{
  "schema_version": 2,
  "method": "memcon",
  "manifest_ref": "agent-research:memcon",
  "dataset": "12 generated openly observable records; not a capability benchmark",
  "seeds": [
    42,
    43,
    44
  ],
  "diagnostic_only": true,
  "promotion_eligible": false,
  "evaluation_tier": "l1_mechanism",
  "evaluation_protocol": {
    "tier": "l1_mechanism",
    "seeds": [
      42,
      43,
      44
    ],
    "formal_comparison": false,
    "diagnostic_only": true,
    "claim_policy": "Controller/backend execution regression only; seeds vary record order, not independent capability tasks; do not claim agent improvement",
    "gold_access": false,
    "llm_calls": 0,
    "episodes_per_seed": 60,
    "memory_decisions_per_episode": 2,
    "feedback": "Presence of requested openly observable record in actual retrieval output"
  },
  "runs": [
    {
      "seed": 42,
      "retrieval_hit_rate": 0.8666666666666667,
      "mean_terminal_reward": 1.0799999999999998,
      "action_counts": [
        17,
        21,
        16,
        15,
        11,
        8,
        8,
        16,
        8
      ],
      "policy_updates": 60,
      "state_rows": 7,
      "memory_entries": 65
    },
    {
      "seed": 43,
      "retrieval_hit_rate": 0.8666666666666667,
      "mean_terminal_reward": 1.0799999999999998,
      "action_counts": [
        17,
        21,
        16,
        15,
        11,
        8,
        8,
        16,
        8
      ],
      "policy_updates": 60,
      "state_rows": 7,
      "memory_entries": 65
    },
    {
      "seed": 44,
      "retrieval_hit_rate": 0.8666666666666667,
      "mean_terminal_reward": 1.0799999999999998,
      "action_counts": [
        17,
        21,
        16,
        15,
        11,
        8,
        8,
        16,
        8
      ],
      "policy_updates": 60,
      "state_rows": 7,
      "memory_entries": 65
    }
  ],
  "provenance": {
    "artifact_path": "docs/agent-research/2607.13591-memcon/metrics/cpu-mechanism-seeds42-44.json",
    "dataset_fingerprint": "sha256:e847139b5f5a250d9e426dafc1c0c75c2c224bb9c0e26d5298b50cfe550b3a9a",
    "source_file_sha256": "d92415ef0105a7403623eaed1e8578a92a16ae83218b062c75d019411fc18655",
    "cross_python_revalidation": {
      "validated_on": "2026-10-14",
      "python_versions": [
        "3.11",
        "3.12"
      ],
      "seeds": [
        42,
        43,
        44
      ],
      "protocol_ref": "evaluation_protocol",
      "exact_json_match": true,
      "aggregation": "math.fsum(rewards) / episodes",
      "measurement_changed": false,
      "scope": "CPU diagnostic replay only; no new LLM/GPU run or capability claim"
    },
    "base_commit": "fcc81a35cbcc81ff0975d87004af007da99efdef",
    "command": "python -c 'from auto_research.agent_research.memcon import run_diagnostic; [run_diagnostic(seed=s, episodes=60) for s in (42,43,44)]'",
    "paper_revision": "2607.13591v3",
    "raw_predictions_committed": false,
    "checkpoint_committed": false
  }
}