Assessment methodology

Each exercise has a versioned rubric of weighted dimensions. A submission is scored per dimension by an automated evaluator; the output must pass validation (schema, scores within bounds, quoted evidence that really appears in your text) or the assessment fails rather than guessing.

The overall level is computed by the platform from the weighted score, not chosen by the model. Automatic results are capped at Level 3 and marked provisional. Levels 4–5, suspected prompt-injection, and low-confidence results go to a human trainer; only a trainer review marks a skill as verified.

The evaluator has not yet been calibrated against expert-labelled data. Treat scores as feedback, not as a verdict.