{
  "census_existing_logs": {
    "detector": "fingerprint, published thresholds (>=4 repeats, >=60 chars, period<=1024)",
    "note": "Census over existing GREEDY base-model gym completions (think@1024 atoms; multi-turn episode turns). This is the empirical basis for descoping the v1 loop-repair arm at deployed budgets; loops dominate only at 16k+ per the verified-macro ladder.",
    "sources": {
      "eval_gym_base_atoms_greedy_think1024": {
        "total": 1200,
        "flagged": 1,
        "rate": 0.0008333333333333334,
        "mean_think_chars": 2864.8141666666666,
        "examples": [
          {
            "id": "foundry_ledger-L2-s90001-0014",
            "period": 21,
            "repeats": 89,
            "snippet": " Guild\"\n    - \"Torbek"
          }
        ]
      },
      "eval_gym_base_episode_turns_greedy": {
        "total": 786,
        "flagged": 0,
        "rate": 0.0
      },
      "harvest_round1_atoms_temp0.8": {
        "total": 160,
        "flagged": 1,
        "rate": 0.00625,
        "mean_think_chars": 4709.08125,
        "examples": [
          {
            "id": "burrowmaze-L2-s11001-0013",
            "period": 35,
            "repeats": 5,
            "snippet": "            |\n                     "
          }
        ]
      }
    }
  },
  "harvest_extension_slices": {
    "prompts": 800,
    "slices": 7,
    "groups_with_node": 196,
    "census_group_rate": 0.245,
    "mixed_outcome_rate": 0.4925,
    "mined_rows_estimate": 237,
    "chosen_per_row_hist": {
      "1": 224,
      "2": 12,
      "3": 1
    },
    "p0_gate": [
      0.15,
      0.3
    ],
    "p0_pass": true,
    "wall_s": 7958.538434505463,
    "sampling": {
      "temperature": 0.6,
      "top_p": 0.95,
      "top_k": 20,
      "n": 16,
      "think_budget": 1024
    },
    "band_cells_used": [
      {
        "source": "code",
        "family": "listxform",
        "level": 2,
        "n": 12,
        "success": 0.16666666666666666,
        "in_band": true
      },
      {
        "source": "gym",
        "family": "caravan",
        "level": 3,
        "n": 12,
        "success": 0.16666666666666666,
        "in_band": true
      },
      {
        "source": "gym",
        "family": "ferrier",
        "level": 3,
        "n": 12,
        "success": 0.5833333333333334,
        "in_band": true
      },
      {
        "source": "gym",
        "family": "foundry_ledger",
        "level": 2,
        "n": 12,
        "success": 0.3333333333333333,
        "in_band": true
      },
      {
        "source": "gym",
        "family": "gatepost",
        "level": 2,
        "n": 12,
        "success": 0.75,
        "in_band": true
      },
      {
        "source": "gym",
        "family": "gatepost",
        "level": 3,
        "n": 12,
        "success": 0.8333333333333334,
        "in_band": true
      },
      {
        "source": "gym",
        "family": "glyphgate",
        "level": 1,
        "n": 12,
        "success": 0.25,
        "in_band": true
      },
      {
        "source": "gym",
        "family": "kilnrite",
        "level": 2,
        "n": 12,
        "success": 0.75,
        "in_band": true
      },
      {
        "source": "gym",
        "family": "kilnrite",
        "level": 3,
        "n": 12,
        "success": 0.5,
        "in_band": true
      },
      {
        "source": "gym",
        "family": "loomfix",
        "level": 1,
        "n": 12,
        "success": 0.5,
        "in_band": true
      },
      {
        "source": "gym",
        "family": "loomfix",
        "level": 2,
        "n": 12,
        "success": 0.5,
        "in_band": true
      },
      {
        "source": "gym",
        "family": "loomfix",
        "level": 3,
        "n": 12,
        "success": 0.3333333333333333,
        "in_band": true
      },
      {
        "source": "gym",
        "family": "runeward",
        "level": 2,
