{
  "generated_at": "2026-08-28",
  "total_tasks": 87,
  "source_worktrees": [
    "c1 v1/v5 (task-by-task-43)",
    "c2 v2 (6 reruns)",
    "c3 v1 (bucket-1/2/3 unblocks)",
    "mac-v2 (7 CCC-blocked tasks retried)"
  ],
  "tasks": [
    {
      "task": "3d-scan-calc",
      "category": "industrial-physical-systems",
      "subcategory": "3d-printing-mass-calculation",
      "source": "c1-v1",
      "status": "KILLED_saturated",
      "seed": 1.0,
      "cand_0001": null,
      "cand_0002": null,
      "cand_0003": null,
      "cand_0004": null,
      "best": 1.0,
      "best_tag": "seed",
      "iters_spent": null,
      "final_test": null,
      "delta": 0.0,
      "a0f": null,
      "a0f_delta": null
    },
    {
      "task": "ada-bathroom-plan-repair",
      "category": "industrial-physical-systems",
      "subcategory": "architectural-design",
      "source": "c3-v1",
      "status": "KILLED_ceiling",
      "seed": 0.4,
      "cand_0001": 0.5,
      "cand_0002": 0.8,
      "cand_0003": 1.0,
      "cand_0004": null,
      "best": 1.0,
      "best_tag": "cand_0003",
      "iters_spent": 3,
      "final_test": null,
      "delta": 0.6,
      "a0f": null,
      "a0f_delta": null
    },
    {
      "task": "adaptive-cruise-control",
      "category": "industrial-physical-systems",
      "subcategory": "vehicle-control",
      "source": "c1-tA",
      "status": "DONE",
      "seed": 0.7,
      "cand_0001": 0.7,
      "cand_0002": 0.9,
      "cand_0003": 1.0,
      "cand_0004": 1.0,
      "best": 1.0,
      "best_tag": "cand_0003",
      "iters_spent": null,
      "final_test": null,
      "delta": 0.30000000000000004,
      "a0f": 0.7,
      "a0f_delta": 0.0
    },
    {
      "task": "azure-bgp-oscillation-route-leak",
      "category": "software-engineering",
      "subcategory": "network-engineering",
      "source": "c3-v1",
      "status": "DONE",
      "seed": 0.0,
      "cand_0001": 1.0,
      "cand_0002": 1.0,
      "cand_0003": 1.0,
      "cand_0004": 1.0,
      "best": 1.0,
      "best_tag": "cand_0001",
      "iters_spent": 4,
      "final_test": 1.0,
      "delta": 1.0,
      "a0f": null,
      "a0f_delta": null
    },
    {
      "task": "bike-rebalance",
      "category": "mathematics-or-formal-reasoning",
      "subcategory": "operations-research-vehicle-routing",
      "source": "c3-v1",
      "status": "KILLED_ceiling",
      "seed": 0.1,
      "cand_0001": 0.8,
      "cand_0002": 1.0,
      "cand_0003": 1.0,
      "cand_0004": 1.0,
      "best": 1.0,
      "best_tag": "cand_0002",
      "iters_spent": 4,
      "final_test": null,
      "delta": 0.9,
      "a0f": null,
      "a0f_delta": null
    },
    {
      "task": "citation-check",
      "category": "office-white-collar",
      "subcategory": "academic-bibliography-verification",
      "source": "c1-v1",
      "status": "KILLED_saturated",
      "seed": 1.0,
      "cand_0001": null,
      "cand_0002": null,
      "cand_0003": null,
      "cand_0004": null,
      "best": 1.0,
      "best_tag": "seed",
      "iters_spent": null,
      "final_test": null,
      "delta": 0.0,
      "a0f": null,
      "a0f_delta": null
    },
    {
      "task": "civ6-adjacency-optimizer",
      "category": "mathematics-or-formal-reasoning",
      "subcategory": "combinatorial-optimization",
      "source": "c3-v1",
      "status": "DONE",
      "seed": 0.31,
      "cand_0001": 0.79,
      "cand_0002": 1.0,
      "cand_0003": 1.0,
      "cand_0004": 1.0,
      "best": 1.0,
      "best_tag": "cand_0002",
      "iters_spent": 4,
      "final_test": 1.0,
      "delta": 0.69,
      "a0f": 0.325,
      "a0f_delta": 0.015
    },
    {
      "task": "court-form-filling",
      "category": "office-white-collar",
      "subcategory": "legal-form-filling",
      "source": "c1-v1",
      "status": "DONE",
      "seed": 0.3,
      "cand_0001": 0.1,
      "cand_0002": 1.0,
      "cand_0003": null,
      "cand_0004": null,
      "best": 1.0,
      "best_tag": "cand_0002",
      "iters_spent": null,
      "final_test": null,
      "delta": 0.7,
      "a0f": null,
      "a0f_delta": null
    },
    {
      "task": "crystallographic-wyckoff-position-analysis",
      "category": "natural-science",
      "subcategory": "crystallography",
      "source": "c3-v1",
      "status": "DONE",
      "seed": 0.626,
      "cand_0001": 0.517,
      "cand_0002": 0.835,
      "cand_0003": 0.627,
      "cand_0004": 0.445,
      "best": 0.835,
      "best_tag": "cand_0002",
      "iters_spent": 4,
      "final_test": 0.317,
      "delta": 0.20899999999999996,
      "a0f": null,
      "a0f_delta": null
    },
    {
      "task": "dapt-intrusion-detection",
      "category": "cybersecurity",
      "subcategory": "network-intrusion-detection",
      "source": "c1-v1",
      "status": "KILLED_saturated",
      "seed": 1.0,
      "cand_0001": null,
      "cand_0002": null,
      "cand_0003": null,
      "cand_0004": null,
      "best": 1.0,
      "best_tag": "seed",
      "iters_spent": null,
      "final_test": null,
      "delta": 0.0,
      "a0f": null,
      "a0f_delta": null
    },
    {
      "task": "data-to-d3",
      "category": "software-engineering",
      "subcategory": "data-visualization-frontend",
