Methodology

Every checker on this site is a deterministic rules engine over published data. No model guesses at your answer; the same inputs produce the same result every time, and the working is auditable.

Where the numbers come from

ACC classification rates are machine-extracted from the published Levy Guidebook PDFs (2024-25, 2025-26, 2026-27), audited by re-locating sampled units in the raw document, then spot-checked by hand before the file is marked verified. KiwiSaver and ESCT figures come from IRD's published guides; weekly compensation figures from ACC's published rates and the Accident Compensation Act 2001. Every data file carries its source URL and the date it was verified — and until a file is verified, the engine reaches its conclusion but refuses to show a figure. Stale data (past its validity date) gets the same treatment. Silent wrongness is the failure mode nobody would catch, so it is engineered out.

Order of operations

Disqualifying questions come before any dollar figure, always. An unanswered question is treated as unanswered — never as a "no". Arithmetic checks run before judgement checks. Where two answers both fit, the checker stops and says so; "ring ACC, it's free" is a result we are content to give.

The classification method

The checker never picks a code from a description directly. It classifies the situation into one of six archetypes, asks that archetype's fixed disqualifiers, and only then applies ACC's published highest-rate rule:

Year-by-year comparisons use each year's own published rate. Part years are flagged, never apportioned — ACC apportions from its own records. Estimates are shown as a deliberately asymmetric band (0.75×–1.15×), and some findings carry no estimate at all because the number belongs to ACC's records, not ours.

What we cannot see

Your agency file, Experience Rating adjustments, employment agreements, IRD's processing queues. Each checker lists what it could not see on the result itself.

Auditability

The engine, its guardrails and its test gates are exercised before every deploy: twenty-two gates covering everything from "no number may exist in code" to "the ACC-bound letter must contain no dollar figure", plus a proof that every guard actually fires when its invariant is broken. The changelog records every data refresh.