{
  "schema_version": 2,
  "seeds_by_adapter": {
    "coral": [
      42,
      43,
      44
    ]
  },
  "budget": "paper-specific",
  "papers": {
    "coral": {
      "manifest": {
        "adapter_key": "coral",
        "arxiv_id": "2609.02730",
        "title": "CORAL: An LLM-Native Harness for Production Recommender Systems",
        "paper_url": "https://arxiv.org/abs/2609.02730",
        "track": "recommendation",
        "organization": "Meta AI",
        "published": "2026-09-02",
        "code_url": null,
        "topics": [
          "auto-research",
          "agent",
          "retrieval-budget",
          "constrained-optimization"
        ],
        "local_code_dir": "src/auto_research/reproductions/coral",
        "fidelity": "core_mechanism",
        "evaluation_tier": "l2_public_dataset",
        "datasets": [
          "MovieLens 100K"
        ],
        "baseline": "fixed equal retrieval allocation",
        "metrics": [
          "hit_at_10",
          "ndcg_at_10",
          "budget feasibility",
          "cycle effect"
        ],
        "default_seeds": [
          42,
          43,
          44
        ],
        "budget": "adapter-defined fixed budget",
        "device_capabilities": [
          "cpu"
        ],
        "requires_gpu_validation": false,
        "gpu_validation_artifact": null,
        "online_evidence": [
          {
            "product": "Meta large-scale video service",
            "metric": "video-viewing sessions",
            "lift_percent": 0.16,
            "traffic": "millions-user online A/B; deployed configuration",
            "source_url": "https://arxiv.org/pdf/2609.02730v1",
            "source_location": "Section 5.1, Table 1",
            "retrieved_at": "2026-09-06"
          },
          {
            "product": "Meta serving-capacity service",
            "metric": "annualized capacity savings",
            "lift_percent": 44.0,
            "traffic": "second A/B round increased first-round savings by 44%; engagement neutral",
            "source_url": "https://arxiv.org/pdf/2609.02730v1",
            "source_location": "Section 5.2",
            "retrieved_at": "2026-09-06"
          }
        ],
        "selection_exception": null,
        "evolve_operators": [
          "controller:coral-constrained-feedback-loop"
        ]
      },
      "seed_results": [
        {
          "paper": {
            "arxiv_id": "2609.02730",
            "title": "CORAL: An LLM-Native Harness for Production Recommender Systems",
            "url": "https://arxiv.org/abs/2609.02730",
            "organization": "Meta AI"
          },
          "dataset": {
            "name": "MovieLens 100K",
            "users": 220,
            "items": 360
          },
          "setup": {
            "adapter": "coral",
            "same_split_and_candidates": true
          },
          "baseline": {
            "name": "fixed equal retrieval allocation",
            "hit_at_10": 0.10909090909090909,
            "ndcg_at_10": 0.05400546778896325,
            "fresh_hit_at_10": 0.031746031746031744,
            "head_share_at_10": 0.23772727272727273
          },
          "method": {
            "name": "five-cycle CORAL allocation (blend=0.3)",
            "hit_at_10": 0.1,
            "ndcg_at_10": 0.05021363472290861,
            "fresh_hit_at_10": 0.031746031746031744,
            "head_share_at_10": 0.23045454545454547
          },
          "relative": {
            "hit_at_10_percent": -8.333333333333325,
            "ndcg_at_10_percent": -7.021202150996931,
            "fresh_hit_at_10_percent": 0.0,
            "head_share_at_10_percent": -3.059273422562138
          },
          "stages": {
            "cycles": 5,
            "memory_horizon": 3,
            "all_budget_feasible": true,
            "trace": [
              {
                "cycle": 1,
                "configuration": [
                  1.1551439011387066,
                  1.107160638020637,
                  0.9689604967557849,
                  0.7687349640848716
                ],
                "effect": 0.04564307309317406,
                "budget": 4.0
              },
              {
                "cycle": 2,
                "configuration": [
                  1.2715538077536952,
                  1.1838292026174924,
                  0.9494978587289522,
                  0.5951191308998603
                ],
                "effect": 0.03404820053984758,
                "budget": 4.0
              },
              {
                "cycle": 3,
                "configuration": [
                  1.3878533634937729,