      "source": "c1-v1",
      "status": "KILLED_ceiling",
      "seed": 0.5,
      "cand_0001": 0.9,
      "cand_0002": 1.0,
      "cand_0003": null,
      "cand_0004": null,
      "best": 1.0,
      "best_tag": "cand_0002",
      "iters_spent": null,
      "final_test": null,
      "delta": 0.5,
      "a0f": null,
      "a0f_delta": null
    },
    {
      "task": "debug-trl-grpo",
      "category": "software-engineering",
      "subcategory": "debugging",
      "source": "c3-v1",
      "status": "KILLED_ceiling",
      "seed": 0.425,
      "cand_0001": 0.865,
      "cand_0002": 0.9199999999999999,
      "cand_0003": 0.925,
      "cand_0004": 1.0,
      "best": 1.0,
      "best_tag": "cand_0004",
      "iters_spent": 4,
      "final_test": null,
      "delta": 0.575,
      "a0f": null,
      "a0f_delta": null
    },
    {
      "task": "dialogue-parser",
      "category": "software-engineering",
      "subcategory": "parser-implementation",
      "source": "c3-v1",
      "status": "KILLED_saturated",
      "seed": 1.0,
      "cand_0001": 1.0,
      "cand_0002": 1.0,
      "cand_0003": 1.0,
      "cand_0004": 1.0,
      "best": 1.0,
      "best_tag": "seed",
      "iters_spent": 4,
      "final_test": 1.0,
      "delta": 0.0,
      "a0f": null,
      "a0f_delta": null
    },
    {
      "task": "drone-planning-control",
      "category": "industrial-physical-systems",
      "subcategory": "robot-control",
      "source": "c1-v5",
      "status": "DONE",
      "seed": 0.7200001,
      "cand_0001": 0.0,
      "cand_0002": 0.0,
      "cand_0003": 0.3,
      "cand_0004": 0.3,
      "best": 0.7200001,
      "best_tag": "seed",
      "iters_spent": null,
      "final_test": null,
      "delta": 0.0,
      "a0f": null,
      "a0f_delta": null
    },
    {
      "task": "dynamic-object-aware-egomotion",
      "category": "industrial-physical-systems",
      "subcategory": "egomotion-and-dynamic-object-segmentation",
      "source": "c1-v1",
      "status": "KILLED_ceiling",
      "seed": 0.0,
      "cand_0001": 1.0,
      "cand_0002": null,
      "cand_0003": null,
      "cand_0004": null,
      "best": 1.0,
      "best_tag": "cand_0001",
      "iters_spent": null,
      "final_test": null,
      "delta": 1.0,
      "a0f": null,
      "a0f_delta": null
    },
    {
      "task": "earthquake-phase-association",
      "category": "natural-science",
      "subcategory": "seismology",
      "source": "mac-v2",
      "status": "NO_SIGNAL",
      "seed": 0.0,
      "cand_0001": 0.0,
      "cand_0002": 0.0,
      "cand_0003": 0.0,
      "cand_0004": 0.0,
      "best": 0.0,
      "best_tag": "seed",
      "iters_spent": 4,
      "final_test": 0.0,
      "delta": 0.0,
      "a0f": null,
      "a0f_delta": null
    },
    {
      "task": "earthquake-plate-calculation",
      "category": "natural-science",
      "subcategory": "geophysics-plate-tectonics",
      "source": "c3-v1",
      "status": "DONE",
      "seed": 0.9,
      "cand_0001": 1.0,
      "cand_0002": 1.0,
      "cand_0003": 1.0,
      "cand_0004": 1.0,
      "best": 1.0,
      "best_tag": "cand_0001",
      "iters_spent": 4,
      "final_test": 1.0,
      "delta": 0.09999999999999998,
      "a0f": 1.0,
      "a0f_delta": 0.1
    },
    {
      "task": "econ-detrending-correlation",
      "category": "finance-economics",
      "subcategory": "macroeconomic-time-series",
      "source": "c3-v1",
      "status": "KILLED_saturated",
      "seed": 1.0,
      "cand_0001": 1.0,
      "cand_0002": 1.0,
      "cand_0003": 1.0,
      "cand_0004": 1.0,
      "best": 1.0,
      "best_tag": "seed",
      "iters_spent": 4,
      "final_test": 1.0,
      "delta": 0.0,
      "a0f": null,
      "a0f_delta": null
    },
    {
      "task": "edit-pdf",
      "category": "office-white-collar",
      "subcategory": "pdf-editing",
      "source": "c1-v1",
      "status": "DONE",
      "seed": 0.9,
      "cand_0001": 0.9,
      "cand_0002": 1.0,
      "cand_0003": null,
      "cand_0004": null,
      "best": 1.0,
      "best_tag": "cand_0002",
      "iters_spent": null,
      "final_test": null,
      "delta": 0.09999999999999998,
      "a0f": null,
      "a0f_delta": null
    },
    {
      "task": "energy-ac-optimal-power-flow",
      "category": "industrial-physical-systems",
      "subcategory": "ac-optimal-power-flow",
      "source": "c2-v2",
      "status": "KILLED_val_1.0",
      "seed": 0.1,
      "cand_0001": 0.5,
      "cand_0002": 0.5,
      "cand_0003": 1.0,
      "cand_0004": null,
      "best": 1.0,
      "best_tag": "cand_0003",
      "iters_spent": 3,
      "final_test": null,
      "delta": 0.9,
      "a0f": null,
      "a0f_delta": null
    },
    {
      "task": "energy-market-pricing",
      "category": "industrial-physical-systems",
      "subcategory": "electricity-market-pricing",
      "source": "c2-v2",
      "status": "DONE",
      "seed": 0.0,
      "cand_0001": 0.5,
      "cand_0002": 0.9,
      "cand_0003": 0.6,
      "cand_0004": 0.8,
      "best": 0.9,
      "best_tag": "cand_0002",
      "iters_spent": 4,
      "final_test": 0.8,
      "delta": 0.9,
      "a0f": null,
      "a0f_delta": null
    },
    {
      "task": "energy-unit-commitment",
