mgsa

PolicyGrader is an agentic system that helps make robots safer to operate alongside humans. It automatically stress-test and analyse robot control policies, end-to-end. It compresses a process that costs robotics teams weeks of frame-by-frame video review into minutes. You describe an evaluation goal in plain English; then a swarm of specialized Claude Opus 4.7 Managed Agents — a planner, a rollout worker, K-parallel vision judges, and a reporter — design the test suite, execute it in simulation, point at each failure with pixel accuracy (or honestly abstain), and cluster the deployment findings into actionable patterns using Opus 4.7's 1M-context window. The differentiator is trust: every run includes a calibration cohort whose ground truth comes from a human labeling a sampled subset, and the judge's measured per-label precision is attached as a confidence chip to every deployment finding. No vibes with safety. The judge is auditable from the dashboard's runtime.json and findings.jsonl. Submitted to the Anthropic Opus 4.7 Hackathon.