Write the knowledge behind an item family once — the variables, their values, and which answer options each value allows. CAFA AIG² enumerates every valid item, verifies each one, and certifies the set. Any domain.
The numbers on the right are computed in this page by the same code the app runs. Switch the domain to see another example counted live.
Three artefacts. The first two you write; the third is generated and checked by code, so the correct answer of every item is decided by logic, never by a language model.
The stem with variable positions, and the fixed answer options.
For every variable: its values, and the options each value is compatible with.
One of the valid combinations, with its answer — and every generated item is checked, not sampled.
The intersection rule: the answer to a combination is the intersection of its values' compatibility sets. A combination is a usable item only when exactly one option remains; an empty intersection is a contradiction and is dropped.
The rule only checks that the answer is unique. It cannot know that two values never occur together in the real domain. Untick a combination and watch the counts change — this is exactly what an author does in the app.
The fingerprint changes when you exclude a combination: an exclusion is part of the model, and every generated set records the fingerprint it came from.
When an author asks for a certified set, the server regenerates it from the stored model with the same deterministic core the browser runs, fires three positive controls on a deliberately corrupted copy (a leftover placeholder, a duplicate, a bad answer position — each must be caught, or the verdict is refused), verifies every item, and records the certificate. The browser has no permission to write one.
A certificate that hides its limits is worth less than none. The human review stays where it belongs.
Verify this certificate — anyone can, with the delivery id printed on every export.
A real certificate, issued 2026-09-29 for the nursing COPD family:
Licensing bodies, certification programs, and universities need more than generated items: they need to know who wrote what, from which version, and what was checked.