      "category": "industrial-physical-systems",
      "subcategory": "unit-commitment",
      "source": "c3-v1",
      "status": "KILLED_ceiling",
      "seed": 0.1,
      "cand_0001": 0.0,
      "cand_0002": 1.0,
      "cand_0003": 1.0,
      "cand_0004": 1.0,
      "best": 1.0,
      "best_tag": "cand_0002",
      "iters_spent": 4,
      "final_test": 0.1,
      "delta": 0.9,
      "a0f": 0.0,
      "a0f_delta": -0.1
    },
    {
      "task": "enterprise-information-search",
      "category": "office-white-collar",
      "subcategory": "business-reporting",
      "source": "c1-v1",
      "status": "KILLED_ceiling",
      "seed": 0.0,
      "cand_0001": 1.0,
      "cand_0002": 1.0,
      "cand_0003": 1.0,
      "cand_0004": null,
      "best": 1.0,
      "best_tag": "cand_0001",
      "iters_spent": null,
      "final_test": null,
      "delta": 1.0,
      "a0f": null,
      "a0f_delta": null
    },
    {
      "task": "exam-block-sequencing",
      "category": "mathematics-or-formal-reasoning",
      "subcategory": "mathematical-optimization",
      "source": "c3-v1",
      "status": "KILLED_ceiling",
      "seed": 0.1,
      "cand_0001": 0.9,
      "cand_0002": 1.0,
      "cand_0003": null,
      "cand_0004": null,
      "best": 1.0,
      "best_tag": "cand_0002",
      "iters_spent": 2,
      "final_test": null,
      "delta": 0.9,
      "a0f": null,
      "a0f_delta": null
    },
    {
      "task": "exceltable-in-ppt",
      "category": "office-white-collar",
      "subcategory": "presentation-editing",
      "source": "c1-v1",
      "status": "KILLED_ceiling",
      "seed": 0.9,
      "cand_0001": 1.0,
      "cand_0002": null,
      "cand_0003": null,
      "cand_0004": null,
      "best": 1.0,
      "best_tag": "cand_0001",
      "iters_spent": null,
      "final_test": null,
      "delta": 0.09999999999999998,
      "a0f": null,
      "a0f_delta": null
    },
    {
      "task": "exoplanet-detection-period",
      "category": "natural-science",
      "subcategory": "astronomy",
      "source": "c3-v1",
      "status": "KILLED_saturated",
      "seed": 1.0,
      "cand_0001": 1.0,
      "cand_0002": 0.9,
      "cand_0003": null,
      "cand_0004": null,
      "best": 1.0,
      "best_tag": "seed",
      "iters_spent": 2,
      "final_test": null,
      "delta": 0.0,
      "a0f": null,
      "a0f_delta": null
    },
    {
      "task": "financial-modeling-qa",
      "category": "finance-economics",
      "subcategory": "financial-modeling",
      "source": "c1-v5",
      "status": "NO_SIGNAL",
      "seed": 0.0,
      "cand_0001": 0.0,
      "cand_0002": 0.0,
      "cand_0003": 0.0,
      "cand_0004": 0.0,
      "best": 0.0,
      "best_tag": "seed",
      "iters_spent": null,
      "final_test": null,
      "delta": 0.0,
      "a0f": null,
      "a0f_delta": null
    },
    {
      "task": "fix-build-agentops",
      "category": "software-engineering",
      "subcategory": "build-repair",
      "source": "c3-v1",
      "status": "NO_SIGNAL",
      "seed": 0.0,
      "cand_0001": 0.0,
      "cand_0002": 0.0,
      "cand_0003": 0.0,
      "cand_0004": 0.0,
      "best": 0.0,
      "best_tag": "seed",
      "iters_spent": 4,
      "final_test": 0.0,
      "delta": 0.0,
      "a0f": null,
      "a0f_delta": null
    },
    {
      "task": "fix-build-google-auto",
      "category": "software-engineering",
      "subcategory": "build-repair",
      "source": "mac-v2",
      "status": "NO_SIGNAL",
      "seed": 0.0,
      "cand_0001": 0.0,
      "cand_0002": 0.0,
      "cand_0003": 0.0,
      "cand_0004": 0.0,
      "best": 0.0,
      "best_tag": "seed",
      "iters_spent": 4,
      "final_test": 0.0,
      "delta": 0.0,
      "a0f": null,
      "a0f_delta": null
    },
    {
      "task": "fix-druid-loophole-cve",
      "category": "cybersecurity",
      "subcategory": "vulnerability-analysis",
      "source": "c3-v1",
      "status": "NO_SIGNAL",
      "seed": 0.0,
      "cand_0001": 0.0,
      "cand_0002": 0.0,
      "cand_0003": 0.0,
      "cand_0004": 0.0,
      "best": 0.0,
      "best_tag": "seed",
      "iters_spent": 4,
      "final_test": 0.0,
      "delta": 0.0,
      "a0f": null,
      "a0f_delta": null
    },
    {
      "task": "fix-erlang-ssh-cve",
      "category": "cybersecurity",
      "subcategory": "vulnerability-analysis",
      "source": "c3-v1",
      "status": "DONE",
      "seed": 0.6,
      "cand_0001": 0.4,
      "cand_0002": 0.3,
      "cand_0003": 0.9,
      "cand_0004": 0.8,
      "best": 0.9,
      "best_tag": "cand_0003",
      "iters_spent": 4,
      "final_test": 1.0,
      "delta": 0.30000000000000004,
      "a0f": null,
      "a0f_delta": null
    },
    {
      "task": "fix-visual-stability",
      "category": "software-engineering",
      "subcategory": "performance-optimization",
      "source": "mac-v2",
      "status": "NO_SIGNAL",
      "seed": 0.0,
      "cand_0001": 0.0,
      "cand_0002": 0.0,
      "cand_0003": 0.0,
      "cand_0004": 0.0,
      "best": 0.0,
      "best_tag": "seed",
      "iters_spent": 4,
      "final_test": 0.0,
      "delta": 0.0,
      "a0f": null,
      "a0f_delta": null
    },
    {
      "task": "flink-query",
      "category": "software-engineering",
