{
  "description": "Navigation index only; each paper keeps its own raw metric artifact.",
  "seed": 42,
  "papers": [
    {
      "batch": "B07",
      "domain": "foundation-models",
      "key": "tcab",
      "arxiv_id": "2608.12831",
      "seed": 42,
      "metric_path": "docs/reproductions/2608.12831-tcab/metrics/public-seed42.json",
      "local_result": {
        "baseline": {
          "name": "independent A/B/n feedback",
          "accuracy": 1.0,
          "relative_cost": 1.0,
          "reward_queries": 480.0
        },
        "method": {
          "name": "tcab",
          "accuracy": 1.0,
          "memory_fraction": 0.625,
          "relative_cost": 0.625,
          "reward_queries": 300.0,
          "independent_queries": 480.0
        }
      }
    },
    {
      "batch": "B07",
      "domain": "foundation-models",
      "key": "olmpool-long-context",
      "arxiv_id": "2608.10296",
      "seed": 42,
      "metric_path": "docs/reproductions/2608.10296-olmpool-long-context/metrics/public-seed42.json",
      "local_result": {
        "baseline": {
          "name": "recent-window attention",
          "accuracy": 0.078125,
          "memory_fraction": 0.25,
          "relative_cost": 1.0
        },
        "method": {
          "name": "olmpool-long-context",
          "accuracy": 1.0,
          "memory_fraction": 1.0,
          "relative_cost": 1.0,
          "controlled_choices": 4.0
        }
      }
    },
    {
      "batch": "B07",
      "domain": "foundation-models",
      "key": "distillcache",
      "arxiv_id": "2608.08878",
      "seed": 42,
      "metric_path": "docs/reproductions/2608.08878-distillcache/metrics/public-seed42.json",
      "local_result": {
        "baseline": {
          "name": "recent-window attention",
          "accuracy": 0.078125,
          "memory_fraction": 0.25,
          "relative_cost": 1.0
        },
        "method": {
          "name": "distillcache",
          "accuracy": 1.0,
          "memory_fraction": 0.25,
          "relative_cost": 0.54,
          "kl_guided_evictions": 4608.0
        }
      }
    },
    {
      "batch": "B07",
      "domain": "foundation-models",
      "key": "autonomy-heads",
      "arxiv_id": "2608.06849",
      "seed": 42,
      "metric_path": "docs/reproductions/2608.06849-autonomy-heads/metrics/public-seed42.json",
      "local_result": {
        "baseline": {
          "name": "recent-window attention",
          "accuracy": 0.078125,
          "memory_fraction": 0.25,
          "relative_cost": 1.0
        },
        "method": {
          "name": "autonomy-heads",
          "accuracy": 1.0,
          "memory_fraction": 0.5,
          "relative_cost": 0.59,
          "retrieval_heads": 2.0
        }
      }
    },
    {
      "batch": "B07",
      "domain": "foundation-models",
      "key": "physics-mm-pretraining",
      "arxiv_id": "2608.05000",
      "seed": 42,
      "metric_path": "docs/reproductions/2608.05000-physics-mm-pretraining/metrics/public-seed42.json",
      "local_result": {
        "baseline": {
          "name": "single-modality representation",
          "accuracy": 1.0,
          "relative_cost": 1.0,
          "modalities_used": 1.0
        },
        "method": {
          "name": "physics-mm-pretraining",
          "accuracy": 1.0,
          "memory_fraction": 1.0,
          "relative_cost": 0.05,
          "modality_synergy": 9.825241935367616
        }
      }
    },
    {
      "batch": "B07",
      "domain": "foundation-models",
      "key": "ttcd",
      "arxiv_id": "2608.01672",
      "seed": 42,
      "metric_path": "docs/reproductions/2608.01672-ttcd/metrics/public-seed42.json",
      "local_result": {
        "baseline": {
          "name": "recent-window attention",
          "accuracy": 0.078125,
          "memory_fraction": 0.25,
          "relative_cost": 1.0
        },
        "method": {
          "name": "ttcd",
          "accuracy": 0.859375,
          "memory_fraction": 0.25,
          "relative_cost": 0.42,
          "context_distillation_updates": 64.0
        }
      }
    },
    {
      "batch": "B07",
      "domain": "foundation-models",
      "key": "dart",
      "arxiv_id": "2608.02032",
      "seed": 42,
      "metric_path": "docs/reproductions/2608.02032-dart/metrics/public-seed42.json",
