Checkpoint preflight

flowedge-inspect validates a .safetensors checkpoint before it reaches an inference service or robot.

flowedge-inspect models/mamba_flow.safetensors
flowedge-inspect models/mamba_flow.safetensors --json

The JSON output is versioned with schema_version: 1 and contains the model family, precision, tensor count, FlowEdge digest, backbone/action dimensions, weight bytes, arena bytes, persistent decode-state bytes, Diffusion Policy configuration fields, missing required tensors, unsupported tensors, and a compatibility result. The human-readable format is intended for deployment logs; JSON is intended for admission checks.

The command returns zero only for a supported checkpoint. It returns non-zero for malformed safetensors, an unknown model family, missing required tensors, incompatible dimensions, unsupported tensor dtypes, or a checkpoint that the runtime weight loader cannot open.

The report estimates the existing fixed runtime arena. It does not start worker threads or initialize persistent decode state for a deployed engine. Observation preprocessing and model training remain outside FlowEdge.