplete. The new measure also:
  - counts taxable_miscellaneous_income (uc_individual_earned_income_before_mif.py:56, open question UC counts royalties, trust income and overseas pensions in miscellaneous_income as earned income #1961);
  - floors a trading loss at nil instead of netting it off wages;
  - floors each person at nil, so one partner's tax on unearned income no longer offsets the other's earnings;
  - drops Class 3 NI;
  - counts reg 77 company income, on current main after Model Universal Credit reg 77 for owner-managers of a company #1965.
- It calls Benefit cap, targeted childcare and local CTR use the wrong UC earned income figure #1986 a "separate uprating issue". Item 1 of Benefit cap, targeted childcare and local CTR use the wrong UC earned income figure #1986 is this formula, and this PR fixes it apart from the threshold.
- It does not mention Pin benefit cap exemption rules with regression tests and legal references (#1818) #1820, which the hub note asked for.
[should-fix] Basic income now counts toward the cap's earnings exception under a reform toggle.
- Where: uc_individual_earned_income_before_mif.py:61-63, when gov.contrib.ubi_center.basic_income.interactions.include_in_means_tests is true (default false).
- Example by reading, not run: a couple with no earnings and £5,200 of basic income each totals 10,400, at least 10,152, so the cap lifts.
- The old formula never counted it. This is neither tested nor mentioned. Exclude it here or record it for Pin benefit cap exemption rules with regression tests and legal references (#1818) #1820, which has the same expression.
[should-fix] Nothing has run from a wheel since the fix.
- The release will be the first wheel-mode run of the two edited property tests.
- test_award_never_rises… keeps its seed, because Hypothesis strips decorators from the digest. The release therefore replays the same sequence, including examples after the one that failed, which have never been evaluated.
- test_statutory_pay_counts_as_wages… has a changed body, so it has a new seed in both modes; only source mode has run.
- By reading I found no in-strategy counterexample (check 4), but that is not execution.
[should-fix] The branch is 19 commits behind main.
- Its base f64155a predates Model Universal Credit reg 77 for owner-managers of a company #1965 (de51192), which rewrote uc_individual_earned_income_before_mif.
- CI tested the merge with a98f8f0, so the test results are valid.
- The impact figures were measured on the pre-Model Universal Credit reg 77 for owner-managers of a company #1965 tree. Merge main in with the fix for finding 1.
[nit] The hard-coded threshold is last year's. is_benefit_cap_exempt_earnings.py:24,36 keeps 10,152. For 2026, reg 82(1)(a) gives 16 × £12.71 × 52/12 = £881.23 a month (10,574.72 a year). This is pre-existing and left to Pin benefit cap exemption rules with regression tests and legal references (#1818) #1820, but the impact numbers inherit it.
[nit] One YAML comment overstates the law. benefit_cap_earnings_exception_earned_income.yaml:86-88 attributes "none is in respect of the employment" to reg 55(5)(b). The lowest-slice allocation is a model convention, as uc_income_tax_on_earnings.py itself says.
[nit] The retained "more pay never ends an exemption" assertion has a hole that predates this PR. By reading only, not executed:
- Family: a mixed-age couple, 62 (recipient, no wages, pension £6,000) and 68 (wages £6,000, pension £6,000), one child aged 1, childcare £6,000, rent 0, pay £0 → £300.
- At low pay: maximum 8,003.64 + 3,647.28 is under 12,000 of unearned income, so UC is nil and housing_benefit_pension_age_regulations_apply exempts.
- At high pay: the childcare element (5,100) opens, UC is positive and the exemption ends.
- The draw probability is negligible. It is not a reason to block.