      "local_result": {
        "baseline": {
          "name": "recent-window attention",
          "accuracy": 0.078125,
          "memory_fraction": 0.25,
          "relative_cost": 1.0
        },
        "method": {
          "name": "dart",
          "accuracy": 0.109375,
          "memory_fraction": 0.125,
          "relative_cost": 0.31,
          "state_memories": 768.0
        }
      }
    },
    {
      "batch": "B07",
      "domain": "foundation-models",
      "key": "transmem",
      "arxiv_id": "2607.29032",
      "seed": 42,
      "metric_path": "docs/reproductions/2607.29032-transmem/metrics/public-seed42.json",
      "local_result": {
        "baseline": {
          "name": "recent-window attention",
          "accuracy": 0.078125,
          "memory_fraction": 0.25,
          "relative_cost": 1.0
        },
        "method": {
          "name": "transmem",
          "accuracy": 0.203125,
          "memory_fraction": 0.125,
          "relative_cost": 0.22,
          "latent_interventions": 64.0
        }
      }
    },
    {
      "batch": "B07",
      "domain": "foundation-models",
      "key": "c2kv",
      "arxiv_id": "2607.17715",
      "seed": 42,
      "metric_path": "docs/reproductions/2607.17715-c2kv/metrics/public-seed42.json",
      "local_result": {
        "baseline": {
          "name": "recent-window attention",
          "accuracy": 0.078125,
          "memory_fraction": 0.25,
          "relative_cost": 1.0
        },
        "method": {
          "name": "c2kv",
          "accuracy": 0.84375,
          "memory_fraction": 0.25,
          "relative_cost": 0.18,
          "composable_chunks": 1536.0
        }
      }
    },
    {
      "batch": "B08",
      "domain": "post-training",
      "key": "r2-opd",
      "arxiv_id": "2608.19408",
      "seed": 42,
      "metric_path": "docs/post-training/2608.19408-r2-opd/metrics/arithmetic-smoke-seed42.json",
      "local_result": {
        "baseline": {
          "accuracy": 0.1953125,
          "mean_reward": 0.359689058826365,
          "entropy": 1.791759469222057,
          "kl_from_reference": 0.0
        },
        "final": {
          "accuracy": 0.6640625,
          "mean_reward": 0.732082365939624,
          "entropy": 1.517513949749238,
          "kl_from_reference": 0.27424551947281656
        }
      }
    },
    {
      "batch": "B08",
      "domain": "post-training",
      "key": "sr-opsd",
      "arxiv_id": "2608.09745",
      "seed": 42,
      "metric_path": "docs/post-training/2608.09745-sr-opsd/metrics/arithmetic-smoke-seed42.json",
      "local_result": {
        "baseline": {
          "accuracy": 0.1953125,
          "mean_reward": 0.359689058826365,
          "entropy": 1.791759469222057,
          "kl_from_reference": 0.0
        },
        "final": {
          "accuracy": 0.6796875,
          "mean_reward": 0.7427208776867426,
          "entropy": 1.6135261163171033,
          "kl_from_reference": 0.1782333529049517
        }
      }
    },
    {
      "batch": "B08",
      "domain": "post-training",
      "key": "opd2",
      "arxiv_id": "2608.05802",
      "seed": 42,
      "metric_path": "docs/post-training/2608.05802-opd2/metrics/arithmetic-smoke-seed42.json",
      "local_result": {
        "baseline": {
          "accuracy": 0.1953125,
          "mean_reward": 0.359689058826365,
          "entropy": 1.791759469222057,
          "kl_from_reference": 0.0
        },
        "final": {
          "accuracy": 0.640625,
          "mean_reward": 0.71869501737969,
          "entropy": 1.6078316719094041,
          "kl_from_reference": 0.18392779731265102
        }
      }
    },
    {
      "batch": "B08",
      "domain": "post-training",
      "key": "causal-opd",
      "arxiv_id": "2608.03673",
      "seed": 42,
      "metric_path": "docs/post-training/2608.03673-causal-opd/metrics/arithmetic-smoke-seed42.json",
      "local_result": {
        "baseline": {
          "accuracy": 0.1953125,
          "mean_reward": 0.359689058826365,
          "entropy": 1.791759469222057,
          "kl_from_reference": 0.0
        },
        "final": {
          "accuracy": 0.6328125,
          "mean_reward": 0.714514301150916,
          "entropy": 1.432505672344747,
          "kl_from_reference": 0.35925379687730924
        }
