{
  "schema_version": 1,
  "audience": "Robotics infrastructure engineers deciding whether RoboWBC is credible enough to adopt for the next whole-body-control wrapper or deployment stack.",
  "decision": "Adopt RoboWBC when it stays inside an acceptable matched-path latency band while materially reducing setup friction, config complexity, and model-switch cost.",
  "thresholds": {
    "favorable": "RoboWBC is at or near CPU parity on the matched path and still easier to integrate or switch across models.",
    "neutral": "RoboWBC is slower on CPU but still comfortably inside the control budget and the DX or portability win is substantial.",
    "unfavorable": "RoboWBC is materially slower on a critical matched path or cannot expose the same path semantics without extra architecture work."
  },
  "fairness_rules": [
    "Run both stacks on the same host machine and record the host fingerprint beside every artifact.",
    "Match the execution provider for the row being compared; do not compare CPU against TensorRT under one label.",
    "Record the upstream NVIDIA commit, RoboWBC commit, and model revision beside every result.",
    "Use the exact command fixture and warmup policy named by the case registry.",
    "Publish one row per matched path; do not collapse cold, warm, replan, tracking, walk, and balance into one aggregate number.",
    "Pair every latency row with an explicit user-facing interpretation about why the result matters."
  ],
  "cases": [
    {
      "case_id": "gear_sonic_velocity/cold_start_tick",
      "family": "gear_sonic",
      "kind": "microbenchmark",
      "description": "First velocity tick after reset; forces a fresh planner path.",
      "criterion_id": "policy/gear_sonic_velocity/cold_start_tick",
      "command_fixture": {
        "kind": "velocity",
        "data": [0.3, 0.0, 0.0]
      },
      "warmup_policy": "reset_then_single_tick",
      "robowbc_command": "GEAR_SONIC_MODEL_DIR=models/gear-sonic cargo bench -p robowbc-ort --bench inference -- --output-format bencher 'policy/gear_sonic_velocity/cold_start_tick'",
      "official_command": "python3 scripts/bench_nvidia_official.py --case gear_sonic_velocity/cold_start_tick",
      "interpretation": "If this row regresses, RoboWBC still pays a measurable planner cold-start tax before any DX win can matter."
    },
    {
      "case_id": "gear_sonic_velocity/warm_steady_state_tick",
      "family": "gear_sonic",
      "kind": "microbenchmark",
      "description": "Velocity interpolation tick after an initial plan; planner stays idle.",
      "criterion_id": "policy/gear_sonic_velocity/warm_steady_state_tick",
      "command_fixture": {
        "kind": "velocity",
        "data": [0.3, 0.0, 0.0]
      },
      "warmup_policy": "reset_then_one_warmup_tick_then_measure",
      "robowbc_command": "GEAR_SONIC_MODEL_DIR=models/gear-sonic cargo bench -p robowbc-ort --bench inference -- --output-format bencher 'policy/gear_sonic_velocity/warm_steady_state_tick'",
      "official_command": "python3 scripts/bench_nvidia_official.py --case gear_sonic_velocity/warm_steady_state_tick",
      "interpretation": "This is the steady-state path most likely to dominate the average control budget for locomotion."
    },
    {
      "case_id": "gear_sonic_velocity/replan_tick",
      "family": "gear_sonic",
      "kind": "microbenchmark",
      "description": "Velocity tick that crosses the replan boundary and executes planner_sonic.onnx.",
      "criterion_id": "policy/gear_sonic_velocity/replan_tick",
      "command_fixture": {
        "kind": "velocity",
        "data": [0.3, 0.0, 0.0]
      },
      "warmup_policy": "reset_then_prime_until_replan_boundary",
      "robowbc_command": "GEAR_SONIC_MODEL_DIR=models/gear-sonic cargo bench -p robowbc-ort --bench inference -- --output-format bencher 'policy/gear_sonic_velocity/replan_tick'",
      "official_command": "python3 scripts/bench_nvidia_official.py --case gear_sonic_velocity/replan_tick",
      "interpretation": "If replan ticks miss parity, the planner path needs attention before stronger NVIDIA comparison claims."