      "subcategory": "implementation",
      "source": "c3-v1",
      "status": "DONE",
      "seed": 0.0,
      "cand_0001": 0.1,
      "cand_0002": 0.1,
      "cand_0003": 0.5,
      "cand_0004": 0.9,
      "best": 0.9,
      "best_tag": "cand_0004",
      "iters_spent": 4,
      "final_test": 0.9,
      "delta": 0.9,
      "a0f": null,
      "a0f_delta": null
    },
    {
      "task": "flood-risk-analysis",
      "category": "natural-science",
      "subcategory": "hydrology",
      "source": "c1-v1",
      "status": "KILLED_saturated",
      "seed": 1.0,
      "cand_0001": null,
      "cand_0002": null,
      "cand_0003": null,
      "cand_0004": null,
      "best": 1.0,
      "best_tag": "seed",
      "iters_spent": null,
      "final_test": null,
      "delta": 0.0,
      "a0f": null,
      "a0f_delta": null
    },
    {
      "task": "glm-lake-mendota",
      "category": "natural-science",
      "subcategory": "hydrology",
      "source": "mac-v2",
      "status": "KILLED_saturated",
      "seed": 1.0,
      "cand_0001": null,
      "cand_0002": null,
      "cand_0003": null,
      "cand_0004": null,
      "best": 1.0,
      "best_tag": "seed",
      "iters_spent": 0,
      "final_test": null,
      "delta": 0.0,
      "a0f": null,
      "a0f_delta": null
    },
    {
      "task": "gravitational-wave-detection",
      "category": "natural-science",
      "subcategory": "astronomy",
      "source": "c3-v1",
      "status": "KILLED_saturated",
      "seed": 1.0,
      "cand_0001": 1.0,
      "cand_0002": 1.0,
      "cand_0003": 1.0,
      "cand_0004": null,
      "best": 1.0,
      "best_tag": "seed",
      "iters_spent": 3,
      "final_test": null,
      "delta": 0.0,
      "a0f": null,
      "a0f_delta": null
    },
    {
      "task": "grid-dispatch-operator",
      "category": "industrial-physical-systems",
      "subcategory": "grid-dispatch",
      "source": "c1-v1",
      "status": "DONE",
      "seed": 0.0,
      "cand_0001": 0.9,
      "cand_0002": 1.0,
      "cand_0003": null,
      "cand_0004": null,
      "best": 1.0,
      "best_tag": "cand_0002",
      "iters_spent": null,
      "final_test": null,
      "delta": 1.0,
      "a0f": null,
      "a0f_delta": null
    },
    {
      "task": "hvac-control",
      "category": "industrial-physical-systems",
      "subcategory": "control-systems",
      "source": "c1-v1",
      "status": "DONE",
      "seed": 1.0,
      "cand_0001": null,
      "cand_0002": null,
      "cand_0003": null,
      "cand_0004": null,
      "best": 1.0,
      "best_tag": "seed",
      "iters_spent": null,
      "final_test": null,
      "delta": 0.0,
      "a0f": 0.5,
      "a0f_delta": -0.5
    },
    {
      "task": "invoice-fraud-detection",
      "category": "finance-economics",
      "subcategory": "fraud-detection",
      "source": "c1-tA",
      "status": "DONE",
      "seed": 0.3,
      "cand_0001": 0.2,
      "cand_0002": 0.4,
      "cand_0003": 0.2,
      "cand_0004": 0.2,
      "best": 0.4,
      "best_tag": "cand_0002",
      "iters_spent": null,
      "final_test": null,
      "delta": 0.10000000000000003,
      "a0f": null,
      "a0f_delta": null
    },
    {
      "task": "jax-computing-basics",
      "category": "software-engineering",
      "subcategory": "library-api-usage",
      "source": "c1-v5",
      "status": "DONE",
      "seed": 0.9,
      "cand_0001": 0.0,
      "cand_0002": 0.0,
      "cand_0003": 0.4,
      "cand_0004": 0.1,
      "best": 0.9,
      "best_tag": "seed",
      "iters_spent": null,
      "final_test": null,
      "delta": 0.0,
      "a0f": null,
      "a0f_delta": null
    },
    {
      "task": "jpg-ocr-stat",
      "category": "office-white-collar",
      "subcategory": "ocr",
      "source": "c1-v1",
      "status": "KILLED_ceiling",
      "seed": 0.8,
      "cand_0001": 1.0,
      "cand_0002": null,
      "cand_0003": null,
      "cand_0004": null,
      "best": 1.0,
      "best_tag": "cand_0001",
      "iters_spent": null,
      "final_test": null,
      "delta": 0.19999999999999996,
      "a0f": 0.7,
      "a0f_delta": -0.1
    },
    {
      "task": "lab-unit-harmonization",
      "category": "natural-science",
      "subcategory": "lab-unit-harmonization",
      "source": "c3-v1",
      "status": "DONE",
      "seed": 0.9978999999999999,
      "cand_0001": 1.0,
      "cand_0002": 1.0,
      "cand_0003": 1.0,
      "cand_0004": 1.0,
      "best": 1.0,
      "best_tag": "cand_0001",
      "iters_spent": 4,
      "final_test": 1.0,
      "delta": 0.002100000000000102,
      "a0f": null,
      "a0f_delta": null
    },
    {
      "task": "lake-warming-attribution",
      "category": "natural-science",
      "subcategory": "hydrology",
      "source": "c1-v1",
      "status": "KILLED_ceiling",
      "seed": 0.1,
      "cand_0001": 0.6,
      "cand_0002": 1.0,
      "cand_0003": null,
      "cand_0004": null,
      "best": 1.0,
      "best_tag": "cand_0002",
      "iters_spent": null,
      "final_test": null,
      "delta": 0.9,
      "a0f": null,
      "a0f_delta": null
    },
    {
      "task": "latex-formula-extraction",
      "category": "office-white-collar",
      "subcategory": "pdf-formula-extraction",
