{
  "schema_version": 2,
  "date": "2026-08-20",
  "scope": [
    "MM-003",
    "AG-001",
    "AG-002",
    "EV-002"
  ],
  "embodied_post_training": {
    "policy_id": "lerobot/smolvla_base",
    "policy_revision": "c83c3163b8ca9b7e67c509fffd9121e66cb96205",
    "vlm_id": "HuggingFaceTB/SmolVLM2-500M-Video-Instruct",
    "vlm_revision": "7b375e1b73b11138ff12fe22c8f2822d8fe03467",
    "dataset_id": "lerobot/svla_so100_pickplace",
    "dataset_revision": "728583b5eaf9e739a7f119e2def466fa1d552402",
    "device_class": "single NVIDIA A30",
    "lerobot_package": "0.4.4",
    "steps": 1,
    "batch_size": 1,
    "loss": 0.297,
    "gradient_norm": 6.4,
    "learnable_parameters": 99880992,
    "total_parameters": 450046176,
    "dataset_frames": 19631,
    "dataset_episodes": 50,
    "dataset_manifest_sha256": "d5d507229bdb25e39df93ec46e7f7f856f65d425d7ceac49981ef508860fc4b9",
    "status": "completed",
    "claim_boundary": "真实视频 batch 的前向、反向与 checkpoint 保存链路通过；未运行机器人或 simulator episode，不报告 success rate。"
  },
  "agent_lightning_checkpoint": {
    "model_id": "HuggingFaceTB/SmolLM2-135M-Instruct",
    "model_revision": "12fd25f77366fa6b3b4b768ec3050bf629380bac",
    "device_class": "single NVIDIA A30",
    "steps": 3,
    "episodes": 3,
    "credit_updates": 6,
    "transition_spans": 9,
    "losses": [
      0.5013298988,
      0.4723831713,
      0.6141988635
    ],
    "baseline_joint_success": 1.0,
    "final_joint_success": 1.0,
    "conclusion": "真实 checkpoint 反传与 executor 闭环通过，但该饱和 micro-suite 没有能力提升。"
  },
  "executor_matrix": {
    "episodes_per_run": 12,
    "seeds": [
      42,
      43,
      44
    ],
    "maximum_subprocess_commands_per_episode": 4,
    "methods": {
      "direct": {
        "joint_success": 0.0,
        "average_cost": 2.0,
        "reuse_rate": 0.0
      },
      "critic": {
        "joint_success": 1.0,
        "average_cost": 4.0,
        "reuse_rate": 0.0
      },
      "agent-lightning": {
        "joint_success": 1.0,
        "average_cost": 3.625,
        "reuse_rate": 0.75
      },
      "swe-agent": {
        "joint_success": 1.0,
        "average_cost": 2.5,
        "reuse_rate": 0.0
      },
      "openhands": {
        "joint_success": 1.0,
        "average_cost": 2.5,
        "reuse_rate": 0.0
      }
    }
  },
  "post_training_evolve": {
    "dataset": "arithmetic-smoke",
    "generations": 2,
    "population": 5,
    "seed": 42,
    "champion": "g2-t1",
    "champion_objective": "kto",
    "validation_accuracy": 0.734375,
    "selection_fitness": 0.734374,
    "claim_boundary": "小型 candidate-policy 验证组合控制器；不等同于 checkpoint 能力收益。"
  },
  "manifest_ref": "experiments:roadmap-4-7",
  "evaluation_protocol": {
    "tier": "l1_mechanism",
    "seeds": [
      42
    ],
    "formal_comparison": false,
    "claim_policy": "single/few-seed smoke result; do not claim a stable improvement"
  },
  "provenance": {
    "artifact_path": "docs/experiments/roadmap-4-7.json",
    "dataset_fingerprint": {
      "svla_manifest_sha256": "d5d507229bdb25e39df93ec46e7f7f856f65d425d7ceac49981ef508860fc4b9",
      "agent_executor_suite": "local-code-benchmark-v1:3-deterministic-tasks",
      "post_training_suite": "arithmetic-smoke:seed42:maximum-examples32"
    }
  }
}