{
  "source_revision": "07e50d0985bb91c4f613017df7a30e9ea68394c0",
  "source_date": "2026-10-18",
  "simulation": {
    "title": "Corrected single-arm simulation comparison",
    "task": "Red block into blue container",
    "fps": 20,
    "seeds": [
      20000,
      20001,
      20002,
      20003,
      20004
    ],
    "train_episodes": 64,
    "validation_episodes": 16,
    "frames": 32220,
    "models": [
      {
        "name": "Zero-shot",
        "contact": 5,
        "bilateral": 0,
        "lift": 0,
        "placement": 0,
        "max_lift_cm": 1.92
      },
      {
        "name": "Raw LoRA",
        "contact": 5,
        "bilateral": 3,
        "lift": 2,
        "placement": 0,
        "max_lift_cm": 15.69
      },
      {
        "name": "EMA LoRA",
        "contact": 2,
        "bilateral": 0,
        "lift": 0,
        "placement": 0,
        "max_lift_cm": 1.28
      }
    ],
    "n": 5,
    "lift_threshold_cm": 3.5,
    "sources": [
      "docs/validation/single-arm-v1.md",
      "experiments/single-arm-v1/simulation-results.json"
    ]
  },
  "hardware": {
    "official_trials": 16,
    "execution_requests": 11,
    "trials_with_commands": 9,
    "independent_task_success_labels": 0,
    "interpretation": "No hardware task success rate can be inferred from these engineering trials.",
    "sources": [
      "experiments/single-arm-v1/real-robot-results.json"
    ]
  },
  "twin": {
    "generated": 576,
    "accepted": 243,
    "green": 133,
    "red": 110,
    "evaluations_per_checkpoint": 20,
    "placements_per_checkpoint": 0,
    "interpretation": "Scripted training demonstrations and learned policy evaluation are different outcomes.",
    "sources": [
      "media/README.md",
      "examples/dual_so101_sim/twin/README.md"
    ]
  },
  "real_pen_highlights": {
    "task": "Put the red pen in its slot in the cardboard box",
    "model": "FLUX red-pen LoRA RAW 5000",
    "actions_per_request": 16,
    "blocks": [
      {
        "label": "Early morning block",
        "reported_successes": 3,
        "attempts": 5
      },
      {
        "label": "07:40 block",
        "reported_successes": 2,
        "attempts": 3
      }
    ],
    "interpretation": "Team-reported small session blocks with retained successful single-arm camera videos. Not a matched benchmark, a pooled hardware success rate or a full learned bimanual demonstration. Separate from the older 16-trial official index.",
    "sources": [
      "media/README.md",
      "media/real_robot/flux_red_pen_runs/"
    ]
  }
}