Solo-Arise

MLDoctor reproduces, investigates, patches, and verifies broken ML training scripts. Training fails quietly: it diverges, or gives different results every run despite a fixed seed. MLDoctor runs the script repeatedly under telemetry wrappers, classifies the symptom (healthy, non_reproducible, diverged, error, leakage) from real parameter hashes and loss curves, asks a model for a strict-schema diagnosis and exact-string patch, then re-runs the patched script to earn the verdict. An anti-masking gate rejects cheating patches (clamped losses, removed shuffling, hardcoded metrics) before re-running. Only real re-runs earn a PASS; the model never decides. Ships as a one-process d ashboard (prebuilt Next.js frontend + FastAPI backend), Docker image, and CI reproducibility gate.