{
  "title": "Adroit fixed-start versus random-reset training: descriptive campaign comparison",
  "model": {
    "name": "gpt-6-astra",
    "reasoning": "max"
  },
  "rows": [
    {
      "task": "Door",
      "fixed": {
        "successes": 14,
        "episodes": 30,
        "search_rollouts": 247,
        "adaptation_env_steps": 80800,
        "provider_requests": 134
      },
      "random": {
        "successes": 29,
        "episodes": 30,
        "search_rollouts": 8858,
        "adaptation_env_steps": 1789920,
        "provider_requests": 671,
        "sampled_episodes_by_purpose": {
          "heldout": 30,
          "learning": 16,
          "search": 8858,
          "validation": 85
        }
      }
    },
    {
      "task": "Hammer",
      "fixed": {
        "successes": 25,
        "episodes": 30,
        "search_rollouts": 638,
        "adaptation_env_steps": 149200,
        "provider_requests": 124
      },
      "random": {
        "successes": 30,
        "episodes": 30,
        "search_rollouts": 875,
        "adaptation_env_steps": 186910,
        "provider_requests": 165,
        "sampled_episodes_by_purpose": {
          "heldout": 30,
          "learning": 15,
          "search": 875,
          "validation": 55
        }
      }
    },
    {
      "task": "Pen",
      "fixed": {
        "successes": 7,
        "episodes": 30,
        "search_rollouts": 1250,
        "adaptation_env_steps": 269400,
        "provider_requests": 248
      },
      "random": {
        "successes": 30,
        "episodes": 30,
        "search_rollouts": 6095,
        "adaptation_env_steps": 1233120,
        "provider_requests": 822,
        "sampled_episodes_by_purpose": {
          "heldout": 30,
          "learning": 14,
          "search": 6095,
          "validation": 68
        }
      }
    },
    {
      "task": "Relocate",
      "fixed": {
        "successes": 12,
        "episodes": 30,
        "search_rollouts": 394,
        "adaptation_env_steps": 115570,
        "provider_requests": 199
      },
      "random": {
        "successes": 30,
        "episodes": 30,
        "search_rollouts": 1312,
        "adaptation_env_steps": 275225,
        "provider_requests": 229,
        "sampled_episodes_by_purpose": {
          "heldout": 30,
          "learning": 18,
          "search": 1312,
          "validation": 55
        }
      }
    }
  ],
  "protocols": {
    "fixed": "seq10 L3+papers: learn each of 10 fixed starts in order and pass the cumulative fixed prefix.",
    "random": "adroit_seq10_l3pr: fresh starts for learning, search, and evaluation; stage k needs k successes in one frozen-policy evaluation. This archived campaign predates incoming-transfer tests."
  },
  "nominal_limits_per_stage": {
    "provider_requests": 300,
    "tool_calls": 1000,
    "learning_resets": 100,
    "evaluation_submissions": 250,
    "search_rollouts": 20000,
    "wall_hours": 12
  },
  "caveats": [
    "One training chain per task and protocol; this is not a replicated causal ablation.",
    "Training guidance, reset protocol, held-out seed sets, and actual interaction costs differ.",
    "Nominal stage limits match, but actual search and adaptation costs are greater for the random-reset campaign.",
    "Both protocols increase the evaluation batch size across stages; the fixed condition reuses a growing fixed prefix and the random condition samples fresh starts.",
    "The random-reset data have only a final held-out test; no held-out learning curve for stages 1\u20139 is implied.",
    "Outcomes concern native simulator reset distributions, not real-world or arbitrary perturbation robustness."
  ],
  "provenance": {
    "repository": "Schema-Robo-v7-r1",
    "sources": [
      {
        "path": "docs/data/astra_vs_opus.json",
        "sha256": "6adfa4050b93b69b5646da34534d7325aade8daff6e7385186754cf5fa977af1"
      },
      {
        "path": "experiment_results/adroit_seq10_l3pr/report.json",
        "sha256": "48c88acbb9d130aad72909833bb9d12443d8a8b7d3e1d86ee7a6567ac3c7a64c"
      }
    ],
    "fixed_crosscheck": "docs/research.md section 5.2",
    "random_crosscheck": "docs/figures/adroit_seq10_l3pr/REPORT.md",
    "random_protocol_source": "experiment_results/adroit_seq10_l3pr/inputs/door_po_seq10_L3PR.json",
    "fixed_protocol_source": "configs/c4/seq10/door_po_L3P_model_driven_papers.json"
  }
}