Brief
Diagnose the provided curve bundle (3 cases) and submit the fix per case with re-measured validation evidence.
Dataset
Curve bundle: loss histories + configs for underfit / overfit / unstable cases.
Starter guardrails (88)
- Template slots per case: diagnosis -> prescribed fix -> re-run evidence.
- One variable changed per re-run (no bundled fixes).
Deterministic checks
Diagnoses match the key; re-run configs differ by the prescribed variable only.
100-pt rubric
- Diagnosis correctness per case (3 x 20 pts): 60
- Fix appropriateness (ordered menu): 20
- Re-measurement evidence: 10
- Single-variable discipline: 10