      }
    },
    {
      "batch": "B08",
      "domain": "post-training",
      "key": "smopd",
      "arxiv_id": "2608.03092",
      "seed": 42,
      "metric_path": "docs/post-training/2608.03092-smopd/metrics/arithmetic-smoke-seed42.json",
      "local_result": {
        "baseline": {
          "accuracy": 0.1953125,
          "mean_reward": 0.359689058826365,
          "entropy": 1.791759469222057,
          "kl_from_reference": 0.0
        },
        "final": {
          "accuracy": 0.3125,
          "mean_reward": 0.4806971630154059,
          "entropy": 1.436163275620846,
          "kl_from_reference": 0.3555961936012087
        }
      }
    },
    {
      "batch": "B08",
      "domain": "post-training",
      "key": "rstg",
      "arxiv_id": "2608.00782",
      "seed": 42,
      "metric_path": "docs/post-training/2608.00782-rstg/metrics/arithmetic-smoke-seed42.json",
      "local_result": {
        "baseline": {
          "accuracy": 0.1953125,
          "mean_reward": 0.359689058826365,
          "entropy": 1.791759469222057,
          "kl_from_reference": 0.0
        },
        "final": {
          "accuracy": 0.6796875,
          "mean_reward": 0.739174087613786,
          "entropy": 1.6401155901760665,
          "kl_from_reference": 0.15164387904598867
        }
      }
    },
    {
      "batch": "B09",
      "domain": "post-training",
      "key": "sa-mrpo",
      "arxiv_id": "2608.16072",
      "seed": 42,
      "metric_path": "docs/post-training/2608.16072-sa-mrpo/metrics/arithmetic-smoke-seed42.json",
      "local_result": {
        "baseline": {
          "accuracy": 0.1953125,
          "mean_reward": 0.359689058826365,
          "entropy": 1.791759469222057,
          "kl_from_reference": 0.0
        },
        "final": {
          "accuracy": 0.28125,
          "mean_reward": 0.45564911291653953,
          "entropy": 1.6061716491012128,
          "kl_from_reference": 0.18558782012084224
        }
      }
    },
    {
      "batch": "B09",
      "domain": "post-training",
      "key": "rubric-dropout",
      "arxiv_id": "2608.11669",
      "seed": 42,
      "metric_path": "docs/post-training/2608.11669-rubric-dropout/metrics/arithmetic-smoke-seed42.json",
      "local_result": {
        "baseline": {
          "accuracy": 0.1953125,
          "mean_reward": 0.359689058826365,
          "entropy": 1.791759469222057,
          "kl_from_reference": 0.0
        },
        "final": {
          "accuracy": 0.28125,
          "mean_reward": 0.45564911291653953,
          "entropy": 1.27859848597865,
          "kl_from_reference": 0.5131609832434061
        }
      }
    },
    {
      "batch": "B09",
      "domain": "post-training",
      "key": "erils",
      "arxiv_id": "2608.01717",
      "seed": 42,
      "metric_path": "docs/post-training/2608.01717-erils/metrics/arithmetic-smoke-seed42.json",
      "local_result": {
        "baseline": {
          "accuracy": 0.1953125,
          "mean_reward": 0.359689058826365,
          "entropy": 1.791759469222057,
          "kl_from_reference": 0.0
        },
        "final": {
          "accuracy": 0.5703125,
          "mean_reward": 0.6722060847364586,
          "entropy": 1.6980760466104274,
          "kl_from_reference": 0.09368342261162747
        }
      }
    },
    {
      "batch": "B09",
      "domain": "post-training",
      "key": "crpo",
      "arxiv_id": "2607.28026",
      "seed": 42,
      "metric_path": "docs/post-training/2607.28026-crpo/metrics/arithmetic-smoke-seed42.json",
      "local_result": {
        "baseline": {
          "accuracy": 0.1953125,
          "mean_reward": 0.359689058826365,
          "entropy": 1.791759469222057,
          "kl_from_reference": 0.0
        },
        "final": {
          "accuracy": 0.6484375,
          "mean_reward": 0.7228659073790651,
          "entropy": 1.7866790808910213,
          "kl_from_reference": 0.0050803883310339935
        }
      }
    },
    {
      "batch": "B09",
      "domain": "post-training",
      "key": "serpo",
      "arxiv_id": "2607.26873",
      "seed": 42,
      "metric_path": "docs/post-training/2607.26873-serpo/metrics/arithmetic-smoke-seed42.json",
      "local_result": {
        "baseline": {
          "accuracy": 0.1953125,
          "mean_reward": 0.359689058826365,
