{
  "metadata": {
    "generated_at": "2026-10-24T17:57:28.283413",
    "schema_version": "1.0.0",
    "spec_version": "v4",
    "git_commit": "0000000000000000000000000000000000000000"
  },
  "forecast": {
    "status": "completed",
    "h1": {
      "auc": 0.571,
      "ci_low": 0.511,
      "ci_high": 0.632,
      "threshold": 0.6,
      "supported": false,
      "confidence": "medium"
    },
    "h2": {
      "auc_obs": 0.571,
      "auc_obs_bridge": 0.567,
      "delta": -0.004,
      "delong_p": 0.54,
      "perm_p": 0.61,
      "supported": false,
      "confidence": "medium"
    },
    "gradient_boosting": {
      "auc": 0.49,
      "brier": 0.034,
      "base_rate_brier": 0.0341,
      "confidence": "medium"
    },
    "forecast_csv": {
      "path": "forecaster/out/forecast.csv",
      "sha256": "2c9a9d86ec22b205717723ce04ad9d8d6e013372f0f79b897e6813b3f1245981",
      "problem_count": 593,
      "expected_3mo": 7.8,
      "model": "hgb_all"
    },
    "repo_url": "https://github.com/Questboy23/ai-math-forecast",
    "forecast_sha256": "2c9a9d86ec22b205717723ce04ad9d8d6e013372f0f79b897e6813b3f1245981",
    "backtest_summary": {
      "test_origins": 7,
      "rows": 4458,
      "events": 152,
      "problems": 68
    }
  },
  "formal_conjectures": {
    "status": "available",
    "erdos_problems": {
      "count": 909,
      "path": "~/euled-experiments/forecaster/data/formal-conjectures/FormalConjectures/ErdosProblems/",
      "theorems": {
        "total": 2918,
        "proved": 2114,
        "open": 804,
        "pending": 0
      }
    },
    "categories": [
      "Arxiv",
      "Books",
      "ErdosProblems",
      "GreensOpenProblems",
      "HilbertProblems",
      "Kourovka",
      "LittProblems",
      "Mathoverflow",
      "Millennium",
      "OEIS",
      "OpenQuantumProblems",
      "OptimizationConstants",
      "Other",
      "Paper",
      "Subsets",
      "Wikipedia",
      "WrittenOnTheWallII"
    ],
    "lean_toolchain": "leanprover/lean4:v4.33.1",
    "verification": {
      "axle_verify_proof": false,
      "axiom_audit": false,
      "lake_build": false,
      "allowed_axioms": [
        "propext",
        "Quot.sound",
        "Classical.choice"
      ],
      "rejected_axioms": []
    }
  },
  "fme": {
    "status": "pending",
    "conjecture_count": 68,
    "repo_url": "https://github.com/epoch-research/LeanOpenProblems",
    "lean_toolchain": "pending",
    "verification": {
      "axle_verify_proof": false,
      "axiom_audit": false,
      "lake_build": false
    }
  },
  "alphaproof_nexus": {
    "status": "pending",
    "repo_url": "https://github.com/google-deepmind/alphaproof-nexus-results",
    "proofs": [],
    "verification": {
      "axle_verify_proof": false,
      "axiom_audit": false,
      "lake_build": false
    }
  },
  "putnambench": {
    "status": "verified",
    "repo_url": "https://github.com/trishullab/PutnamBench",
    "solved_count": 6,
    "verification": {
      "axle_verify_proof": false,
      "axiom_audit": true,
      "lake_build": true
    },
    "solved": [
      {
        "id": "putnam_1969_b6",
        "source": "PutnamBench",
        "file": "bench/proofs/putnam_1969_b6.lean",
        "sha256_16": "6f01ac3ca6ea6569",
        "verified": {
          "axle_verify_proof": true,
          "lake_env_lean": true,
          "rule2_harness": false
        },
        "axioms": [
          "propext",
          "Classical.choice",
          "Quot.sound"
        ],
        "proved_at": "2026-10-24 16:45-17:25 EDT",
        "method": "Max (Sonnet 5.5) run, informal sketch then Lean proof, AXLE-checked"
      },
      {
        "id": "putnam_1972_a1",
        "source": "PutnamBench",
        "file": "bench/proofs/putnam_1972_a1.lean",
        "sha256_16": "8200fad87a9d035e",
        "verified": {
          "axle_verify_proof": true,
          "lake_env_lean": true,
          "rule2_harness": true
        },
        "axioms": [
          "propext",
          "Classical.choice",
          "Quot.sound"
        ],
        "proved_at": "2026-10-24 16:45-17:25 EDT",
        "method": "Max (Sonnet 5.5) run, informal sketch then Lean proof, AXLE-checked"
      },
      {
        "id": "putnam_2000_a4",
        "source": "PutnamBench",
        "file": "bench/proofs/putnam_2000_a4.lean",
        "sha256_16": "d1ce3da2fa57e1d3",
        "verified": {
          "axle_verify_proof": true,
          "lake_env_lean": true,
          "rule2_harness": false
        },
        "axioms": [
          "propext",
          "Classical.choice",
          "Quot.sound"
        ],
        "proved_at": "2026-10-24 16:45-17:25 EDT",
        "method": "Max (Sonnet 5.5) run, informal sketch then Lean proof, AXLE-checked"
      },
      {
        "id": "putnam_2010_a4",
        "source": "PutnamBench",