                  1.2566137899775471,
                  0.9339563015989184,
                  0.4215765449297618
                ],
                "effect": 0.033849244859569964,
                "budget": 4.0
              },
              {
                "cycle": 4,
                "configuration": [
                  1.50366950553129,
                  1.3250024150658977,
                  0.9213280794028128,
                  0.25
                ],
                "effect": 0.03341295175710572,
                "budget": 4.0
              },
              {
                "cycle": 5,
                "configuration": [
                  1.5622879231527524,
                  1.3323314740416796,
                  0.8553806028055684,
                  0.25
                ],
                "effect": 0.005453075788960237,
                "budget": 4.0
              }
            ]
          },
          "paper_results": {
            "watch_time_percent": 0.15,
            "sessions_percent": 0.16,
            "new_low_signal_sessions_percent": 0.23,
            "second_round_savings_increase_percent": 44.0
          },
          "scope": "公开 MovieLens 上运行固定控制流、三周期记忆、约束投影和结果反馈；不接入 Meta 私有遥测、LLM 或线上实验系统。",
          "runtime": {
            "requested_device": "auto",
            "cpu_threads": null,
            "platform": "Darwin arm64",
            "resolved_device": "cpu",
            "torch_version": "2.13.0",
            "accelerator": "arm"
          },
          "seed": 42,
          "schema_version": 2,
          "manifest": {
            "adapter_key": "coral",
            "arxiv_id": "2609.02730",
            "title": "CORAL: An LLM-Native Harness for Production Recommender Systems",
            "paper_url": "https://arxiv.org/abs/2609.02730",
            "track": "recommendation",
            "organization": "Meta AI",
            "published": "2026-09-02",
            "code_url": null,
            "topics": [
              "auto-research",
              "agent",
              "retrieval-budget",
              "constrained-optimization"
            ],
            "local_code_dir": "src/auto_research/reproductions/coral",
            "fidelity": "core_mechanism",
            "evaluation_tier": "l2_public_dataset",
            "datasets": [
              "MovieLens 100K"
            ],
            "baseline": "fixed equal retrieval allocation",
            "metrics": [
              "hit_at_10",
              "ndcg_at_10",
              "budget feasibility",
              "cycle effect"
            ],
            "default_seeds": [
              42,
              43,
              44
            ],
            "budget": "adapter-defined fixed budget",
            "device_capabilities": [
              "cpu"
            ],
            "requires_gpu_validation": false,
            "gpu_validation_artifact": null,
            "online_evidence": [
              {
                "product": "Meta large-scale video service",
                "metric": "video-viewing sessions",
                "lift_percent": 0.16,
                "traffic": "millions-user online A/B; deployed configuration",
                "source_url": "https://arxiv.org/pdf/2609.02730v1",
                "source_location": "Section 5.1, Table 1",
                "retrieved_at": "2026-09-06"
              },
              {
                "product": "Meta serving-capacity service",
                "metric": "annualized capacity savings",
                "lift_percent": 44.0,
                "traffic": "second A/B round increased first-round savings by 44%; engagement neutral",
                "source_url": "https://arxiv.org/pdf/2609.02730v1",
                "source_location": "Section 5.2",
                "retrieved_at": "2026-09-06"
              }
            ],
            "selection_exception": null,
            "evolve_operators": [
              "controller:coral-constrained-feedback-loop"
            ]
          },
          "provenance": {
            "created_at": "2026-09-05T19:38:33.597719+00:00",
            "code_commit": "6d97d3b2028b651e46229d2596e866d2a6f567c9",
            "python": "3.12.9",
            "platform": "macOS-26.5.2-arm64-arm-64bit",
            "dataset_dir": "/Users/bytedance/Documents/git_daiwk/auto-research/data",
            "dataset_fingerprint": "e66c261317a7b3179720ef3e3f5f79d7a204b135d1f217e2911b153ad2ce50a0",
            "packages": {
              "auto-research": "0.1.0",
              "numpy": "2.5.2",
              "torch": "2.13.0",
              "transformers": "5.14.1"
            }
          },
          "evaluation_protocol": {
