{
  "schema_version": 2,
  "date": "2026-08-25",
  "scope": [
    "2608.23493-srpo",
    "2608.23311-erpo",
    "2608.23318-agent-g2",
    "2608.23041-autosaddler"
  ],
  "results": {
    "srpo": {
      "source": "../post-training/2608.23493-srpo/metrics/arithmetic-smoke-seed42.json",
      "baseline_accuracy": 0.1953125,
      "final_accuracy": 0.5703125,
      "reflection_patches_last_batch": 3
    },
    "erpo": {
      "source": "../post-training/2608.23311-erpo/metrics/arithmetic-smoke-seed42.json",
      "baseline_accuracy": 0.1953125,
      "final_accuracy": 0.609375,
      "query_kl_last_batch": 0.0056632246,
      "response_policy_kl_coefficient": 0.0
    },
    "agent-g2": {
      "source": "../agent-research/2608.23318-agent-g2/metrics/planbench-mini-seed42.json",
      "joint_success": 1.0,
      "average_cost": 0.8252638889,
      "gaussian_guidance_rollouts": 120
    },
    "autosaddler": {
      "source": "../agent-research/2608.23041-autosaddler/metrics/planbench-mini-seed42.json",
      "joint_success": 1.0,
      "average_cost": 0.544,
      "durable_updates": 12,
      "harness_reuses": 108
    }
  },
  "claim_boundary": "This is an index over per-paper L1 mechanism artifacts, not a cross-task leaderboard. Candidate-policy and deterministic Agent mini-suite results must not be presented as paper-scale LLM or official benchmark reproductions.",
  "manifest_ref": "experiments:latest-20260825-seed42",
  "evaluation_protocol": {
    "tier": "l1_mechanism",
    "seeds": [
      42
    ],
    "formal_comparison": false,
    "claim_policy": "single-seed mechanism verification; do not claim stable generalization"
  },
  "provenance": {
    "artifact_path": "docs/experiments/latest-20260825-seed42.json",
    "dataset_fingerprint": {
      "srpo": "arithmetic-smoke:100-steps:seed42",
      "erpo": "arithmetic-smoke:100-steps:seed42",
      "agent-g2": "planbench-mini:120-episodes:seed42",
      "autosaddler": "planbench-mini:120-episodes:seed42"
    }
  }
}