        "n": 12,
        "success": 0.6666666666666666,
        "in_band": true
      },
      {
        "source": "gym",
        "family": "runeward",
        "level": 3,
        "n": 12,
        "success": 0.16666666666666666,
        "in_band": true
      }
    ]
  },
  "build_rows_pivot": {
    "arm": "pivot",
    "counts": {
      "groups": 2800,
      "groups_with_node": 717,
      "rows": 879,
      "chosen_case_variant_dropped": 3,
      "no_chosen_after_filters": 3,
      "pool_rows": 879,
      "rows_dropped_empty_chosen": 0,
      "train_rows": 615
    },
    "train_rows": 615,
    "pool_rows": 879,
    "min_train_rows_gate": 600,
    "gate_pass": true,
    "chosen_per_row_hist": {
      "1": 582,
      "2": 28,
      "3": 5
    },
    "top_rejected_surfaces": [
      [
        "specific",
        19
      ],
      [
        "a",
        18
      ],
      [
        "the",
        16
      ],
      [
        "and",
        16
      ],
      [
        "`",
        15
      ],
      [
        ".",
        14
      ],
      [
        "to",
        13
      ],
      [
        "\"",
        13
      ],
      [
        "examples",
        13
      ],
      [
        "",
        12
      ],
      [
        "state",
        11
      ],
      [
        "holds",
        11
      ],
      [
        "based",
        10
      ],
      [
        "whole",
        10
      ],
      [
        "has",
        10
      ]
    ],
    "median_context_tokens": 297
  },
  "build_rows_shuffled": {
    "arm": "shuffled",
    "counts": {
      "groups": 2800,
      "groups_with_node": 749,
      "rows": 945,
      "chosen_case_variant_dropped": 2,
      "no_chosen_after_filters": 2,
      "pool_rows": 945,
      "rows_dropped_empty_chosen": 0,
      "train_rows": 661
    },
    "train_rows": 661,
    "pool_rows": 945,
    "min_train_rows_gate": 600,
    "gate_pass": true,
    "chosen_per_row_hist": {
      "1": 633,
      "2": 22,
      "3": 6
    },
    "top_rejected_surfaces": [
      [
        "the",
        20
      ],
      [
        "a",
        18
      ],
      [
        "`",
        17
      ],
      [
        "and",
        17
      ],
      [
        ".",
        14
      ],
      [
        "",
        14
      ],
      [
        "has",
        14
      ],
      [
        "examples",
        14
      ],
      [
        "to",
        13
      ],
      [
        "holds",
        13
      ],
      [
        "\"",
        13
      ],
      [
        "specific",
        13
      ],
      [
        "based",
        11
      ],
      [
        "input",
        11
      ],
      [
        "determine",
        10
      ]
    ],
    "median_context_tokens": 296
  },
  "whitebox": {
    "base_main": {
      "arm": "base",
      "stage": "main",
      "model": "base",
      "think@1024": {
        "n": 490,
        "success": 0.49795918367346936,
        "mean_score": 0.49795918367346936,
        "loop_rate": 0.0020408163265306124,
        "unresolved_rate": 0.9836734693877551,
        "answer_limit_rate": 0.44285714285714284,
        "natural_close_rate": 0.014285714285714285,
        "mean_think_tokens": 1020.8938775510204
      },
      "think@1024_per_source": {
        "gym": {
          "n": 455,
          "success": 0.5142857142857142,
          "mean_score": 0.5142857142857142,
          "loop_rate": 0.002197802197802198,
          "unresolved_rate": 0.9824175824175824,
          "answer_limit_rate": 0.43956043956043955,
          "natural_close_rate": 0.015384615384615385,
          "mean_think_tokens": 1020.6549450549451
        },
        "code": {
          "n": 35,
          "success": 0.2857142857142857,
          "mean_score": 0.2857142857142857,
          "loop_rate": 0.0,
          "unresolved_rate": 1.0,
          "answer_limit_rate": 0.4857142857142857,
          "natural_close_rate": 0.0,
          "mean_think_tokens": 1024.0
        }
      },
