[juaristi22]

Fixes #1799

Reduce the savings starting-rate band from the first pound of taxable non-savings income. The formula previously deducted the maximum personal allowance plus the starting-rate limit from income that already had allowances deducted, leaving the full £5,000 exemption available too far up the income range.

The correction uses the existing non-savings taxable-income aggregate, limit and phase-out parameters. It preserves property-income aggregation, personal-allowance allocation, the personal savings allowance and the remaining savings tax bands. Dividends do not directly reduce the starting-rate band; they can change personal-allowance or savings-allowance entitlement through the existing rules. The separate limit-indexation change in #2452 is independent.

The rule follows ITA 2007 s12(1) and (5), s16 income ordering, and HMRC's starting-rate guidance. Examples below use ordinary English 2026–27 inputs, the standard £12,570 personal allowance, £1,000 personal savings allowance, no other reliefs and the 20% savings basic rate:

Annual wages	Savings interest	Correct starting-rate band used	Savings income tax	Total income tax
£14,000	£6,000	£3,570	£286	£572
£30,000	£4,000	£0	£600	£4,086

For the £30,000 wages case, the original model charged £0 savings income tax. The tests use genuine annual income inputs and check final savings and total income tax, with boundaries, pensions, property, unused personal allowance, dividend-ordering and parameter-reform controls. The old five fixtures forced taxable intermediates and encoded the duplicated deduction; they are replaced by 14 meaningful cases with independently derived expected amounts.

Validation: all 75 focused tests passed (14 changed cases and 61 existing regressions; 104.59s pytest / 113.71s bounded), with per-case outcomes and JUnit recorded throughout. The final-tax-first regressions fail on unchanged production (11 failed, 3 controls passed). make format, Ruff and git diff --check pass. Local Python 3.13.14 / Core 3.32.27 differs from locked Core 3.33.2; hosted full CI remains required. Full make test, documentation build and private-data population analysis were not run for this scoped formula change.

Reproducer: policyengine_uk/tests/policy/baseline/finance/tax/income_tax/savings_starter_rate_income.yaml

axiom: needed

A rulespec-uk parity follow-up remains outstanding; no encoded-correctness or population estimate is claimed.

[vahid-ahmadi]

Looks right, fine to approve.

@juaristi22, the new formula matches ITA 2007 s12 and the GOV.UK guidance. non_savings_taxable_income is already net of allowances, so each pound of it now reduces the £5,000 band once. The old code deducted the personal allowance and the band a second time.

1. nit, policyengine_uk/variables/gov/hmrc/income_tax/bracketized_savings_income/savings_starter_rate_income.py:46 reads "ss.12(5) and16". The YAML header comments have the same missing spaces: savings_starter_rate_income.yaml lines 1, 7, 8, 11 and 210 ("rate20%", "or£500", "band£500", "future2027-28", "less50% of£1,430"). Worth tidying, because these comments are the hand-workings a reviewer reads.

What I checked

• Read the issue, the PR body, the full diff (including all 14 YAML cases), and non_savings_taxable_income and basic_rate_savings_income_pre_starter, to confirm the base already includes property income and is net of allowances.
• Worked two cases by hand. £14,000 wages and £6,000 interest: 1,430 taxable wages, leaving a 3,570 band; (6,000 - 3,570 - 1,000) x 20% = £286, plus wage tax £286, gives £572. £30,000 wages and £4,000 interest: no band left; (4,000 - 1,000) x 20% = £600. Both match the YAML. I also worked one case the PR body doesn't cover. With £10,000 wages and £9,000 interest, 2,570 of unused allowance leaves 6,430 of taxable interest. After the 5,000 band and 1,000 PSA, 430 x 20% = £86. This matches.
• Ran policyengine-core test .../savings_starter_rate_income.yaml: 14 passed. With the formula reverted to main: 11 failed, 3 passed.
• Merged this branch with #2452 locally. Both PRs' tests pass together.
• ruff check and ruff format --check on the formula file: clean.
• Hosted CI on 35f992c is green.
• Process: the changelog fragment is .fixed.md; the Reproducer and axiom lines are present; the Fix evidence check is green.
• Not checked: the rulespec-uk parity (axiom: needed), and any population-level effect.

🤖 Generated with Claude Code

[vahid-ahmadi]

Approving, per my review above. The shoulds and nits can follow.