Check 1: the law

Wording read on legislation.gov.uk (through a summarising fetch):
- Reg 82(1)(a) and 82(4) came back in near-verbatim fragments and match the PR's quotes.
- Reg 55(4)'s list and 55(5)(a)-(c), and reg 57(2)'s steps (a loss gives nil; no set-off against employed earnings), came back as summaries.
- Reg 52 was only partly quoted. Reg 78(2) was not fetched; I relied on the repo's citation.
The quantity is right. "Earned income" is the Part 6 figure after the reg 55(5) and 57(2) deductions. Reg 82(4) removes only reg 62 income, so before the MIF is correct. The work allowance belongs to reg 22, not to earned income, so before the work allowance is correct.
Summing over is_uc_assessed_claimant is right. It gives the claimant or the couple, includes the partner of a reg 3(3) single claim, and excludes dependants.
Included but arguable: miscellaneous income, basic income under the toggle, and reg 77 company income on main.
Omitted, all pre-existing in the UC variable: the other reg 55(4) payments, HMRC refunds, surplus earnings, reg 60 notional earnings, payroll giving, and the reg 82(1)(b) grace period.

Check 2: the falsifying family

Hand arithmetic with the repo's parameters: personal allowance 12,570, basic rate 20%, primary threshold 241.73 × 52 = 12,569.96, Class 1 rate 8%, WEEKS_IN_YEAR = 52.

Statutory sick pay	Old measure	Old exempt	New measure	New exempt
£0	12,570 − 1,200 = 11,370	yes	12,570 (NI 0.003)	yes
£6,000	12,570 − 2,400 − 480 = 9,690	no	18,570 − 1,200 − 480 = 16,890	yes

The old formula flips and the new one does not.
The prior attempt's executed output at e01f2f5 shows the same tax (1,200 / 2,400) and NI (0 / 480).
The regression test asserts 12,570 and 16,890 (tolerance 0.05), both exemption flags true, a nil cap reduction, and UC not rising. These are the right things, and it would fail on main at the high version.
The @example is well formed; explicit examples are not checked against min_size=4.
Both passed in CI, in shards 2 and 3.

Check 3: the six YAML cases

All six are right by hand, and CI's YAML job passed at this head. The single-claimant cap outside London is 14,753.

Sick pay. 3,000 + 8,000 = 11,000, under the allowance and the threshold, so no tax or NI. 11,000 ≥ 10,152: exempt, reduction 0. Right. (Old: 3,000.)
Paternity pay. Gross 11,000. The model does not tax it; the NI base of 11,000 gives nil. 11,000: exempt. Right.
Maternity pay. As case 1. Right.
Private pension. Income tax (21,000 − 12,570) × 20% = 1,686. Taxable earnings are max(0, 8,430 − 10,000) = 0, and NI is nil. 11,000: exempt. Right. (Old: 9,314.)
Pension contributions. 11,000 − 1,000 = 10,000. Tax is nil after relief; NI on 11,000 is nil. 10,000 < 10,152: not exempt; 18,000 − 14,753 = 3,247. Right.
Minimum income floor. Profit 5,000; tax nil; Class 2 rate 0; Class 4 limit 12,570. Before the floor 5,000, after it 20,000. Not exempt; 3,247. Right.

Check 4: the property file

The diff against main removes or weakens no assertion. There is no new skip, xfail, assume, health-check suppression or narrowed strategy.
Additions: one plain test, one @example, and two more variables in the wages-equivalence loop.
The extended equivalence holds inside the strategy, by reading:
- Sick and maternity pay enter taxable_employment_income and ni_class_1_income exactly as wages do, in exact float32 integer sums, so earned income is identical.
- No reported legacy awards exist, so WTC, CTC and working-age HB are nil.
- The LCWRA element and the pension-age flag do not read pay beyond the pre-cap award, which was already asserted equal.
Paternity pay and under-covered pension contributions are excluded by counts_as_wages. That filter also excludes the one childcare mismatch I found (wages 0, profit −4,000, pay ≤ 4,000).
It is not a general invariant outside the strategy (reported HB, carers).