      "think@2048": {
        "n": 490,
        "success": 0.5408163265306123,
        "mean_score": 0.5408163265306123,
        "loop_rate": 0.004081632653061225,
        "unresolved_rate": 0.8285714285714286,
        "answer_limit_rate": 0.4204081632653061,
        "natural_close_rate": 0.1673469387755102,
        "mean_think_tokens": 1954.7571428571428
      },
      "think@2048_per_source": {
        "gym": {
          "n": 455,
          "success": 0.5560439560439561,
          "mean_score": 0.5560439560439561,
          "loop_rate": 0.004395604395604396,
          "unresolved_rate": 0.8197802197802198,
          "answer_limit_rate": 0.4065934065934066,
          "natural_close_rate": 0.17582417582417584,
          "mean_think_tokens": 1950.2087912087911
        },
        "code": {
          "n": 35,
          "success": 0.34285714285714286,
          "mean_score": 0.34285714285714286,
          "loop_rate": 0.0,
          "unresolved_rate": 0.9428571428571428,
          "answer_limit_rate": 0.6,
          "natural_close_rate": 0.05714285714285714,
          "mean_think_tokens": 2013.8857142857144
        }
      },
      "wall_s": 848.3690950870514
    },
    "base_formats": {
      "arm": "base",
      "stage": "formats",
      "model": "base",
      "variant1": {
        "n": 140,
        "success": 0.014285714285714285,
        "mean_score": 0.014285714285714285,
        "loop_rate": 0.0,
        "unresolved_rate": 1.0,
        "answer_limit_rate": 0.5071428571428571,
        "natural_close_rate": 0.0,
        "mean_think_tokens": 1024.0
      },
      "variant2": {
        "n": 140,
        "success": 0.45714285714285713,
        "mean_score": 0.45714285714285713,
        "loop_rate": 0.0,
        "unresolved_rate": 0.9857142857142858,
        "answer_limit_rate": 0.5214285714285715,
        "natural_close_rate": 0.014285714285714285,
        "mean_think_tokens": 1020.7
      },
      "wall_s": 201.48683834075928
    },
    "base_coverage": {
      "arm": "base",
      "stage": "coverage",
      "model": "base",
      "coverage_at_8": 0.7295918367346939,
      "label": "NON-DEPLOYABLE oracle ceiling (best-of-8, no selector)",
      "wall_s": 2006.4867465496063
    },
    "base_collapse": {
      "arm": "base",
      "stage": "collapse",
      "model": "base",
      "greedy": {
        "n": 120,
        "success": 0.058333333333333334,
        "mean_score": 0.058333333333333334,
        "loop_rate": 0.0,
        "unresolved_rate": 1.0,
        "answer_limit_rate": 0.7,
        "natural_close_rate": 0.0,
        "mean_think_tokens": 1024.0
      },
      "pass@8": 0.10833333333333334,
      "wall_s": 867.8146548271179
    },
    "base_nothink": {
      "arm": "base",
      "stage": "nothink",
      "model": "base",
      "nothink": {
        "n": 120,
        "success": 0.36666666666666664,
        "mean_score": 0.36666666666666664,
        "loop_rate": 0.0,
        "unresolved_rate": 0.0,
        "answer_limit_rate": 0.625,
        "natural_close_rate": 1.0,
        "mean_think_tokens": 0.0
      },
      "wall_s": 11.793559074401855
    },
    "pivot_main": {
      "arm": "pivot",
      "stage": "main",
      "model": "../../large_artifacts/qwen35_4b_think_ftpo_round1/merged/pivot",
      "think@1024": {
        "n": 490,
        "success": 0.45918367346938777,
        "mean_score": 0.45918367346938777,
        "loop_rate": 0.0020408163265306124,
        "unresolved_rate": 0.9959183673469387,
        "answer_limit_rate": 0.47551020408163264,
        "natural_close_rate": 0.0020408163265306124,
        "mean_think_tokens": 1023.5979591836734
      },
      "think@1024_per_source": {
        "gym": {
          "n": 455,
          "success": 0.4747252747252747,
          "mean_score": 0.4747252747252747,
          "loop_rate": 0.002197802197802198,
          "unresolved_rate": 0.9956043956043956,
          "answer_limit_rate": 0.4703296703296703,
