Validation Plan

Cheapest first, no build:

1. Get 3–5 real Australians to share (redacted) 12 months of statements + their actual lodged return.
2. Manually run classify, categorise, draft by hand (or with an LLM + rules).
3. Measure: transactions auto-classified correctly; deductions found vs missed; draft total vs real return.
4. Interview: would they trust an auto-draft? would they pay? would they still want a human check?

Pass: over 80% transactions auto-classified correctly, at least one missed deduction found per user, and users say they would use it if it stopped at "copy into myTax".

Kill/pivot: classification too noisy without heavy manual work; users will not share statements; the regulatory/liability burden sinks even the assistant version.