            "tier": "l2_public_dataset",
            "tier_label": "L2 公开数据集训练",
            "seeds": [
              42
            ],
            "budget": "paper-specific",
            "formal_comparison": false,
            "claim_policy": "single/few-seed smoke result; do not claim a stable improvement"
          }
        },
        {
          "paper": {
            "arxiv_id": "2609.02730",
            "title": "CORAL: An LLM-Native Harness for Production Recommender Systems",
            "url": "https://arxiv.org/abs/2609.02730",
            "organization": "Meta AI"
          },
          "dataset": {
            "name": "MovieLens 100K",
            "users": 220,
            "items": 360
          },
          "setup": {
            "adapter": "coral",
            "same_split_and_candidates": true
          },
          "baseline": {
            "name": "fixed equal retrieval allocation",
            "hit_at_10": 0.10909090909090909,
            "ndcg_at_10": 0.05400546778896325,
            "fresh_hit_at_10": 0.031746031746031744,
            "head_share_at_10": 0.23772727272727273
          },
          "method": {
            "name": "five-cycle CORAL allocation (blend=0.9)",
            "hit_at_10": 0.09090909090909091,
            "ndcg_at_10": 0.04580523905657015,
            "fresh_hit_at_10": 0.015873015873015872,
            "head_share_at_10": 0.22727272727272727
          },
          "relative": {
            "hit_at_10_percent": -16.66666666666666,
            "ndcg_at_10_percent": -15.18407129521972,
            "fresh_hit_at_10_percent": -50.0,
            "head_share_at_10_percent": -4.397705544933084
          },
          "stages": {
            "cycles": 5,
            "memory_horizon": 3,
            "all_budget_feasible": true,
            "trace": [
              {
                "cycle": 1,
                "configuration": [
                  1.1465989548247164,
                  1.1204842325012874,
                  0.9621808741456788,
                  0.7707359385283175
                ],
                "effect": 0.05002072930924285,
                "budget": 4.0
              },
              {
                "cycle": 2,
                "configuration": [
                  1.2566952886464655,
                  1.2074370880936784,
                  0.9373267150639234,
                  0.5985409081959325
                ],
                "effect": 0.0372672142863525,
                "budget": 4.0
              },
              {
                "cycle": 3,
                "configuration": [
                  1.3667991682734866,
                  1.2907893640104764,
                  0.9160946613080732,
                  0.42631680640796343
                ],
                "effect": 0.03699984429122709,
                "budget": 4.0
              },
              {
                "cycle": 4,
                "configuration": [
                  1.4774059413153626,
                  1.3709561004461677,
                  0.898348954310779,
                  0.2532890039276905
                ],
                "effect": 0.03693717037569675,
                "budget": 4.0
              },
              {
                "cycle": 5,
                "configuration": [
                  1.531515130760818,
                  1.3909274067292117,
                  0.8275574625099709,
                  0.25
                ],
                "effect": 0.006896004126088141,
                "budget": 4.0
              }
            ]
          },
          "paper_results": {
            "watch_time_percent": 0.15,
            "sessions_percent": 0.16,
            "new_low_signal_sessions_percent": 0.23,
            "second_round_savings_increase_percent": 44.0
          },
          "scope": "公开 MovieLens 上运行固定控制流、三周期记忆、约束投影和结果反馈；不接入 Meta 私有遥测、LLM 或线上实验系统。",
          "runtime": {
            "requested_device": "auto",
            "cpu_threads": null,
            "platform": "Darwin arm64",
            "resolved_device": "cpu",
            "torch_version": "2.13.0",
            "accelerator": "arm"
          },
          "seed": 43,
          "schema_version": 2,
          "manifest": {
            "adapter_key": "coral",
            "arxiv_id": "2609.02730",
            "title": "CORAL: An LLM-Native Harness for Production Recommender Systems",
            "paper_url": "https://arxiv.org/abs/2609.02730",
            "track": "recommendation",
            "organization": "Meta AI",
            "published": "2026-09-02",
            "code_url": null,
            "topics": [
              "auto-research",
              "agent",