      "source": "c3-v1",
      "status": "DONE",
      "seed": 0.2,
      "cand_0001": 0.0,
      "cand_0002": 0.5,
      "cand_0003": 0.7,
      "cand_0004": 0.8,
      "best": 0.8,
      "best_tag": "cand_0004",
      "iters_spent": 4,
      "final_test": 0.8,
      "delta": 0.6000000000000001,
      "a0f": null,
      "a0f_delta": null
    },
    {
      "task": "lean4-proof",
      "category": "mathematics-or-formal-reasoning",
      "subcategory": "formal-proof",
      "source": "c3-v1",
      "status": "DONE",
      "seed": 0.9,
      "cand_0001": 1.0,
      "cand_0002": 1.0,
      "cand_0003": 0.9,
      "cand_0004": 1.0,
      "best": 1.0,
      "best_tag": "cand_0001",
      "iters_spent": 4,
      "final_test": 1.0,
      "delta": 0.09999999999999998,
      "a0f": null,
      "a0f_delta": null
    },
    {
      "task": "llm-prefix-cache-replay",
      "category": "software-engineering",
      "subcategory": "performance-optimization",
      "source": "c3-v1",
      "status": "KILLED_saturated",
      "seed": 1.0,
      "cand_0001": 1.0,
      "cand_0002": 1.0,
      "cand_0003": 1.0,
      "cand_0004": 1.0,
      "best": 1.0,
      "best_tag": "seed",
      "iters_spent": 4,
      "final_test": 1.0,
      "delta": 0.0,
      "a0f": null,
      "a0f_delta": null
    },
    {
      "task": "manufacturing-codebook-normalization",
      "category": "industrial-physical-systems",
      "subcategory": "defect-analysis",
      "source": "c3-v1",
      "status": "KILLED_ceiling",
      "seed": 0.2,
      "cand_0001": 0.8,
      "cand_0002": 1.0,
      "cand_0003": 1.0,
      "cand_0004": 1.0,
      "best": 1.0,
      "best_tag": "cand_0002",
      "iters_spent": 4,
      "final_test": null,
      "delta": 0.8,
      "a0f": null,
      "a0f_delta": null
    },
    {
      "task": "manufacturing-equipment-maintenance",
      "category": "industrial-physical-systems",
      "subcategory": "maintenance",
      "source": "c3-v1",
      "status": "KILLED_ceiling",
      "seed": 0.7,
      "cand_0001": 0.9,
      "cand_0002": 0.8,
      "cand_0003": 1.0,
      "cand_0004": 1.0,
      "best": 1.0,
      "best_tag": "cand_0003",
      "iters_spent": 4,
      "final_test": null,
      "delta": 0.30000000000000004,
      "a0f": null,
      "a0f_delta": null
    },
    {
      "task": "manufacturing-fjsp-optimization",
      "category": "industrial-physical-systems",
      "subcategory": "production-scheduling",
      "source": "c3-v1",
      "status": "DONE",
      "seed": 0.8,
      "cand_0001": 1.0,
      "cand_0002": 1.0,
      "cand_0003": 1.0,
      "cand_0004": 1.0,
      "best": 1.0,
      "best_tag": "cand_0001",
      "iters_spent": 4,
      "final_test": 1.0,
      "delta": 0.19999999999999996,
      "a0f": null,
      "a0f_delta": null
    },
    {
      "task": "mario-coin-counting",
      "category": "media-content-production",
      "subcategory": "video-processing",
      "source": "c3-v1",
      "status": "KILLED_saturated",
      "seed": 1.0,
      "cand_0001": 1.0,
      "cand_0002": 1.0,
      "cand_0003": null,
      "cand_0004": null,
      "best": 1.0,
      "best_tag": "seed",
      "iters_spent": 2,
      "final_test": null,
      "delta": 0.0,
      "a0f": 0.1,
      "a0f_delta": -0.9
    },
    {
      "task": "mars-clouds-clustering",
      "category": "natural-science",
      "subcategory": "astronomy",
      "source": "c3-v1",
      "status": "KILLED_saturated",
      "seed": 1.0,
      "cand_0001": null,
      "cand_0002": null,
      "cand_0003": null,
      "cand_0004": null,
      "best": 1.0,
      "best_tag": "seed",
      "iters_spent": 0,
      "final_test": null,
      "delta": 0.0,
      "a0f": null,
      "a0f_delta": null
    },
    {
      "task": "multilingual-video-dubbing",
      "category": "media-content-production",
      "subcategory": "video-processing",
      "source": "c1-tA",
      "status": "DONE",
      "seed": 0.7,
      "cand_0001": 0.9,
      "cand_0002": 0.9,
      "cand_0003": 0.9,
      "cand_0004": 1.0,
      "best": 1.0,
      "best_tag": "cand_0004",
      "iters_spent": null,
      "final_test": null,
      "delta": 0.30000000000000004,
      "a0f": null,
      "a0f_delta": null
    },
    {
      "task": "offer-letter-generator",
      "category": "office-white-collar",
      "subcategory": "document-editing",
      "source": "mac-v2",
      "status": "KILLED_ceiling",
      "seed": 1.0,
      "cand_0001": 1.0,
      "cand_0002": 1.0,
      "cand_0003": 1.0,
      "cand_0004": 0.6666666666666666,
      "best": 1.0,
      "best_tag": "seed",
      "iters_spent": 4,
      "final_test": 1.0,
      "delta": 0.0,
      "a0f": null,
      "a0f_delta": null
    },
    {
      "task": "organize-messy-files",
      "category": "office-white-collar",
      "subcategory": "document-classification",
      "source": "c1-v1",
      "status": "DONE",
      "seed": 0.5,
      "cand_0001": 0.6,
      "cand_0002": 0.6,
      "cand_0003": 0.9,
      "cand_0004": 0.9,
      "best": 0.9,
      "best_tag": "cand_0003",
      "iters_spent": null,
      "final_test": null,
      "delta": 0.4,
      "a0f": 0.5,
      "a0f_delta": 0.0
    },
    {
      "task": "paper-anonymizer",
