Black Box

Black Box is a forensic copilot for robots and embodied systems. You hand it a field recording — whatever the platform produces: ROS1/ROS2 bags, raw telemetry logs, multi-camera video, LiDAR/IMU traces, GPS/RTK data, controller logs, or repo context — and it returns a grounded post-mortem: ranked root-cause hypotheses, timestamped multimodal evidence, an NTSB-style PDF, and a scoped code patch. The core insight is that robotics failures are rarely obvious from one log or one operator's account. In our hero case, the operator filed "tunnel caused the anomaly." Black Box's grounding gate cross-checked telemetry, video, and source code and refuted the operator: RTK heading was already degraded 43 minutes before tunnel entry, drive-by-wire never engaged, so the tunnel could not plausibly be the cause. The refutation ships as a ranked hypothesis with its own confidence and patch hint — not as agreement. We built around long-horizon agentic investigation rather than chat. The system includes a FastAPI/HTMX UI with streaming reasoning, a deterministic grounding gate that drops weakly-supported hypotheses, an append-only 4-layer memory stack (case → platform → taxonomy → eval), a frozen 7-class bug taxonomy enforced at parse time by Pydantic, an HITL approve/reject gate before any patch is applied, and Claude Managed Agents for persistent forensic sessions with tool use, checkpoints, time-travel rollback, and operator steering.