              "retrieval-budget",
              "constrained-optimization"
            ],
            "local_code_dir": "src/auto_research/reproductions/coral",
            "fidelity": "core_mechanism",
            "evaluation_tier": "l2_public_dataset",
            "datasets": [
              "MovieLens 100K"
            ],
            "baseline": "fixed equal retrieval allocation",
            "metrics": [
              "hit_at_10",
              "ndcg_at_10",
              "budget feasibility",
              "cycle effect"
            ],
            "default_seeds": [
              42,
              43,
              44
            ],
            "budget": "adapter-defined fixed budget",
            "device_capabilities": [
              "cpu"
            ],
            "requires_gpu_validation": false,
            "gpu_validation_artifact": null,
            "online_evidence": [
              {
                "product": "Meta large-scale video service",
                "metric": "video-viewing sessions",
                "lift_percent": 0.16,
                "traffic": "millions-user online A/B; deployed configuration",
                "source_url": "https://arxiv.org/pdf/2609.02730v1",
                "source_location": "Section 5.1, Table 1",
                "retrieved_at": "2026-09-06"
              },
              {
                "product": "Meta serving-capacity service",
                "metric": "annualized capacity savings",
                "lift_percent": 44.0,
                "traffic": "second A/B round increased first-round savings by 44%; engagement neutral",
                "source_url": "https://arxiv.org/pdf/2609.02730v1",
                "source_location": "Section 5.2",
                "retrieved_at": "2026-09-06"
              }
            ],
            "selection_exception": null,
            "evolve_operators": [
              "controller:coral-constrained-feedback-loop"
            ]
          },
          "provenance": {
            "created_at": "2026-09-05T19:38:33.941861+00:00",
            "code_commit": "6d97d3b2028b651e46229d2596e866d2a6f567c9",
            "python": "3.12.9",
            "platform": "macOS-26.5.2-arm64-arm-64bit",
            "dataset_dir": "/Users/bytedance/Documents/git_daiwk/auto-research/data",
            "dataset_fingerprint": "e66c261317a7b3179720ef3e3f5f79d7a204b135d1f217e2911b153ad2ce50a0",
            "packages": {
              "auto-research": "0.1.0",
              "numpy": "2.5.2",
              "torch": "2.13.0",
              "transformers": "5.14.1"
            }
          },
          "evaluation_protocol": {
            "tier": "l2_public_dataset",
            "tier_label": "L2 公开数据集训练",
            "seeds": [
              43
            ],
            "budget": "paper-specific",
            "formal_comparison": false,
            "claim_policy": "single/few-seed smoke result; do not claim a stable improvement"
          }
        },
        {
          "paper": {
            "arxiv_id": "2609.02730",
            "title": "CORAL: An LLM-Native Harness for Production Recommender Systems",
            "url": "https://arxiv.org/abs/2609.02730",
            "organization": "Meta AI"
          },
          "dataset": {
            "name": "MovieLens 100K",
            "users": 220,
            "items": 360
          },
          "setup": {
            "adapter": "coral",
            "same_split_and_candidates": true
          },
          "baseline": {
            "name": "fixed equal retrieval allocation",
            "hit_at_10": 0.10909090909090909,
            "ndcg_at_10": 0.05400546778896325,
            "fresh_hit_at_10": 0.031746031746031744,
            "head_share_at_10": 0.23772727272727273
          },
          "method": {
            "name": "five-cycle CORAL allocation (blend=1.0)",
            "hit_at_10": 0.1,
            "ndcg_at_10": 0.053887345968488334,
            "fresh_hit_at_10": 0.015873015873015872,
            "head_share_at_10": 0.21954545454545454
          },
          "relative": {
            "hit_at_10_percent": -8.333333333333325,
            "ndcg_at_10_percent": -0.2187219652211878,
            "fresh_hit_at_10_percent": -50.0,
            "head_share_at_10_percent": -7.648183556405356
          },
          "stages": {
            "cycles": 5,
            "memory_horizon": 3,
            "all_budget_feasible": true,
            "trace": [
              {
                "cycle": 1,
                "configuration": [
                  1.1559477015674835,
                  1.1114904746757255,
                  0.9597918721350773,
                  0.7727699516217132