      "category": "office-white-collar",
      "subcategory": "document-editing",
      "source": "c1-v5",
      "status": "DONE",
      "seed": 0.6,
      "cand_0001": 0.0,
      "cand_0002": 0.0,
      "cand_0003": 0.2,
      "cand_0004": 0.0,
      "best": 0.6,
      "best_tag": "seed",
      "iters_spent": null,
      "final_test": null,
      "delta": 0.0,
      "a0f": null,
      "a0f_delta": null
    },
    {
      "task": "parallel-tfidf-search",
      "category": "software-engineering",
      "subcategory": "performance-optimization",
      "source": "c3-v1",
      "status": "KILLED_saturated",
      "seed": 1.0,
      "cand_0001": 1.0,
      "cand_0002": 0.9,
      "cand_0003": null,
      "cand_0004": null,
      "best": 1.0,
      "best_tag": "seed",
      "iters_spent": 2,
      "final_test": null,
      "delta": 0.0,
      "a0f": null,
      "a0f_delta": null
    },
    {
      "task": "paratransit-routing",
      "category": "mathematics-or-formal-reasoning",
      "subcategory": "mathematical-optimization",
      "source": "c3-v1",
      "status": "KILLED_ceiling",
      "seed": 0.0,
      "cand_0001": 1.0,
      "cand_0002": 0.9,
      "cand_0003": 1.0,
      "cand_0004": null,
      "best": 1.0,
      "best_tag": "cand_0001",
      "iters_spent": 3,
      "final_test": null,
      "delta": 1.0,
      "a0f": null,
      "a0f_delta": null
    },
    {
      "task": "pddl-airport-planning",
      "category": "mathematics-or-formal-reasoning",
      "subcategory": "formal-planning",
      "source": "c1-v1",
      "status": "DONE",
      "seed": 0.0,
      "cand_0001": 0.0,
      "cand_0002": 1.0,
      "cand_0003": null,
      "cand_0004": null,
      "best": 1.0,
      "best_tag": "cand_0002",
      "iters_spent": null,
      "final_test": null,
      "delta": 1.0,
      "a0f": null,
      "a0f_delta": null
    },
    {
      "task": "pddl-tpp-planning",
      "category": "mathematics-or-formal-reasoning",
      "subcategory": "formal-planning",
      "source": "c3-v1",
      "status": "KILLED_saturated",
      "seed": 1.0,
      "cand_0001": 1.0,
      "cand_0002": 1.0,
      "cand_0003": null,
      "cand_0004": null,
      "best": 1.0,
      "best_tag": "seed",
      "iters_spent": 2,
      "final_test": null,
      "delta": 0.0,
      "a0f": null,
      "a0f_delta": null
    },
    {
      "task": "pdf-excel-diff",
      "category": "office-white-collar",
      "subcategory": "spreadsheet-workflow",
      "source": "c1-v1",
      "status": "KILLED_ceiling",
      "seed": 0.9,
      "cand_0001": 1.0,
      "cand_0002": 1.0,
      "cand_0003": null,
      "cand_0004": null,
      "best": 1.0,
      "best_tag": "cand_0001",
      "iters_spent": null,
      "final_test": null,
      "delta": 0.09999999999999998,
      "a0f": null,
      "a0f_delta": null
    },
    {
      "task": "powerlifting-coef-calc",
      "category": "office-white-collar",
      "subcategory": "spreadsheet-workflow",
      "source": "c3-v1",
      "status": "KILLED_saturated",
      "seed": 1.0,
      "cand_0001": 1.0,
      "cand_0002": 1.0,
      "cand_0003": 1.0,
      "cand_0004": 1.0,
      "best": 1.0,
      "best_tag": "seed",
      "iters_spent": 4,
      "final_test": 1.0,
      "delta": 0.0,
      "a0f": null,
      "a0f_delta": null
    },
    {
      "task": "pptx-reference-formatting",
      "category": "office-white-collar",
      "subcategory": "presentation-editing",
      "source": "c1-v1",
      "status": "KILLED_saturated",
      "seed": 1.0,
      "cand_0001": null,
      "cand_0002": null,
      "cand_0003": null,
      "cand_0004": null,
      "best": 1.0,
      "best_tag": "seed",
      "iters_spent": null,
      "final_test": null,
      "delta": 0.0,
      "a0f": null,
      "a0f_delta": null
    },
    {
      "task": "protein-expression-analysis",
      "category": "natural-science",
      "subcategory": "protein-expression",
      "source": "c1-v1",
      "status": "KILLED_saturated",
      "seed": 1.0,
      "cand_0001": null,
      "cand_0002": null,
      "cand_0003": null,
      "cand_0004": null,
      "best": 1.0,
      "best_tag": "seed",
      "iters_spent": null,
      "final_test": null,
      "delta": 0.0,
      "a0f": null,
      "a0f_delta": null
    },
    {
      "task": "python-scala-translation",
      "category": "software-engineering",
      "subcategory": "code-translation",
      "source": "c1-v1",
      "status": "NO_SIGNAL",
      "seed": 0.0,
      "cand_0001": 0.0,
      "cand_0002": 0.0,
      "cand_0003": null,
      "cand_0004": null,
      "best": 0.0,
      "best_tag": "seed",
      "iters_spent": null,
      "final_test": null,
      "delta": 0.0,
      "a0f": null,
      "a0f_delta": null
    },
    {
      "task": "quantum-numerical-simulation",
      "category": "natural-science",
      "subcategory": "quantum-simulation",
      "source": "c3-v1",
      "status": "NO_SIGNAL",
      "seed": 0.0,
      "cand_0001": 0.0,
      "cand_0002": 0.0,
      "cand_0003": 0.0,
      "cand_0004": 0.0,
      "best": 0.0,
      "best_tag": "seed",
      "iters_spent": 4,
      "final_test": 0.0,
      "delta": 0.0,
      "a0f": null,
      "a0f_delta": null
    },
    {
