{
  "title": "CosmicBrain physical AI evaluation protocol template",
  "version": "0.1",
  "status": "proposed_design_template_not_run_results",
  "published_date": "2026-10-02",
  "protocol": {
    "run_id": null,
    "task_manifest_hash": null,
    "task_instruction": null,
    "success_predicates": [],
    "terminal_state_required": true,
    "time_limit_seconds": null,
    "maximum_retries": null,
    "initiated_attempt_definition": null,
    "assistance_policy": null,
    "operating_envelope": null,
    "stop_rules": [],
    "planned_attempts": null,
    "primary_metric": "unassisted_success / valid_initiated_attempts",
    "release_thresholds": null,
    "pre_registered_exclusion_rules": []
  },
  "system": {
    "model_checkpoint_hash": null,
    "training_data_manifest": null,
    "known_pretraining_overlap": null,
    "adaptation_budget": null,
    "robot_platform": null,
    "gripper": null,
    "action_space": null,
    "controller": null,
    "control_frequency_hz": null,
    "camera_manifest": null,
    "calibration_hash": null,
    "inference_hardware": null,
    "runtime_commit": null,
    "simulation_version_if_applicable": null
  },
  "split": {
    "unit": "whole_episode",
    "group_by": [
      "site",
      "recording_session",
      "object_instance"
    ],
    "development_manifest": null,
    "validation_manifest": null,
    "held_out_test_manifest": null,
    "generalization_slices": [
      "objects",
      "environment",
      "instruction",
      "temporal",
      "embodiment"
    ]
  },
  "comparison": {
    "baseline_checkpoint": null,
    "matched_condition_blocks": [],
    "order_randomization_seed": null,
    "blind_outcome_review": null
  },
  "episode_template": {
    "episode_id": null,
    "site_id": null,
    "session_id": null,
    "task_id": null,
    "slice_ids": [],
    "initial_state_or_seed": null,
    "control_ownership_log": null,
    "video_artifacts": [],
    "observation_action_artifact": null,
    "start_timestamp": null,
    "terminal_timestamp": null,
    "outcome": null,
    "allowed_outcomes": [
      "unassisted_success",
      "assisted_completion",
      "policy_failure",
      "timeout",
      "policy_caused_stop",
      "infrastructure_invalid"
    ],
    "subgoals_completed": null,
    "subgoals_required": null,
    "interventions": [],
    "failure_category": null,
    "infrastructure_exclusion_reason": null,
    "reviewer_id": null,
    "adjudication_record": null
  },
  "analysis": {
    "successes": null,
    "valid_initiated_attempts": null,
    "excluded_infrastructure_attempts": null,
    "assisted_completions": null,
    "confidence_method": "Wilson 95% only when independent Bernoulli trials are defensible; otherwise use cluster-aware analysis",
    "aggregation": "publish per-task and per-site results, macro-average and pooled counts with declared weights",
    "cycle_time": "successful-run median/p95 with timeout fraction separately",
    "latency": "declare timestamp boundaries, p50/p95 and deadline misses",
    "throughput": "include declared reset time and downtime",
    "judge_audit": "rubric/version, inputs, outputs and held-out human reference"
  },
  "note": "Null means unmeasured or to be specified. Do not fill this template with invented evaluation results."
}