          "natural_close_rate": 0.002197802197802198,
          "mean_think_tokens": 1023.567032967033
        },
        "code": {
          "n": 35,
          "success": 0.2571428571428571,
          "mean_score": 0.2571428571428571,
          "loop_rate": 0.0,
          "unresolved_rate": 1.0,
          "answer_limit_rate": 0.5428571428571428,
          "natural_close_rate": 0.0,
          "mean_think_tokens": 1024.0
        }
      },
      "think@2048": {
        "n": 490,
        "success": 0.46530612244897956,
        "mean_score": 0.46530612244897956,
        "loop_rate": 0.0020408163265306124,
        "unresolved_rate": 0.9224489795918367,
        "answer_limit_rate": 0.4897959183673469,
        "natural_close_rate": 0.07551020408163266,
        "mean_think_tokens": 2013.5816326530612
      },
      "think@2048_per_source": {
        "gym": {
          "n": 455,
          "success": 0.47912087912087914,
          "mean_score": 0.47912087912087914,
          "loop_rate": 0.002197802197802198,
          "unresolved_rate": 0.9164835164835164,
          "answer_limit_rate": 0.47912087912087914,
          "natural_close_rate": 0.08131868131868132,
          "mean_think_tokens": 2010.934065934066
        },
        "code": {
          "n": 35,
          "success": 0.2857142857142857,
          "mean_score": 0.2857142857142857,
          "loop_rate": 0.0,
          "unresolved_rate": 1.0,
          "answer_limit_rate": 0.6285714285714286,
          "natural_close_rate": 0.0,
          "mean_think_tokens": 2048.0
        }
      },
      "wall_s": 858.7652988433838
    },
    "pivot_formats": {
      "arm": "pivot",
      "stage": "formats",
      "model": "../../large_artifacts/qwen35_4b_think_ftpo_round1/merged/pivot",
      "variant1": {
        "n": 140,
        "success": 0.007142857142857143,
        "mean_score": 0.007142857142857143,
        "loop_rate": 0.0,
        "unresolved_rate": 1.0,
        "answer_limit_rate": 0.5785714285714286,
        "natural_close_rate": 0.0,
        "mean_think_tokens": 1024.0
      },
      "variant2": {
        "n": 140,
        "success": 0.39285714285714285,
        "mean_score": 0.39285714285714285,
        "loop_rate": 0.0,
        "unresolved_rate": 1.0,
        "answer_limit_rate": 0.5642857142857143,
        "natural_close_rate": 0.0,
        "mean_think_tokens": 1024.0
      },
      "wall_s": 201.5766885280609
    },
    "pivot_collapse": {
      "arm": "pivot",
      "stage": "collapse",
      "model": "../../large_artifacts/qwen35_4b_think_ftpo_round1/merged/pivot",
      "greedy": {
        "n": 120,
        "success": 0.06666666666666667,
        "mean_score": 0.06666666666666667,
        "loop_rate": 0.008333333333333333,
        "unresolved_rate": 0.9916666666666667,
        "answer_limit_rate": 0.7083333333333334,
        "natural_close_rate": 0.0,
        "mean_think_tokens": 1024.0
      },
      "pass@8": 0.1,
      "wall_s": 838.4791285991669
    },
    "pivot_nothink": {
      "arm": "pivot",
      "stage": "nothink",
      "model": "../../large_artifacts/qwen35_4b_think_ftpo_round1/merged/pivot",
      "nothink": {
        "n": 120,
        "success": 0.4083333333333333,
        "mean_score": 0.4083333333333333,
        "loop_rate": 0.0,
        "unresolved_rate": 0.0,
        "answer_limit_rate": 0.5916666666666667,
        "natural_close_rate": 1.0,
        "mean_think_tokens": 0.0
      },
      "wall_s": 11.63565969467163
    },
    "shuffled_main": {
      "arm": "shuffled",
      "stage": "main",
      "model": "../../large_artifacts/qwen35_4b_think_ftpo_round1/merged/shuffled",
      "think@1024": {
        "n": 490,
        "success": 0.47551020408163264,
        "mean_score": 0.47551020408163264,
        "loop_rate": 0.0020408163265306124,
        "unresolved_rate": 0.9959183673469387,
        "answer_limit_rate": 0.47959183673469385,