                ],
                "effect": 0.04691773037540472,
                "budget": 3.999999999999999
              },
              {
                "cycle": 2,
                "configuration": [
                  1.273022433946683,
                  1.1915275854817484,
                  0.9333875247557902,
                  0.6020624558157779
                ],
                "effect": 0.03496947835993952,
                "budget": 3.999999999999999
              },
              {
                "cycle": 3,
                "configuration": [
                  1.3900537165871714,
                  1.2677965117909877,
                  0.9108528485604858,
                  0.4312969230613547
                ],
                "effect": 0.03473485418457347,
                "budget": 3.999999999999999
              },
              {
                "cycle": 4,
                "configuration": [
                  1.5075604872821324,
                  1.3406806664810298,
                  0.8920412296018816,
                  0.2597176166349555
                ],
                "effect": 0.034694439042462034,
                "budget": 3.999999999999999
              },
              {
                "cycle": 5,
                "configuration": [
                  1.5711376144435165,
                  1.355810423730501,
                  0.8230519618259828,
                  0.25
                ],
                "effect": 0.008005459101508516,
                "budget": 4.0
              }
            ]
          },
          "paper_results": {
            "watch_time_percent": 0.15,
            "sessions_percent": 0.16,
            "new_low_signal_sessions_percent": 0.23,
            "second_round_savings_increase_percent": 44.0
          },
          "scope": "公开 MovieLens 上运行固定控制流、三周期记忆、约束投影和结果反馈；不接入 Meta 私有遥测、LLM 或线上实验系统。",
          "runtime": {
            "requested_device": "auto",
            "cpu_threads": null,
            "platform": "Darwin arm64",
            "resolved_device": "cpu",
            "torch_version": "2.13.0",
            "accelerator": "arm"
          },
          "seed": 44,
          "schema_version": 2,
          "manifest": {
            "adapter_key": "coral",
            "arxiv_id": "2609.02730",
            "title": "CORAL: An LLM-Native Harness for Production Recommender Systems",
            "paper_url": "https://arxiv.org/abs/2609.02730",
            "track": "recommendation",
            "organization": "Meta AI",
            "published": "2026-09-02",
            "code_url": null,
            "topics": [
              "auto-research",
              "agent",
              "retrieval-budget",
              "constrained-optimization"
            ],
            "local_code_dir": "src/auto_research/reproductions/coral",
            "fidelity": "core_mechanism",
            "evaluation_tier": "l2_public_dataset",
            "datasets": [
              "MovieLens 100K"
            ],
            "baseline": "fixed equal retrieval allocation",
            "metrics": [
              "hit_at_10",
              "ndcg_at_10",
              "budget feasibility",
              "cycle effect"
            ],
            "default_seeds": [
              42,
              43,
              44
            ],
            "budget": "adapter-defined fixed budget",
            "device_capabilities": [
              "cpu"
            ],
            "requires_gpu_validation": false,
            "gpu_validation_artifact": null,
            "online_evidence": [
              {
                "product": "Meta large-scale video service",
                "metric": "video-viewing sessions",
                "lift_percent": 0.16,
                "traffic": "millions-user online A/B; deployed configuration",
                "source_url": "https://arxiv.org/pdf/2609.02730v1",
                "source_location": "Section 5.1, Table 1",
                "retrieved_at": "2026-09-06"
              },
              {
                "product": "Meta serving-capacity service",
                "metric": "annualized capacity savings",
                "lift_percent": 44.0,
                "traffic": "second A/B round increased first-round savings by 44%; engagement neutral",
                "source_url": "https://arxiv.org/pdf/2609.02730v1",
                "source_location": "Section 5.2",
                "retrieved_at": "2026-09-06"
              }
            ],
            "selection_exception": null,
            "evolve_operators": [
              "controller:coral-constrained-feedback-loop"
            ]
          },
          "provenance": {