    },
    {
      "case_id": "gear_sonic_tracking/standing_placeholder_tick",
      "family": "gear_sonic",
      "kind": "microbenchmark",
      "description": "Encoder plus decoder standing-placeholder tracking tick with planner loaded but idle.",
      "criterion_id": "policy/gear_sonic_tracking/standing_placeholder_tick",
      "command_fixture": {
        "kind": "standing_placeholder_tracking",
        "data": []
      },
      "warmup_policy": "reset_then_single_tracking_tick",
      "robowbc_command": "GEAR_SONIC_MODEL_DIR=models/gear-sonic cargo bench -p robowbc-ort --bench inference -- --output-format bencher 'policy/gear_sonic_tracking/standing_placeholder_tick'",
      "official_command": "python3 scripts/bench_nvidia_official.py --case gear_sonic_tracking/standing_placeholder_tick",
      "interpretation": "A tracking-path miss suggests the motion-reference contract is the real optimization target, not the planner path."
    },
    {
      "case_id": "decoupled_wbc/walk_predict",
      "family": "decoupled_wbc",
      "kind": "microbenchmark",
      "description": "GR00T G1 history contract with a movement command that routes to the walk checkpoint.",
      "criterion_id": "policy/decoupled_wbc/walk_predict",
      "command_fixture": {
        "kind": "velocity",
        "data": [0.25, 0.0, 0.05]
      },
      "warmup_policy": "reset_then_single_tick_with_motion_command",
      "robowbc_command": "DECOUPLED_WBC_MODEL_DIR=models/decoupled-wbc cargo bench -p robowbc-ort --bench inference -- --output-format bencher 'policy/decoupled_wbc/walk_predict'",
      "official_command": "python3 scripts/bench_nvidia_official.py --case decoupled_wbc/walk_predict",
      "interpretation": "This row captures the locomotion path people hit while actually moving, not the standing fallback."
    },
    {
      "case_id": "decoupled_wbc/balance_predict",
      "family": "decoupled_wbc",
      "kind": "microbenchmark",
      "description": "GR00T G1 history contract with a near-zero command that routes to the balance checkpoint.",
      "criterion_id": "policy/decoupled_wbc/balance_predict",
      "command_fixture": {
        "kind": "velocity",
        "data": [0.0, 0.0, 0.0]
      },
      "warmup_policy": "reset_then_single_tick_with_balance_command",
      "robowbc_command": "DECOUPLED_WBC_MODEL_DIR=models/decoupled-wbc cargo bench -p robowbc-ort --bench inference -- --output-format bencher 'policy/decoupled_wbc/balance_predict'",
      "official_command": "python3 scripts/bench_nvidia_official.py --case decoupled_wbc/balance_predict",
      "interpretation": "This row makes the command-magnitude model switch explicit instead of hiding it behind one average number."
    },
    {
      "case_id": "gear_sonic/end_to_end_cli_loop",
      "family": "gear_sonic",
      "kind": "end_to_end",
      "description": "RoboWBC CLI loop running configs/sonic_g1.toml through the synthetic transport with a JSON report.",
      "config_path": "configs/sonic_g1.toml",
      "command_fixture": {
        "kind": "velocity",
        "data": [0.3, 0.0, 0.0]
      },
      "warmup_policy": "default_cli_loop_200_ticks",
      "robowbc_command": "python3 scripts/bench_robowbc_compare.py --case gear_sonic/end_to_end_cli_loop",
      "official_command": "python3 scripts/bench_nvidia_official.py --case gear_sonic/end_to_end_cli_loop",
      "interpretation": "This row answers the operator-facing question: can the whole deployment loop hold the target frequency, not just one inference call?"
    },
    {
      "case_id": "decoupled_wbc/end_to_end_cli_loop",
      "family": "decoupled_wbc",
      "kind": "end_to_end",
      "description": "RoboWBC CLI loop running configs/decoupled_g1.toml through the synthetic transport with a JSON report.",
      "config_path": "configs/decoupled_g1.toml",
      "command_fixture": {
        "kind": "velocity",
        "data": [0.25, 0.0, 0.05]
      },
      "warmup_policy": "default_cli_loop_200_ticks",
      "robowbc_command": "python3 scripts/bench_robowbc_compare.py --case decoupled_wbc/end_to_end_cli_loop",
      "official_command": "python3 scripts/bench_nvidia_official.py --case decoupled_wbc/end_to_end_cli_loop",
      "interpretation": "If the loop holds up here, the comparison story can move beyond one policy tick and toward deployable control loops."
    }
  ]
}