          "entropy": 1.791759469222057,
          "kl_from_reference": 0.0
        },
        "final": {
          "accuracy": 0.328125,
          "mean_reward": 0.49280061959272564,
          "entropy": 1.557815149442971,
          "kl_from_reference": 0.23394431977908514
        }
      }
    },
    {
      "batch": "B09",
      "domain": "post-training",
      "key": "iso-rlvr",
      "arxiv_id": "2607.19331",
      "seed": 42,
      "metric_path": "docs/post-training/2607.19331-iso-rlvr/metrics/arithmetic-smoke-seed42.json",
      "local_result": {
        "baseline": {
          "accuracy": 0.1953125,
          "mean_reward": 0.359689058826365,
          "entropy": 1.791759469222057,
          "kl_from_reference": 0.0
        },
        "final": {
          "accuracy": 0.5390625,
          "mean_reward": 0.65004577576101,
          "entropy": 1.689655629751765,
          "kl_from_reference": 0.10210383947028875
        }
      }
    },
    {
      "batch": "B10",
      "domain": "agent-research",
      "key": "sapo",
      "arxiv_id": "2608.19842",
      "seed": 42,
      "metric_path": "docs/agent-research/2608.19842-sapo/metrics/planbench-mini-seed42.json",
      "local_result": {
        "baseline": {
          "answer_accuracy": 1.0,
          "plan_success": 1.0,
          "joint_success": 1.0,
          "average_cost": 64.5
        },
        "method": {
          "answer_accuracy": 1.0,
          "plan_success": 1.0,
          "joint_success": 1.0,
          "average_cost": 0.6199999999999997
        }
      }
    },
    {
      "batch": "B10",
      "domain": "agent-research",
      "key": "spade",
      "arxiv_id": "2608.19197",
      "seed": 42,
      "metric_path": "docs/agent-research/2608.19197-spade/metrics/planbench-mini-seed42.json",
      "local_result": {
        "baseline": {
          "answer_accuracy": 1.0,
          "plan_success": 1.0,
          "joint_success": 1.0,
          "average_cost": 64.5
        },
        "method": {
          "answer_accuracy": 1.0,
          "plan_success": 1.0,
          "joint_success": 1.0,
          "average_cost": 0.739999999999999
        }
      }
    },
    {
      "batch": "B10",
      "domain": "agent-research",
      "key": "rtpo",
      "arxiv_id": "2608.18682",
      "seed": 42,
      "metric_path": "docs/agent-research/2608.18682-rtpo/metrics/planbench-mini-seed42.json",
      "local_result": {
        "baseline": {
          "answer_accuracy": 1.0,
          "plan_success": 1.0,
          "joint_success": 1.0,
          "average_cost": 64.5
        },
        "method": {
          "answer_accuracy": 1.0,
          "plan_success": 1.0,
          "joint_success": 1.0,
          "average_cost": 0.6800000000000013
        }
      }
    },
    {
      "batch": "B10",
      "domain": "agent-research",
      "key": "planpo",
      "arxiv_id": "2608.17289",
      "seed": 42,
      "metric_path": "docs/agent-research/2608.17289-planpo/metrics/planbench-mini-seed42.json",
      "local_result": {
        "baseline": {
          "answer_accuracy": 1.0,
          "plan_success": 1.0,
          "joint_success": 1.0,
          "average_cost": 64.5
        },
        "method": {
          "answer_accuracy": 1.0,
          "plan_success": 1.0,
          "joint_success": 1.0,
          "average_cost": 0.6400000000000005
        }
      }
    },
    {
      "batch": "B10",
      "domain": "agent-research",
      "key": "trca",
      "arxiv_id": "2608.16156",
      "seed": 42,
      "metric_path": "docs/agent-research/2608.16156-trca/metrics/planbench-mini-seed42.json",
      "local_result": {
        "baseline": {
          "answer_accuracy": 1.0,
          "plan_success": 1.0,
          "joint_success": 1.0,
          "average_cost": 64.5
        },
        "method": {
          "answer_accuracy": 1.0,
          "plan_success": 1.0,
          "joint_success": 1.0,
          "average_cost": 0.709999999999999
        }
      }
    },
    {
      "batch": "B10",
      "domain": "agent-research",
      "key": "loongreflect",
      "arxiv_id": "2608.11967",
      "seed": 42,
      "metric_path": "docs/agent-research/2608.11967-loongreflect/metrics/planbench-mini-seed42.json",
      "local_result": {
        "baseline": {
          "answer_accuracy": 1.0,