            "created_at": "2026-09-05T19:38:34.140373+00:00",
            "code_commit": "6d97d3b2028b651e46229d2596e866d2a6f567c9",
            "python": "3.12.9",
            "platform": "macOS-26.5.2-arm64-arm-64bit",
            "dataset_dir": "/Users/bytedance/Documents/git_daiwk/auto-research/data",
            "dataset_fingerprint": "e66c261317a7b3179720ef3e3f5f79d7a204b135d1f217e2911b153ad2ce50a0",
            "packages": {
              "auto-research": "0.1.0",
              "numpy": "2.5.2",
              "torch": "2.13.0",
              "transformers": "5.14.1"
            }
          },
          "evaluation_protocol": {
            "tier": "l2_public_dataset",
            "tier_label": "L2 公开数据集训练",
            "seeds": [
              44
            ],
            "budget": "paper-specific",
            "formal_comparison": false,
            "claim_policy": "single/few-seed smoke result; do not claim a stable improvement"
          }
        }
      ],
      "aggregate_metrics": {
        "dataset.users": {
          "mean": 220.0,
          "std": 0.0,
          "ci95": 0.0,
          "n": 3
        },
        "dataset.items": {
          "mean": 360.0,
          "std": 0.0,
          "ci95": 0.0,
          "n": 3
        },
        "baseline.hit_at_10": {
          "mean": 0.10909090909090909,
          "std": 0.0,
          "ci95": 0.0,
          "n": 3
        },
        "baseline.ndcg_at_10": {
          "mean": 0.05400546778896325,
          "std": 0.0,
          "ci95": 0.0,
          "n": 3
        },
        "baseline.fresh_hit_at_10": {
          "mean": 0.031746031746031744,
          "std": 0.0,
          "ci95": 0.0,
          "n": 3
        },
        "baseline.head_share_at_10": {
          "mean": 0.2377272727272727,
          "std": 0.0,
          "ci95": 0.0,
          "n": 3
        },
        "method.hit_at_10": {
          "mean": 0.09696969696969697,
          "std": 0.0052486388108147815,
          "ci95": 0.005939393939393941,
          "n": 3
        },
        "method.ndcg_at_10": {
          "mean": 0.049968739915989026,
          "std": 0.004046615021689247,
          "ci95": 0.004579175973274977,
          "n": 3
        },
        "method.fresh_hit_at_10": {
          "mean": 0.021164021164021163,
          "std": 0.009164289987136916,
          "ci95": 0.01037037037037037,
          "n": 3
        },
        "method.head_share_at_10": {
          "mean": 0.22575757575757574,
          "std": 0.005610154115628375,
          "ci95": 0.006348487018152525,
          "n": 3
        },
        "relative.hit_at_10_percent": {
          "mean": -11.111111111111105,
          "std": 4.811252243246883,
          "ci95": 5.4444444444444455,
          "n": 3
        },
        "relative.ndcg_at_10_percent": {
          "mean": -7.474665137145947,
          "std": 7.492972818052743,
          "ci95": 8.479096952125268,
          "n": 3
        },
        "relative.fresh_hit_at_10_percent": {
          "mean": -33.333333333333336,
          "std": 28.867513459481287,
          "ci95": 32.666666666666664,
          "n": 3
        },
        "relative.head_share_at_10_percent": {
          "mean": -5.035054174633526,
          "std": 2.3599118650826814,
          "ci95": 2.670491671115785,
          "n": 3
        },
        "stages.cycles": {
          "mean": 5.0,
          "std": 0.0,
          "ci95": 0.0,
          "n": 3
        },
        "stages.memory_horizon": {
          "mean": 3.0,
          "std": 0.0,
          "ci95": 0.0,
          "n": 3
        },
        "paper_results.watch_time_percent": {
          "mean": 0.15,
          "std": 0.0,
          "ci95": 0.0,
          "n": 3
        },
        "paper_results.sessions_percent": {
          "mean": 0.16,
          "std": 0.0,
          "ci95": 0.0,
          "n": 3
        },
        "paper_results.new_low_signal_sessions_percent": {
          "mean": 0.23,
          "std": 0.0,
          "ci95": 0.0,
          "n": 3
        },
        "paper_results.second_round_savings_increase_percent": {
          "mean": 44.0,
          "std": 0.0,
          "ci95": 0.0,
          "n": 3
        }
      },
      "formal_comparison": true
    }
  },
  "manifest_ref": "reproduction:coral",
  "evaluation_protocol": {
    "tier": "l2_public_dataset",
    "seeds": [
      42,
      43,
      44
    ],
    "formal_comparison": true,
    "claim_policy": "formal multi-seed comparison"
  },
  "provenance": {
    "artifact_path": "docs/reproductions/2609.02730-coral/metrics/public-seeds42-44.json",
    "dataset_fingerprint": "not recorded in historical artifact"
  }
}