Check 5: who moves

Only relievable pension contributions can remove an exemption, and reg 55(5)(a) requires that. Every other difference adds to the measure.
impact-comparison.json matches the description:
- UC +8,924,132; cap reductions −8,924,132; HB 0.
- 2 records, 1,693 weighted households gaining, none losing.
- 400,060 households gain an earnings-exempt unit; 8,818 lose their only one.
The file also shows 7,288 households losing every exempt unit overall, with no cash change.
Plausible, but two records is a thin base.
Unsupported claims are listed in finding 2.

Check 6: overlap with #1820

Pin benefit cap exemption rules with regression tests and legal references (#1818) #1820 at e9b20e5 uses the identical expression: benunit.sum(uc_individual_earned_income_before_mif * is_uc_assessed_claimant).
It differs only in the threshold (a parameter, 881 a month for 2026), label, comments and references.
Fix benefit-cap statutory-pay earnings exemption #2399 is a subset. The conflict in this file is resolved by taking Pin benefit cap exemption rules with regression tests and legal references (#1818) #1820's version. Fixing finding 1 will add a second trivial conflict in the capital-income test file.
The YAML cases at 11,000 and 10,000 still hold under Pin benefit cap exemption rules with regression tests and legal references (#1818) #1820's threshold.
Pin benefit cap exemption rules with regression tests and legal references (#1818) #1820's CHANGES_REQUESTED review was on an older head and asked for this expression.

Check 7: the CI-gap claim

Verified by reading the saved Hypothesis source:
- is_local_module_file rejects any path containing /site-packages/ (constants_ast.py:260-284).
- _get_local_constants mines only local modules (providers.py:325-380).
- _maybe_draw_constant spends PRNG draws that depend on that pool.
- function_digest hashes cleaned source and signature only.
Verified in the repo: wheel mode installs under $RUNNER_TEMP/venv/…/site-packages, and core 3.33.2 registers variable files in sys.modules.
Could not verify:
- scrutineer.py is missing from the saved tree, so I did not read the "source checkout counts as local" half.
- The tree has no version file; only the directory name and uv.lock say 6.169.0.
- I did not re-read the Install Lint and the documentation build from uv.lock #2394 or release logs.
Issue Release test fails: test_award_never_rises_with_a_claimants_statutory_pay passes from source, fails from the installed wheel #2397's comment reports the divergence empirically (constants mined from source, none from the wheel, examples differ from the third).

Check 8: anything else

Readers. Only is_benefit_cap_exempt reads the variable. No test embeds the old formula.
Cycles. None new. A cycle would need benefit_cap_reduction among the dependencies of the earned-income variable, and that would already be a cycle on main, because the pre-cap award reads it. Main's added dependencies from Model Universal Credit reg 77 for owner-managers of a company #1965 read inputs only.
Performance. No extra cost; UC already computes the variable.
Changelog. changelog.d/2397.fixed.md has the right name and type. The Fix evidence check passes.
PR state. It is still a draft.

- The basic-income case now enters an award of 25,000, above the couple cap outside London (22,020), so its zero cap reduction shows the exemption working. A control case without the reform toggle shows basic income not counting: no exemption, reduction 2,980. - Say that counting basic income in the earnings exception is new in this PR (main never counted it), not existing behaviour. - The capital-income property compares the capped award only where the exemption is unchanged because capital income can change another exemption route, not because of the allowance taper. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

The release workflow tests the installed wheel without coverage, so test_a_simulation_leaves_coverage_no_new_code_objects skips there ("coverage is not measuring with sys.monitoring"). It had no allowlist entry, so every wheel-mode Release test / Result failed on it, behind the statutory-pay failure this PR fixes. The entry is limited to wheel mode and to that skip reason. The test still runs and passes in source mode on every PR. Replaying this PR's own release run (38479445636) through check_test_shards.py with the entry: every collected test ran in exactly one shard and passed, apart from allowlisted outcomes. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

MaxGhenis · 2026-10-14T01:34:10Z

Merged by the UK hub (audit). Merged with a merge commit at the tested head 8b997692f (expectedHeadOid, through the main merge queue). This unblocks PE-UK releases, which had two blockers since 2026-10-12 (PyPI was stuck at 2.127.0):

every main release run failed test_award_never_rises_with_a_claimants_statutory_pay;
behind that, wheel mode skipped test_a_simulation_leaves_coverage_no_new_code_objects with no allowlist entry.