        "file": "bench/proofs/putnam_2010_a4.lean",
        "sha256_16": "b02142102c4401e0",
        "verified": {
          "axle_verify_proof": true,
          "lake_env_lean": true,
          "rule2_harness": false
        },
        "axioms": [
          "propext",
          "Classical.choice",
          "Quot.sound"
        ],
        "proved_at": "2026-10-24 16:45-17:25 EDT",
        "method": "Max (Sonnet 5.5) run, informal sketch then Lean proof, AXLE-checked"
      },
      {
        "id": "putnam_2015_b1",
        "source": "PutnamBench",
        "file": "bench/proofs/putnam_2015_b1.lean",
        "sha256_16": "3202817c51b4c668",
        "verified": {
          "axle_verify_proof": true,
          "lake_env_lean": true,
          "rule2_harness": true
        },
        "axioms": [
          "propext",
          "Classical.choice",
          "Quot.sound"
        ],
        "proved_at": "2026-10-24 16:45-17:25 EDT",
        "method": "Max (Sonnet 5.5) run, informal sketch then Lean proof, AXLE-checked"
      },
      {
        "id": "putnam_2017_b1",
        "source": "PutnamBench",
        "file": "bench/proofs/putnam_2017_b1.lean",
        "sha256_16": "00a6c0937dafdafb",
        "verified": {
          "axle_verify_proof": true,
          "lake_env_lean": true,
          "rule2_harness": true
        },
        "axioms": [
          "propext",
          "Classical.choice",
          "Quot.sound"
        ],
        "proved_at": "2026-10-24 16:45-17:25 EDT",
        "method": "Max (Sonnet 5.5) run, informal sketch then Lean proof, AXLE-checked"
      }
    ],
    "run": {
      "log": "logs/PROGRESS.md (entries 16:45-17:25 EDT), logs/improve_build.log",
      "seed": "n/a (model-driven); proof integrity via sha256_16 per file",
      "commit": "d43a22a",
      "rerun_command": "bash bench/check.sh <putnam_id> bench/proofs/<putnam_id>.lean",
      "regenerate_command": "bash run_improve3.sh",
      "budget": {
        "max_spent_usd": 3.88,
        "cap_usd": 29.6
      },
      "toolchain": "leanprover/lean4:v4.33.1",
      "checker": "AXLE verify_proof (bench/axle_verify.json) + lake env lean (all 6 exit 0)"
    }
  },
  "minif2f": {
    "status": "pending",
    "repo_url": "https://github.com/...",
    "solved_count": 0,
    "verification": {
      "axle_verify_proof": false,
      "axiom_audit": false,
      "lake_build": false
    }
  },
  "reinvention": {
    "status": "pending",
    "repo_url": "",
    "results": []
  },
  "sweep": {
    "status": "pending",
    "misformalization_candidates": []
  },
  "shortener": {
    "status": "pending",
    "alpha_proof_nexus_proofs": [],
    "new_proofs": [],
    "idempotence_test": {
      "cpu_passes_defect": false,
      "neural_passes_nonconvergence": false,
      "seeds": 3
    }
  },
  "external_anchors": {
    "anchors": []
  },
  "gpu_ledger": {
    "jobs": [
      {
        "job_id": "euled-g0",
        "phase": "kaggle",
        "gpu_hours": 0.09,
        "gpu_type": "2x T4",
        "tokens": 0,
        "status": "failed:no-internet",
        "timestamp": "2026-10-24T06:18:26Z",
        "tag": "kaggle"
      }
    ]
  },
  "verification": {
    "kimina_pool": false,
    "axle_statement_check": false,
    "axiom_audit": false,
    "html_json_check": false,
    "wrapper_check": false
  },
  "compute_matched": {
    "status": "pending",
    "arms": {
      "A": {
        "description": "whole-proof best-of-n sampling (baseline)",
        "gpu_hours": 0,
        "verified_solves": 0,
        "solves_per_gpu_hour": 0,
        "cost_per_solve": 0
      },
      "B": {
        "description": "Euled cascade + two-ended search + lemma bank, uniform effort",
        "gpu_hours": 0,
        "verified_solves": 0,
        "solves_per_gpu_hour": 0,
        "cost_per_solve": 0,
        "lemma_bank_reuse_rate": 0
      },
      "C": {
        "description": "Euled cascade + coverage-per-cost allocation",
        "gpu_hours": 0,
        "verified_solves": 0,
        "solves_per_gpu_hour": 0,
        "cost_per_solve": 0,
        "lemma_bank_reuse_rate": 0
      }
    },
    "power_analysis": {
      "required_solves": 80,
      "predicted_solves": 0,
      "threshold_reached": false,
      "icc": 0,
      "cluster_bootstrap": false
    }
  },
  "exponential_test": {
    "status": "pending",
    "rounds": 0,
    "control_arm": false,
    "aicc_result": {
      "constant_ratio_aicc": 0,
      "logistic_aicc": 0,
      "winner": "insufficient_rounds",
      "fitted_ratio": 0
    }
  },
  "background": {
    "status": "pending",
    "target": "",
    "scores": [],
    "pick": "",
    "progress_metrics": {
      "bridge_distance": 0,
      "verified_sublemmas": 0,
      "lemma_bank_size": 0,
      "lemma_bank_reuse": 0,
      "open_subgoals": 0,
      "cpu_hours": 0,
      "gpu_hours": 0,
      "fcc_calls": 0
    }
  }
}