Charge Class 4 NI on full trading profits, without deducting Class 1

[MaxGhenis]

Fixes #1885

What was wrong

ni_class_4 and ni_class_4_main computed Class 4 profits as self_employment_income - ni_class_1_employee. Trading profits were reduced by the person's employee Class 1 NI before the Lower and Upper Profits Limits applied.

The law has no such deduction. SSCBA 1992 Sch 2 para 2 charges Class 4 on "the full amount of all profits" chargeable under ITTOIA 2005 Part 2 Ch 2. The only reliefs are those in para 3: trade losses under ITA 2007 ss.64, 72, 83 and 89, and the interest and annual payments in para 3(5). Class 1 contributions are not among them. The Class 1 / Class 4 interaction is handled solely by the regulation 100 annual maximum. The model already implements that in ni_class_4_maximum, and its Steps Six and Nine already used undeducted profits. The deduction arrived with a 2021 validation commit (66b6dce) with no comment or reference. The only Class-1-related exclusion in law is reg 94 of the 2001 Regulations (NIM24620). It removes earnings that bear Class 1 but are taxed as trading income, on application; it does not remove Class 1 contributions. It can't arise here, because employment_income and self_employment_income are separate inputs.

Example (2026-27): employment £30,000 and profits £20,000. The statute gives 6% × (20,000 − 12,570) = £445.80, and the annual maximum does not bind. The model returned about £362.

Changes

• ni_class_4, ni_class_4_main: Class 4 profits are self_employment_income.
• ni_class_4_maximum: the Case 1 unused-band test (from Fix Class 4 NI dropping the additional-rate band at non-round thresholds #1884) uses the same profits.
• ni_class_4 gets its own label ("NI Class 4 contributions", previously a copy of ni_class_4_main's) and a s.15 reference.
• Tests:
  • A dual-earner example test (employment £30,000, profits £20,000 in 2026: £445.80).
  • The Fix Class 4 NI dropping the additional-rate band at non-round thresholds #1884 Case 1 boundary test now expects full profits.
  • The Hypothesis differential reference now uses full profits.
  • Invariant 1 now covers everyone the maximum does not bind, not only people without Class 1.
• Docs: the NI notebook's Class 4 row no longer says "self-employment income minus Class 1 employee NI". The saved table output is regenerated to match, since docs builds don't execute notebooks.

Invariants

These hold for any 0 ≤ LPL < UPL, rates ≥ 0 and profits ≥ 0. Comparisons allow float32 storage: 1p + 4×10⁻⁷ × the largest amount involved.

1. Statute when the maximum does not bind. Whenever reg 100 does not apply (no primary Class 1 or Class 2) or its maximum is at least the s.15(3) amount, ni_class_4 = main × clamp(p − LPL, 0, UPL − LPL) + additional × max(p − UPL, 0), with p the full self-employment profit.
2. Differential. For everyone, ni_class_4 equals an exact-rational implementation of s.15(3) capped by the literal reg 100 steps.
3. Bounds. 0 ≤ ni_class_4 ≤ the s.15(3) amount.
4. Monotonicity. With employment income fixed, ni_class_4 is non-decreasing in profits wherever Class 2 does not change. Before April 2024, reg 100 subtracts Class 2, so crossing the small profits threshold can lower Class 4; Class 2 + Class 4 stays non-decreasing when main ≥ additional.
5. Scope of change. Only people with employee Class 1 NI can change, and their Class 4 can only rise. For fixed Class 2 (unchanged here, since it depends on undeducted income both before and after), the pre-maximum amount and the maximum are both non-decreasing in the Class 4 profit base, and this change only raises that base.

Verification

Self-employment × employment grid: 1,732 profit levels (£0–£300k) × 152 employment incomes (£0–£150k), 263,264 cells per year. Base is #1884's merged code (2.102.1), head is this PR.

Year	Cells changed	Changed without Class 1	Decreases	Mean increase (changed cells)	Max increase
2023	9,955	0	0	£102.66	£282.52
2026	7,144	0	0	£41.16	£125.56
2028	7,480	0	0	£42.03	£129.12
2030	7,714	0	0	£44.04	£135.85

In every year, each cell where the maximum does not bind equals the closed-form statute on full profits. Cells where the maximum binds, which is most dual earners with profits well above the LPL, are unchanged. 2023 is larger because of that year's 9% main rate.

Aggregate revenue effect on the FRS-based microsimulation: not computed locally, since the dataset needs credentials. CI's microsimulation tests run on it.

Axiom parity

axiom: uk:statutes/ukpga/1992/4/15 encoded-correct (TheAxiomFoundation/rulespec-uk uk/statutes/ukpga/1992/4/15.yaml charges profits_chargeable_to_class_4_contributions with no Class 1 deduction; companion test 15.test.yaml::class_4_main_and_additional_contributions_apply)

The Axiom-vs-policyengine-uk NICs grid in TheAxiomFoundation/axiom-oracles#565 attributes 4 dual-earner rows to #1885.

Independent review

Claude Opus 5.5 via Subfleet (read-only) returned APPROVE with nothing blocking:

• It confirmed the legal claim from SSCBA Sch 2 paras 2–3, reg 100 and reg 94 / NIM24620.
• It found no other profit deduction in the model code.
• It proved the "only Class 1 payers, only upward" invariant, and reproduced the £445.80 / £362.14 example by hand.

Its notes are addressed in this PR: the stale saved notebook output, the monotonicity caveat, the Axiom repo path, the reg 94 note and the duplicated ni_class_4 label.

Commands run

• ruff format . and ruff check .
• pytest policyengine_uk/tests/test_ni_class_4_maximum.py policyengine_uk/tests/test_ni_class_4_properties.py policyengine_uk/tests/test_parameter_descriptions.py (16 passed)
• policyengine-core test policyengine_uk/tests/policy/baseline/gov/hmrc/national_insurance -c policyengine_uk (38 passed)
• A 300-random-example (non-derandomised) Hypothesis campaign per property, with the model pinned to this branch: both passed.

🤖 Generated with Claude Code