        "natural_close_rate": 0.0020408163265306124,
        "mean_think_tokens": 1023.9795918367347
      },
      "think@1024_per_source": {
        "gym": {
          "n": 455,
          "success": 0.49230769230769234,
          "mean_score": 0.49230769230769234,
          "loop_rate": 0.002197802197802198,
          "unresolved_rate": 0.9956043956043956,
          "answer_limit_rate": 0.4725274725274725,
          "natural_close_rate": 0.002197802197802198,
          "mean_think_tokens": 1023.978021978022
        },
        "code": {
          "n": 35,
          "success": 0.2571428571428571,
          "mean_score": 0.2571428571428571,
          "loop_rate": 0.0,
          "unresolved_rate": 1.0,
          "answer_limit_rate": 0.5714285714285714,
          "natural_close_rate": 0.0,
          "mean_think_tokens": 1024.0
        }
      },
      "think@2048": {
        "n": 490,
        "success": 0.4857142857142857,
        "mean_score": 0.4857142857142857,
        "loop_rate": 0.00816326530612245,
        "unresolved_rate": 0.9163265306122449,
        "answer_limit_rate": 0.48367346938775513,
        "natural_close_rate": 0.07551020408163266,
        "mean_think_tokens": 2013.0795918367346
      },
      "think@2048_per_source": {
        "gym": {
          "n": 455,
          "success": 0.4967032967032967,
          "mean_score": 0.4967032967032967,
          "loop_rate": 0.006593406593406593,
          "unresolved_rate": 0.9120879120879121,
          "answer_limit_rate": 0.4747252747252747,
          "natural_close_rate": 0.08131868131868132,
          "mean_think_tokens": 2010.3934065934066
        },
        "code": {
          "n": 35,
          "success": 0.34285714285714286,
          "mean_score": 0.34285714285714286,
          "loop_rate": 0.02857142857142857,
          "unresolved_rate": 0.9714285714285714,
          "answer_limit_rate": 0.6,
          "natural_close_rate": 0.0,
          "mean_think_tokens": 2048.0
        }
      },
      "wall_s": 857.4539268016815
    },
    "shuffled_collapse": {
      "arm": "shuffled",
      "stage": "collapse",
      "model": "../../large_artifacts/qwen35_4b_think_ftpo_round1/merged/shuffled",
      "greedy": {
        "n": 120,
        "success": 0.058333333333333334,
        "mean_score": 0.058333333333333334,
        "loop_rate": 0.016666666666666666,
        "unresolved_rate": 0.9833333333333333,
        "answer_limit_rate": 0.8333333333333334,
        "natural_close_rate": 0.0,
        "mean_think_tokens": 1024.0
      },
      "pass@8": 0.10833333333333334,
      "wall_s": 849.3442122936249
    }
  },
  "eval_gym": {
    "base": {
      "arm": "base",
      "model": "base",
      "per_family": {
        "brinework": 0.5,
        "burrowmaze": 0.7142857142857143,
        "caravan": 0.5,
        "ferrier": 0.7857142857142857,
        "foundry_ledger": 0.4166666666666667,
        "gatepost": 0.8571428571428571,
        "glyphgate": 0.14285714285714285,
        "kilnrite": 0.7000000000000001,
        "loomfix": 0.21428571428571427,
        "runeward": 0.75,
        "spindle": 0.625,
        "stallwright": 0.0
      },
      "aggregate_all": 0.5171626984126984,
      "aggregate_trained": 0.508095238095238,
      "aggregate_heldout": 0.5625,
      "forced_close_rate_atoms": 0.9166666666666666,
      "n_atoms": 144,
      "n_episodes": 14,
      "wall_s": 203.39925861358643
    },
    "pivot": {
      "arm": "pivot",
      "model": "../../large_artifacts/qwen35_4b_think_ftpo_round1/merged/pivot",
      "per_family": {
        "brinework": 0.4166666666666667,
        "burrowmaze": 0.7142857142857143,
        "caravan": 0.5833333333333334,
        "ferrier": 0.9285714285714286,
        "foundry_ledger": 0.4166666666666667,
        "gatepost": 0.7857142857142857,
        "glyphgate": 0.14285714285714285,
        "kilnrite": 0.6714285714285715,
        "loomfix": 0.07142857142857142,
        "runeward": 0.5833333333333334,
