mirror of
https://github.com/hanzoai/iam.git
synced 2026-11-19 18:37:16 +00:00
The training answer was a bool, so "never asked" and "asked and declined" shared a value. Nothing could tell whether an answer had ever been given, which means no screen could know to ask and no data path could treat silence as refusal. Training becomes a tri-state Answer (unanswered/granted/refused) with the zero value being unanswered, so a missing record, an unparseable blob, a wrong JSON type and an unrecognized token all decode to silence. MayTrain admits exactly one value — an explicit "granted" — and is the ONE predicate; pkg/model aliases the type and re-exports the states so a consumer outside this module shares the same definition rather than re-deriving what granted means. The write path validates at the boundary, so an answer this version does not know is refused rather than persisted for a later reader to interpret, and a change to the answer writes an AuditLog row carrying the prior and new value — a grant and a later revocation are both attributable, which overwriting a JSON field is not. The preferences property is now defined once in schema, where the consent record nested in the same blob reads it, so the two cannot drift apart. scripts/mutants.py drives cloud's strict-scored mutation engine against these properties: 8 rows, 8 KILLED. Co-authored-by: Hanzo Dev <[email protected]>