v2.1.0 — IFRS feature review

Adds a capability the skill did not previously have: reviewing software that produces accounting figures, rather than answering questions about the standards.

What it does

Give it a feature that produces an accounting figure — a billing engine, a revenue report, a balance-sheet generator — together with one real output it produced, and it reports what is wrong in language a developer can act on. It refuses to run on code alone: reading an implementation and pronouncing it correct is guessing at the output rather than looking at it.

Findings carry a class saying who fixes the defect — Non-compliant (accounting logic), Wrong (calculation), Incomplete (data model), Untraceable (logging) — and a severity: Blocking, Needs work, Conforms. The verdict follows arithmetically: any blocking finding means not ready. The reviewer states what a standard requires and what the code does instead. It never recommends a provision rate, an estimate or a policy, because those are decisions for a qualified accountant.

Standard numbers stay out of the body of a finding and sit in a source line at the end, so a review reads as plain English to an engineer and stays checkable by their accountant.

Coverage

Six standards, the ones software actually touches: IFRS 15 revenue, IFRS 9 receivables and expected credit losses, IAS 1 and IFRS 18 presentation, IAS 21 foreign currency, IFRS 16 leases, IAS 7 cash flows. Anything a feature touches outside them is reported as unchecked rather than guessed at — silence would read as approval.

Every review also checks the 1 January 2027 presentation change, since code written today is still running when IFRS 18 replaces IAS 1.

How it was verified

tests/ holds a deliberately defective balance-sheet generator with four planted defects, and an answer key kept outside the fixture directory so a reviewing agent cannot read what it is being tested on.

An agent with no sight of the key found all four defects at the correct severity — and four more that were not planted, including that the statement's Cash line is the sum of paid invoices and therefore cannot be a cash balance. It also declined to recommend a provision rate, which the skill forbids.

That run corrected the key: the sales-tax defect had been classed Wrong, and breaking a requirement is Non-compliant.

scripts/check_citations.py now resolves the citations in answer keys as well as the skill — a key with a wrong paragraph reference would fail a correct review. 4,239 of 4,245 checkable citations resolve (99.86%), across 4,452 in total.

Also in this release

• scripts/check_feature_review.py — checks a review's shape against the rules the skill sets
• CONTEXT.md — the domain glossary
• docs/adr/ — two decisions recorded: that practice notes may be uncited under a fixed heading, and that reviews split by standard rather than by module
• A fourth check in CI, asserting every figure the answer key quotes

Full detail in CHANGELOG.md.

---

This skill provides technical guidance and does not replace professional judgement. Consult qualified professionals for specific accounting decisions.