Repository navigation
A miscellaneous loss is relieved against all other income, not only miscellaneous income (ITA 2007 s. 152(4)) #2326
Description
What the law says
ITA 2007 s. 152 gives relief for a loss on a miscellaneous transaction. Subsection (4) limits where it can go:
(4) But a deduction for that purpose is to be made only from the person's relevant miscellaneous income.
So a miscellaneous loss reduces only miscellaneous income, in the same year and later years. It never reduces employment, pension, property, savings or dividend income.
What the model does
taxable_miscellaneous_income is adds = ["miscellaneous_income"], so a negative miscellaneous_income passes straight through. It is a component of adjusted net income and total income, so the loss is set against every other source of income.
Minimal reproduction
policyengine-uk 2.126.7, and main at 6d89338:
from policyengine_uk import Simulation
def income_tax(**inputs):
s = {
"people": {"p": {"age": {"2026": 45}, **{k: {"2026": v} for k, v in inputs.items()}}},
"benunits": {"b": {"members": ["p"]}},
"households": {"h": {"members": ["p"], "country": {"2026": "ENGLAND"}}},
}
return Simulation(situation=s).calculate("income_tax", 2026)[0]
income_tax(employment_income=50_000) # 7,486
income_tax(employment_income=50_000, miscellaneous_income=-10_000) # 5,486 the law gives 7,486A property loss is handled correctly: Income Tax stays at £7,486, though total_income does fall by the loss.
How it was found
The PolicyBench UK round 2 internal audit (October 2026). The public transfer dataset carries negative miscellaneous income for 964 of 58,848 people, and the round 2 sample includes several. In one, a household with −£2.4m of miscellaneous income pays no tax on a £510,223 salary. The affected audit cells will be added below.
A fix with tests follows in a PR.
Activity
MaxGhenis commentedon Oct 11, 2026
Evidence from PolicyBench UK round 2, an internal audit of policyengine-uk 2.126.7 (references identical on 2.126.8). Six models answered the same 200 households. A reference adversary (a blind GPT-6.1 Sol judge, then reconciliation with the engine log) reviewed consensus disagreements and a sample of single-model misses, and each verdict was adjudicated against the engine code and the law.
Audited cells that turn on this issue: 14. "With #2327" is the reference recomputed on that branch alone; "exact" counts the six models within £1 of the reference, before and after.
All cells, audited or not. Of the run's 2,800 references, #2327 moves 16 (14 of them audited). In those 16 cells, model answers matching the reference go from 17 to 18 of 96.
Source:
audit/adjudication.jsonandaudit/attribution/attribution.jsonin the round 2 record.