The model observes. The rubric decides.
A competency result that cannot name the rubric version it was judged against is not evidence. Ours can.

Scoring is deterministic: rules against an approved rubric, never a language model’s opinion of an answer.
Every rubric criterion carries a reference back to the document that authorised it.
Each attempt is recorded as its own row, so a second attempt after a hint is distinguishable from a first-time correct answer.
What this rests on
- Competency is scored by rules against an approved rubric, not by a language model’s opinionForty-one automated checks, including a deliberate break to prove the checks can fail. The model records what it observed; the rubric decides the mark.
- Every procedure step and rubric criterion carries a reference back to the document it came fromA criterion with no source reference is refused before it can be saved. Not a warning — a rejection.
- Interactive STEM visuals — graphs, geometry, circuits, reactions — that the learner can manipulate54 STEM diagrams and 27 interactive ones ship today, and every one is re-drawn and checked before a release.