        "spindle": 0.5,
        "stallwright": 0.0
      },
      "aggregate_all": 0.4845238095238095,
      "aggregate_trained": 0.4897619047619048,
      "aggregate_heldout": 0.45833333333333337,
      "forced_close_rate_atoms": 0.9652777777777778,
      "n_atoms": 144,
      "n_episodes": 14,
      "wall_s": 212.71113443374634
    },
    "shuffled": {
      "arm": "shuffled",
      "model": "../../large_artifacts/qwen35_4b_think_ftpo_round1/merged/shuffled",
      "per_family": {
        "brinework": 0.4166666666666667,
        "burrowmaze": 0.7142857142857143,
        "caravan": 0.4166666666666667,
        "ferrier": 0.7857142857142857,
        "foundry_ledger": 0.4166666666666667,
        "gatepost": 0.8571428571428571,
        "glyphgate": 0.2857142857142857,
        "kilnrite": 0.6428571428571429,
        "loomfix": 0.2857142857142857,
        "runeward": 0.6666666666666666,
        "spindle": 0.6785714285714286,
        "stallwright": 0.0
      },
      "aggregate_all": 0.513888888888889,
      "aggregate_trained": 0.5071428571428571,
      "aggregate_heldout": 0.5476190476190477,
      "forced_close_rate_atoms": 0.9722222222222222,
      "n_atoms": 144,
      "n_episodes": 14,
      "wall_s": 216.24292087554932
    }
  },
  "gates": {
    "pivot_c49": "PASS (0/8 identical)",
    "shuffled_c49": "see eval_chain.log"
  },
  "headline_table": [
    {
      "arm": "base",
      "budget": "think@1024",
      "success": 0.498,
      "natural_close": 0.0143,
      "loop_rate": 0.002,
      "unresolved": 0.9837,
      "answer_limit": 0.4429
    },
    {
      "arm": "base",
      "budget": "think@2048",
      "success": 0.5408,
      "natural_close": 0.1673,
      "loop_rate": 0.0041,
      "unresolved": 0.8286,
      "answer_limit": 0.4204
    },
    {
      "arm": "pivot",
      "budget": "think@1024",
      "success": 0.4592,
      "natural_close": 0.002,
      "loop_rate": 0.002,
      "unresolved": 0.9959,
      "answer_limit": 0.4755,
      "success_delta_vs_base": -0.0388
    },
    {
      "arm": "pivot",
      "budget": "think@2048",
      "success": 0.4653,
      "natural_close": 0.0755,
      "loop_rate": 0.002,
      "unresolved": 0.9224,
      "answer_limit": 0.4898,
      "success_delta_vs_base": -0.0755
    },
    {
      "arm": "shuffled",
      "budget": "think@1024",
      "success": 0.4755,
      "natural_close": 0.002,
      "loop_rate": 0.002,
      "unresolved": 0.9959,
      "answer_limit": 0.4796,
      "success_delta_vs_base": -0.0224
    },
    {
      "arm": "shuffled",
      "budget": "think@2048",
      "success": 0.4857,
      "natural_close": 0.0755,
      "loop_rate": 0.0082,
      "unresolved": 0.9163,
      "answer_limit": 0.4837,
      "success_delta_vs_base": -0.0551
    }
  ],
  "verdicts": {
    "P0_census": {
      "eligible_rate": null,
      "note": "combined across main+extension; see report"
    },
    "P1_mechanism": {
      "bar": "+0.05 absolute greedy success on held-out band tasks",
      "measured_1024": -0.0388,
      "measured_2048": -0.0755,
      "verdict": "FAIL"
    },
    "control_read": {
      "shuffled_delta_1024": -0.0224,
      "shuffled_delta_2048": -0.0551,
      "note": "shuffled ~= pivot degradation -> damage is generic to the training regime, not the outcome-conditioned signal"
    },
    "collapse_guard": {
      "greedy_rel_change": 0.1429,
      "pass8_rel_change": -0.0769,
      "flag_damaging": false
    },
    "nothink_guard": {
      "base": 0.3667,
      "pivot": 0.4083,
      "flag": false
    },
    "P3_gym": {
      "base_aggregate": 0.5172,
      "pivot_aggregate": 0.4845,
      "shuffled_aggregate": 0.5139,
      "guard_fail": true
    },
    "P4_menagerie": {
      "verdict": "NOT RUN \u2014 preregistered rule: mechanism prediction (P1) failed, so the round is a training-recipe failure with no capability read; blackbox spend cancelled"
    }
  }
}