{
  "schema_version": 2,
  "date": "2026-08-26",
  "scope": [
    "2608.24034-tagr",
    "2608.24053-wemm-embedding",
    "2608.24310-opd-search-plus",
    "2608.24696-opdvr",
    "2608.24870-spo-plus-plus",
    "2608.24747-skillforge",
    "2608.24114-ahead",
    "2608.24571-smith"
  ],
  "results": {
    "tagr": {
      "source": "../reproductions/2608.24034-tagr/metrics/movielens-1m-seed42.json",
      "baseline_ndcg_at_10": 0.1122723,
      "final_ndcg_at_10": 0.0821179,
      "relative_ndcg_at_10": -0.26858
    },
    "wemm-embedding": {
      "source": "../reproductions/2608.24053-wemm-embedding/metrics/movielens-1m-seed42.json",
      "stage1_recall_at_10": 0.561111,
      "refined_recall_at_10": 0.838889,
      "relative_recall_at_10": 0.49505
    },
    "opd-search-plus": {
      "source": "../post-training/2608.24310-opd-search-plus/metrics/arithmetic-smoke-seed42.json",
      "baseline_accuracy": 0.1953125,
      "final_accuracy": 0.6640625
    },
    "opdvr": {
      "source": "../post-training/2608.24696-opdvr/metrics/arithmetic-smoke-seed42.json",
      "baseline_accuracy": 0.1953125,
      "final_accuracy": 0.6640625
    },
    "spo-plus-plus": {
      "source": "../agent-research/2608.24870-spo-plus-plus/metrics/planbench-mini-seed42.json",
      "joint_success": 1.0,
      "average_cost": 0.64
    },
    "skillforge": {
      "source": "../agent-research/2608.24747-skillforge/metrics/planbench-mini-seed42.json",
      "joint_success": 1.0,
      "average_cost": 0.62
    },
    "ahead": {
      "source": "../agent-research/2608.24114-ahead/metrics/planbench-mini-seed42.json",
      "joint_success": 1.0,
      "average_cost": 0.616
    },
    "smith": {
      "source": "../agent-research/2608.24571-smith/metrics/planbench-mini-seed42.json",
      "joint_success": 1.0,
      "average_cost": 0.455
    }
  },
  "claim_boundary": "Index over per-paper L1/public-data mechanism artifacts; not a cross-task leaderboard or a paper-scale reproduction.",
  "manifest_ref": "experiments:latest-20260826-seed42",
  "evaluation_protocol": {
    "tier": "mixed_l1_and_public_dataset",
    "seeds": [
      42
    ],
    "formal_comparison": false,
    "claim_policy": "single-seed mechanism verification; paper and local results remain separate"
  },
  "provenance": {
    "artifact_path": "docs/experiments/latest-20260826-seed42.json",
    "dataset_fingerprint": {
      "tagr": "movielens-1m:seed42",
      "wemm-embedding": "movielens-1m-paired-views:seed42",
      "post-training": "arithmetic-smoke:100-steps:seed42",
      "agent": "planbench-mini:120-episodes:seed42"
    }
  }
}