This PR fixes the first in the model and allowlists the second for wheel mode only.

Gates:

CI: gh pr checks exits 0 at the head, with 29 checks passing. That includes Release test / Result, which ran 5,607 tests from the built wheel: every one ran and passed apart from allowlisted outcomes.
MERGEABLE, not a draft.
Independent review: three in-session Opus 5.5 reviews, run that way because the GPT-6.1 Sol and Claude lanes were all held (disk and ramp holds; Fleet ops approved the fallback).
- e6564306b: REQUEST_CHANGES. Its blocker was a strict xfail for Benefit cap, targeted childcare and local CTR use the wrong UC earned income figure #1986 that the fix turned into an XPASS.
- The round-2 delta: REQUEST_CHANGES, for the PR body and wheel coverage.
- c0f501b5e: APPROVE (~/reviews/uk-hub/reviews/uk2399/review-final.md).
- The final head, 8b997692f (the allowlist entry): APPROVE (review-final-8b997692f.md). One nit, about a test comment, is deferred.
- Every round-1 and round-2 finding is fixed or recorded in the description.
Wheel-mode check (the release's failure mode), from the installed wheel, byte-identical to 05610e4d7:
- the two edited property files: 14 passed;
- the four other property files that read the cap: 42 passed.
Real-run impact: 2026, enhanced FRS from uk-data 1.58.0 (sha256 5ba399ea…), head against main's cap formula.
- UC +£8.92m and benefit cap reductions −£8.92m; HB unchanged.
- About 1.7k households gain, from 2 records; none lose.
- The figures are the same at e6564306b and at 05610e4d7, which includes Model Universal Credit reg 77 for owner-managers of a company #1965.
Law: UC Regs 2013 regs 82(1)(a) and 82(4), 55(4)–(5) and 57.
Follow-ups:
- Relieve trading losses by each programme's rule #2085 rewrites the same file to the same expression and resolves the trivial conflict after this lands.
- Pin benefit cap exemption rules with regression tests and legal references (#1818) #1820 supersedes the file with the NLW threshold parameter.
- Basic income and reg 77 scope are open questions in the description.

#2399 landed the same benefit-cap earnings exception change this branch made (is_benefit_cap_exempt_earnings reads uc_individual_earned_income_before_mif). The only conflict was in that file's comments; take main's version, so this PR no longer touches it. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

Resolve conflicts with #2399, #2076 and #1965: - is_benefit_cap_exempt_earnings: keep main's per-claimant uc_individual_earned_income_before_mif test and documentation; keep this PR's dated threshold parameter in place of 10_152 and its reg 57 and FA 2004 s.188 references. - is_benefit_cap_exempt_health_disability: take main's version (#2076) and add the UC route for AFIP: on a Universal Credit award only a claimant's counts; otherwise Housing Benefit's young person counts too. - is_benefit_cap_exempt: take main's label. - test_uc_single_claim_properties: keep main's AFCS assertions and this PR's partner-only AFIP example. Pin main's young-person AFIP YAML case to the Housing Benefit route, and give the specified-benefits property oracle the family's scheme. Rewrite the changelog fragment to what the PR still changes relative to main. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

…ted-benefits-scoped-awards

…shared-cap-lha-rules #2399 changed the shared earnings exception this PR deletes. Its measure, uc_individual_earned_income_before_mif summed over the assessed claimants (reg. 82(1)(a), excluding MIF income under reg. 82(4)), is carried over unchanged into the UC-only is_uc_benefit_cap_exempt_earnings; the shared file stays deleted. Main's new tests that read the deleted shared names now read the UC ones, with their expected values unchanged: - test_uc_statutory_pay_properties.py (#2399; conflict): main's text and regression case, reading is_uc_benefit_cap_exempt(_earnings). - benefit_cap_earnings_exception_earned_income.yaml (#2399) and uc_trading_loss.yaml (#2085). - test_trading_loss_properties.py (#2085). Targeted YAML (benefit_cap folder, benefit_cap_by_scheme, uc_trading_loss): 121 passed. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

