Use human demonstrations only when their recording metadata includes
demonstration_source: human and a pseudonymous participant_id. Create a
checksum-bound held-out corpus with create_human_benchmark_manifest.
behavioral_fidelity compares camera velocity, acceleration, jerk, and map
occupancy; lower values indicate closer behavior. behavioral_safety_report
reports repeated actions, blind fire, and stationary windows. These are
evaluation signals, not training rewards or claims of human equivalence.
For every model, plan RGB-only, optional privileged-modality, temporal-context, and leave-one-scenario-out runs separately. Never compare those conditions as though they used identical observations or data.