Disclaimer
What a CAFA AIG² certificate proves, what it does not, and what that means for using generated items.
What a certificate means
A CAFA AIG² certificate states that, for a specific set of generated items identified by its hash, the server regenerated the set from a specific model version identified by its fingerprint, fired three positive controls on a deliberately corrupted copy, and then verified every item for three structural properties: no unresolved placeholders, no duplicate conditions, and a valid answer position. If a control failed to fire, the verdict is "refused" and nothing is certified.
What it does not mean
- It does not verify that the values, options or answer keys are correct in the domain (clinically, mathematically, linguistically or otherwise). The answer of each item is the logical consequence of the Ontology Model you wrote; if the model is wrong, every item is consistently wrong.
- It does not verify fit to any curriculum, blueprint or test specification.
- It does not estimate difficulty, discrimination or any psychometric property.
- It does not verify that items which differ only in surface-only variables are equivalent for examinees.
Every certificate carries a "Not verified" line naming these limits. A certificate presented without that line has been altered.
Use in examinations
Generated items are drafts for expert review. Before any item is used where the result matters to a person — a course grade, a licensing decision, a certification — it must be reviewed by a qualified subject-matter expert and, where applicable, follow the test owner's own review and approval process. CAFA Lab does not review item content and accepts no responsibility for its use.
AI drafts
Models drafted by the AI assistant are proposals. They may contain factual errors, outdated practice, or values that are incompatible in the real domain. The app measures a draft's structure (how many items it yields, how the answers are distributed) — it cannot measure whether the draft is true.
Reproducibility
Given the same model version, seed and settings, the service produces the same items. This is a property of the code, not a warranty: it holds for the engine version named on the certificate, and a future engine version may render differently.
Questions
CAFA Lab, Inc. — support@cafalab.com.