Rebased onto main after #2399, which made the benefit cap earnings fix. Only the test-environment change remains: conftest stops Hypothesis drawing literal constants from policyengine_uk's own modules, and a guard test fails if a Hypothesis upgrade bypasses that. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

Picks up #2399 (8565474), which fixed test_award_never_rises_with_a_claimants_statutory_pay, the Release test failure on this branch. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

Fix benefit-cap earnings exemption for statutory pay

e656430

MaxGhenis and others added 4 commits October 13, 2026 14:40

Merge main into statutory-pay release fix

827bbe0

Remove resolved cap xfail and document income interactions

05610e4

MaxGhenis marked this pull request as ready for review October 14, 2026 01:04

MaxGhenis added this pull request to the merge queue Oct 14, 2026

Merged via the queue into main with commit 8565474 Oct 14, 2026
31 of 32 checks passed

MaxGhenis deleted the fix/release-statutory-pay-cap-exemption branch October 14, 2026 01:32

vahid-ahmadi mentioned this pull request Oct 14, 2026

Pin benefit cap exemption rules with regression tests and legal references (#1818) #1820

Merged

MaxGhenis added a commit that referenced this pull request Oct 14, 2026

Merge origin/main (dd64849, with #2085 and #2399) into disable-simula…

e6b3367

…ted-benefits-scoped-awards

MaxGhenis mentioned this pull request Oct 14, 2026

Fix disable_simulated_benefits and read one claimant-or-partner award #2075

Merged

juaristi22 mentioned this pull request Oct 14, 2026

UK: policyengine-uk 2.130.1 and the fourth batch of #1095 ports, with the release's weight year, identifiers and birth months fixed PolicyEngine/microcosm#1271

Merged

juaristi22 mentioned this pull request Oct 23, 2026

Count statutory pay in Income Support income #2595

Merged

This comment has been minimized.

Sign in to view

vahid-ahmadi mentioned this pull request Oct 23, 2026

Benefit cap: model caps about 1.3x DWP's count in 2026-27 (204k vs 160k in May 2026) #1818

Open

This comment has been minimized.

Sign in to view

Navigation Menu

Repository navigation

Fix benefit-cap statutory-pay earnings exemption - #2399

Fix benefit-cap statutory-pay earnings exemption#2399
MaxGhenis merged 5 commits into
mainfrom
fix/release-statutory-pay-cap-exemption

MaxGhenis commented Oct 13, 2026 •

edited

Loading

Uh oh!

MaxGhenis commented Oct 13, 2026

Uh oh!

MaxGhenis commented Oct 13, 2026

Uh oh!

Uh oh!

MaxGhenis commented Oct 14, 2026

Uh oh!

This comment has been minimized.

This comment has been minimized.

This comment has been minimized.

This comment has been minimized.

This comment has been minimized.

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

Conversation

MaxGhenis commented Oct 13, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

MaxGhenis commented Oct 13, 2026

Uh oh!

MaxGhenis commented Oct 13, 2026

Findings

Check 1: the law

Check 2: the falsifying family

Check 3: the six YAML cases

Check 4: the property file

Check 5: who moves

Check 6: overlap with #1820

Check 7: the CI-gap claim

Check 8: anything else

Uh oh!

Uh oh!

MaxGhenis commented Oct 14, 2026

Uh oh!

This comment has been minimized.

This comment has been minimized.

This comment has been minimized.

This comment has been minimized.

This comment has been minimized.

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

MaxGhenis commented Oct 13, 2026 •

edited

Loading