          "plan_success": 1.0,
          "joint_success": 1.0,
          "average_cost": 64.5
        },
        "method": {
          "answer_accuracy": 1.0,
          "plan_success": 1.0,
          "joint_success": 1.0,
          "average_cost": 0.7699999999999998
        }
      }
    },
    {
      "batch": "B11",
      "domain": "agent-research",
      "key": "hymem",
      "arxiv_id": "2608.15703",
      "seed": 42,
      "metric_path": "docs/agent-research/2608.15703-hymem/metrics/planbench-mini-seed42.json",
      "local_result": {
        "baseline": {
          "answer_accuracy": 1.0,
          "plan_success": 1.0,
          "joint_success": 1.0,
          "average_cost": 64.5
        },
        "method": {
          "answer_accuracy": 1.0,
          "plan_success": 1.0,
          "joint_success": 1.0,
          "average_cost": 0.5200000000000014
        }
      }
    },
    {
      "batch": "B11",
      "domain": "agent-research",
      "key": "openloopevolve",
      "arxiv_id": "2608.09380",
      "seed": 42,
      "metric_path": "docs/agent-research/2608.09380-openloopevolve/metrics/planbench-mini-seed42.json",
      "local_result": {
        "baseline": {
          "answer_accuracy": 1.0,
          "plan_success": 1.0,
          "joint_success": 1.0,
          "average_cost": 64.5
        },
        "method": {
          "answer_accuracy": 1.0,
          "plan_success": 1.0,
          "joint_success": 1.0,
          "average_cost": 0.6899999999999988
        }
      }
    },
    {
      "batch": "B11",
      "domain": "agent-research",
      "key": "pmcoder",
      "arxiv_id": "2608.06811",
      "seed": 42,
      "metric_path": "docs/agent-research/2608.06811-pmcoder/metrics/planbench-mini-seed42.json",
      "local_result": {
        "baseline": {
          "answer_accuracy": 1.0,
          "plan_success": 1.0,
          "joint_success": 1.0,
          "average_cost": 64.5
        },
        "method": {
          "answer_accuracy": 1.0,
          "plan_success": 1.0,
          "joint_success": 1.0,
          "average_cost": 0.5799999999999987
        }
      }
    },
    {
      "batch": "B11",
      "domain": "agent-research",
      "key": "toollift",
      "arxiv_id": "2608.03468",
      "seed": 42,
      "metric_path": "docs/agent-research/2608.03468-toollift/metrics/planbench-mini-seed42.json",
      "local_result": {
        "baseline": {
          "answer_accuracy": 1.0,
          "plan_success": 1.0,
          "joint_success": 1.0,
          "average_cost": 64.5
        },
        "method": {
          "answer_accuracy": 1.0,
          "plan_success": 1.0,
          "joint_success": 1.0,
          "average_cost": 0.6099999999999995
        }
      }
    },
    {
      "batch": "B11",
      "domain": "agent-research",
      "key": "hyperagent",
      "arxiv_id": "2608.02650",
      "seed": 42,
      "metric_path": "docs/agent-research/2608.02650-hyperagent/metrics/planbench-mini-seed42.json",
      "local_result": {
        "baseline": {
          "answer_accuracy": 1.0,
          "plan_success": 1.0,
          "joint_success": 1.0,
          "average_cost": 64.5
        },
        "method": {
          "answer_accuracy": 1.0,
          "plan_success": 1.0,
          "joint_success": 1.0,
          "average_cost": 0.5499999999999988
        }
      }
    },
    {
      "batch": "B11",
      "domain": "agent-research",
      "key": "manta",
      "arxiv_id": "2607.28527",
      "seed": 42,
      "metric_path": "docs/agent-research/2607.28527-manta/metrics/planbench-mini-seed42.json",
      "local_result": {
        "baseline": {
          "answer_accuracy": 1.0,
          "plan_success": 1.0,
          "joint_success": 1.0,
          "average_cost": 64.5
        },
        "method": {
          "answer_accuracy": 1.0,
          "plan_success": 1.0,
          "joint_success": 1.0,
          "average_cost": 0.659999999999998
        }
      }
    }
  ],
  "schema_version": 2,
  "manifest_ref": "experiments:historical-b07-b11-seed42",
  "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/historical-b07-b11-seed42.json",
    "dataset_fingerprint": "not recorded in historical artifact",
    "historical_migration": "historical-metrics-v2-2026-08-09",
    "original_code_commit": "not recorded"
  }
}