      "task": "r2r-mpc-control",
      "category": "industrial-physical-systems",
      "subcategory": "control-systems",
      "source": "c1-v1",
      "status": "DONE",
      "seed": 0.3,
      "cand_0001": 1.0,
      "cand_0002": 1.0,
      "cand_0003": null,
      "cand_0004": null,
      "best": 1.0,
      "best_tag": "cand_0001",
      "iters_spent": null,
      "final_test": null,
      "delta": 0.7,
      "a0f": null,
      "a0f_delta": null
    },
    {
      "task": "radar-vital-signs",
      "category": "natural-science",
      "subcategory": "biomedical-analysis",
      "source": "c3-v1",
      "status": "KILLED_saturated",
      "seed": 1.0,
      "cand_0001": 1.0,
      "cand_0002": 1.0,
      "cand_0003": 1.0,
      "cand_0004": 1.0,
      "best": 1.0,
      "best_tag": "seed",
      "iters_spent": 4,
      "final_test": 1.0,
      "delta": 0.0,
      "a0f": null,
      "a0f_delta": null
    },
    {
      "task": "react-performance-debugging",
      "category": "software-engineering",
      "subcategory": "performance-optimization",
      "source": "c3-v1",
      "status": "KILLED_ceiling",
      "seed": 0.6,
      "cand_0001": 1.0,
      "cand_0002": 0.9,
      "cand_0003": 1.0,
      "cand_0004": 1.0,
      "best": 1.0,
      "best_tag": "cand_0001",
      "iters_spent": 4,
      "final_test": null,
      "delta": 0.4,
      "a0f": 0.4,
      "a0f_delta": -0.2
    },
    {
      "task": "reserves-at-risk-calc",
      "category": "finance-economics",
      "subcategory": "risk-analysis",
      "source": "c2-v2",
      "status": "DONE",
      "seed": 0.0,
      "cand_0001": 0.0,
      "cand_0002": 0.7,
      "cand_0003": 0.8,
      "cand_0004": 1.0,
      "best": 1.0,
      "best_tag": "cand_0004",
      "iters_spent": 4,
      "final_test": 1.0,
      "delta": 1.0,
      "a0f": null,
      "a0f_delta": null
    },
    {
      "task": "sales-pivot-analysis",
      "category": "office-white-collar",
      "subcategory": "spreadsheet-workflow",
      "source": "c1-v1",
      "status": "KILLED_ceiling",
      "seed": 0.0,
      "cand_0001": 1.0,
      "cand_0002": null,
      "cand_0003": null,
      "cand_0004": null,
      "best": 1.0,
      "best_tag": "cand_0001",
      "iters_spent": null,
      "final_test": null,
      "delta": 1.0,
      "a0f": null,
      "a0f_delta": null
    },
    {
      "task": "sec-financial-report",
      "category": "finance-economics",
      "subcategory": "sec-filings-analysis",
      "source": "c1-v1",
      "status": "DONE",
      "seed": 0.9,
      "cand_0001": 0.0,
      "cand_0002": 1.0,
      "cand_0003": null,
      "cand_0004": null,
      "best": 1.0,
      "best_tag": "cand_0002",
      "iters_spent": null,
      "final_test": null,
      "delta": 0.09999999999999998,
      "a0f": null,
      "a0f_delta": null
    },
    {
      "task": "seismic-phase-picking",
      "category": "natural-science",
      "subcategory": "seismology",
      "source": "mac-v2",
      "status": "NO_SIGNAL",
      "seed": 0.0,
      "cand_0001": 0.0,
      "cand_0002": 0.0,
      "cand_0003": 0.0,
      "cand_0004": 0.0,
      "best": 0.0,
      "best_tag": "seed",
      "iters_spent": 4,
      "final_test": 0.0,
      "delta": 0.0,
      "a0f": null,
      "a0f_delta": null
    },
    {
      "task": "setup-fuzzing-py",
      "category": "cybersecurity",
      "subcategory": "fuzzing",
      "source": "c3-v1",
      "status": "DONE",
      "seed": 0.364,
      "cand_0001": 0.0,
      "cand_0002": 0.264,
      "cand_0003": 0.513,
      "cand_0004": 0.549,
      "best": 0.549,
      "best_tag": "cand_0004",
      "iters_spent": 4,
      "final_test": 0.881,
      "delta": 0.18500000000000005,
      "a0f": 0.513,
      "a0f_delta": 0.149
    },
    {
      "task": "shock-analysis-demand",
      "category": "finance-economics",
      "subcategory": "macroeconomic-analysis",
      "source": "c2-v2",
      "status": "DONE",
      "seed": 0.0,
      "cand_0001": 0.0,
      "cand_0002": 0.0,
      "cand_0003": 0.4,
      "cand_0004": 0.9,
      "best": 0.9,
      "best_tag": "cand_0004",
      "iters_spent": 4,
      "final_test": 0.9,
      "delta": 0.9,
      "a0f": null,
      "a0f_delta": null
    },
    {
      "task": "shock-analysis-supply",
      "category": "finance-economics",
      "subcategory": "macroeconomic-analysis",
      "source": "c2-v2",
      "status": "DONE",
      "seed": 0.0,
      "cand_0001": 0.2,
      "cand_0002": 0.2,
      "cand_0003": 0.2,
      "cand_0004": 0.3,
      "best": 0.3,
      "best_tag": "cand_0004",
      "iters_spent": 4,
      "final_test": 0.2,
      "delta": 0.3,
      "a0f": null,
      "a0f_delta": null
    },
    {
      "task": "simpo-code-reproduction",
      "category": "software-engineering",
      "subcategory": "paper-to-code-reproduction",
      "source": "c2-v2",
      "status": "KILLED_val_1.0",
      "seed": 0.8,
      "cand_0001": 0.5,
      "cand_0002": 1.0,
      "cand_0003": null,
      "cand_0004": null,
      "best": 1.0,
      "best_tag": "cand_0002",
      "iters_spent": 2,
      "final_test": null,
      "delta": 0.19999999999999996,
      "a0f": null,
      "a0f_delta": null
    },
    {
      "task": "software-dependency-audit",
      "category": "cybersecurity",
      "subcategory": "vulnerability-analysis",
      "source": "c3-v1",
      "status": "DONE",
      "seed": 0.3,
      "cand_0001": 1.0,
      "cand_0002": 1.0,
      "cand_0003": 1.0,
      "cand_0004": 1.0,
      "best": 1.0,
      "best_tag": "cand_0001",
      "iters_spent": 4,
      "final_test": 1.0,
      "delta": 0.7,
      "a0f": null,
      "a0f_delta": null
    },
    {
      "task": "spring-boot-jakarta-migration",
      "category": "software-engineering",
      "subcategory": "migration",
      "source": "c1-v1",
      "status": "NO_SIGNAL",
      "seed": 0.0,
      "cand_0001": 0.0,
      "cand_0002": 0.0,
      "cand_0003": 0.0,
      "cand_0004": 0.0,
      "best": 0.0,
      "best_tag": "seed",
      "iters_spent": null,
      "final_test": null,
      "delta": 0.0,
      "a0f": null,
      "a0f_delta": null
    },
    {
      "task": "suricata-custom-exfil",
      "category": "cybersecurity",
      "subcategory": "intrusion-detection",
      "source": "mac-v2",
      "status": "DONE",
      "seed": 0.6666666666666666,
      "cand_0001": 0.0,
      "cand_0002": 0.0,
      "cand_0003": 0.0,
      "cand_0004": 0.0,
      "best": 0.6666666666666666,
      "best_tag": "seed",
      "iters_spent": 4,
      "final_test": 0.6666666666666666,
      "delta": 0.0,
      "a0f": null,
      "a0f_delta": null
    },
    {
      "task": "syzkaller-ppdev-syzlang",
      "category": "cybersecurity",
      "subcategory": "fuzzing",
      "source": "c1-v1",
      "status": "DONE",
      "seed": 0.7,
      "cand_0001": 0.9,
      "cand_0002": 1.0,
      "cand_0003": null,
      "cand_0004": null,
      "best": 1.0,
      "best_tag": "cand_0002",
      "iters_spent": null,
      "final_test": null,
      "delta": 0.30000000000000004,
      "a0f": null,
      "a0f_delta": null
    },
    {
      "task": "threejs-structure-parser",
      "category": "media-content-production",
      "subcategory": "3d-content",
      "source": "c1-v1",
      "status": "KILLED_ceiling",
      "seed": 0.5,
      "cand_0001": 1.0,
      "cand_0002": 1.0,
      "cand_0003": 1.0,
      "cand_0004": null,
      "best": 1.0,
      "best_tag": "cand_0001",
      "iters_spent": null,
      "final_test": null,
      "delta": 0.5,
      "a0f": null,
      "a0f_delta": null
    },
    {
      "task": "threejs-to-obj",
      "category": "media-content-production",
      "subcategory": "3d-content",
      "source": "c1-v1",
      "status": "KILLED_saturated",
      "seed": 1.0,
      "cand_0001": null,
      "cand_0002": null,
      "cand_0003": null,
      "cand_0004": null,
      "best": 1.0,
      "best_tag": "seed",
      "iters_spent": null,
      "final_test": null,
      "delta": 0.0,
      "a0f": null,
      "a0f_delta": null
    },
    {
      "task": "tictoc-unnecessary-abort-detection",
      "category": "software-engineering",
      "subcategory": "concurrency-control",
      "source": "c3-v1",
      "status": "DONE",
      "seed": 0.885,
      "cand_0001": 1.0,
      "cand_0002": 1.0,
      "cand_0003": 1.0,
      "cand_0004": 1.0,
      "best": 1.0,
      "best_tag": "cand_0001",
      "iters_spent": 4,
      "final_test": 1.0,
      "delta": 0.11499999999999999,
      "a0f": null,
      "a0f_delta": null
    },
    {
      "task": "travel-planning",
      "category": "mathematics-or-formal-reasoning",
      "subcategory": "calendar-scheduling",
      "source": "c3-v1",
      "status": "DONE",
      "seed": 0.0,
      "cand_0001": 0.2,
      "cand_0002": 0.6,
      "cand_0003": 1.0,
      "cand_0004": 1.0,
      "best": 1.0,
      "best_tag": "cand_0003",
      "iters_spent": 4,
      "final_test": 0.8,
      "delta": 1.0,
      "a0f": null,
      "a0f_delta": null
    },
    {
      "task": "video-silence-remover",
      "category": "media-content-production",
      "subcategory": "video-processing",
      "source": "c3-v1",
      "status": "DONE",
      "seed": 0.1,
      "cand_0001": 0.9,
      "cand_0002": 1.0,
      "cand_0003": 1.0,
      "cand_0004": 1.0,
      "best": 1.0,
      "best_tag": "cand_0002",
      "iters_spent": 4,
      "final_test": 1.0,
      "delta": 0.9,
      "a0f": null,
      "a0f_delta": null
    },
    {
      "task": "weighted-gdp-calc",
      "category": "finance-economics",
      "subcategory": "macroeconomic-analysis",
      "source": "c1-v1",
      "status": "KILLED_ceiling",
      "seed": 0.8,
      "cand_0001": 1.0,
      "cand_0002": 1.0,
      "cand_0003": null,
      "cand_0004": null,
      "best": 1.0,
      "best_tag": "cand_0001",
      "iters_spent": null,
      "final_test": null,
      "delta": 0.19999999999999996,
      "a0f": null,
      "a0f_delta": null
    },
    {
      "task": "xlsx-recover-data",
      "category": "finance-economics",
      "subcategory": "financial-modeling",
      "source": "c1-v1",
      "status": "KILLED_ceiling",
      "seed": 0.0,
      "cand_0001": 0.7,
      "cand_0002": 1.0,
      "cand_0003": null,
      "cand_0004": null,
      "best": 1.0,
      "best_tag": "cand_0002",
      "iters_spent": null,
      "final_test": null,
      "delta": 1.0,
      "a0f": null,
      "a0f_delta": null
    }
  ]
}