{
  "model": "Lev-2.6B",
  "checkpoint": "ouro26-ls-s0 (S31, seed 0)",
  "experiment": "E06",
  "split": "development",
  "temperatures": [
    1.148698354997035,
    1.1755479062836087,
    1.148698354997035,
    1.148698354997035
  ],
  "min_agree": 0.98,
  "probability_definition": "softmax(archived fp32 logits / per-depth calibration temperature); probabilities are percentages",
  "stopper": "Reconstructed offline with the unchanged E06 two-fold, group-disjoint agreement-fitting procedure; not a new model run. Calibration plus the opposite development fold train each logistic regressor. The held-out case never enters its stopper fit. τ is fitted on the opposite fold for 98% aggregate agreement.",
  "selection": "Two selected successful examples for mechanism explanation, not a representative accuracy estimate. Easy/Hard describe these observed trajectories, not official dataset difficulty labels.",
  "display": "Original context and option wording retained. The MathQA problem field is displayed as prose, its question is repeated as a short heading, and option keys are uppercased in the UI. Editorial answer explanations are not model reasoning.",
  "features": [
    "top probability",
    "margin",
    "normalised entropy",
    "same answer as the previous depth",
    "KL(previous depth || this depth)",
    "log option count",
    "top probability minus the previous depth's",
    "depth"
  ],
  "fitted_stoppers": [
    {
      "held_out_fold": 0,
      "threshold": 0.9275,
      "coefficients": [
        [
          1.8177726888317134,
          1.2164108045251345,
          -1.6362144493475144,
          0.3142525266617298,
          0.0,
          -0.18248579602295434,
          0.0,
          0.0
        ],
        [
          -0.2568506277304674,
          4.387708854990757,
          -0.34262230108176084,
          0.3295457980462332,
          -0.6526291438700512,
          0.3480192703748077,
          -0.18653352253019215,
          0.010875785906513887
        ],
        [
          0.6949800557353737,
          5.1031299523408,
          0.2808302619392168,
          0.4071217362239693,
          -0.5214717058319877,
          0.7749263762051971,
          1.1634614087378663,
          -0.4322270899806101
        ]
      ],
      "intercepts": [
        0.3986473381559831,
        -0.11519241434714916,
        -0.3246069767044413
      ]
    },
    {
      "held_out_fold": 1,
      "threshold": 0.9225,
      "coefficients": [
        [
          1.5411509109795631,
          1.217183511395416,
          -1.6852543455022044,
          0.39521724421987753,
          0.0,
          -0.19192591152662086,
          0.0,
          0.0
        ],
        [
          -0.025709194547677606,
          3.4586320343197534,
          -0.9311096820769365,
          0.14015794862814734,
          -0.6163221975366481,
          0.21759986385374663,
          0.17685187495769653,
          0.23275961240769946
        ],
        [
          0.8464335990417597,
          5.192660717207364,
          0.11623106022776104,
          0.5445267275712011,
          -0.3260925341065738,
          0.4829585215926536,
          1.8200858789781014,
          -0.3069012588401383
        ]
      ],
      "intercepts": [
        0.476946865197514,
        0.4205803445142968,
        -0.23337231697949073
      ]
    }
  ],
  "cases": {
    "easy": {
      "label": "Easy",
      "kind": "Social IQa",
      "title": "Airport pickup",
      "heading": "What will happen to Robin?",
      "options": [
        "A",
        "B",
        "C"
      ],
      "option_text": [
        "got Robin seasick",
        "got into Cameron's automobile",
        "got lost looking for Cameron"
      ],
      "option_keys": [
        "a",
        "b",
        "c"
      ],
      "state": "Cameron picked Robin up from the airport.",
      "instructions": "What will happen to Robin?",
      "dataset_label": "b",
      "probabilities": [
        [
          0.7608458778888406,
          97.12394043981078,
          2.115213682300366
        ],
        [
          1.597527221920221,
          95.61023977102721,
          2.7922330070525603
        ],
        [
          0.4819611477977534,
          96.63737019088698,
          2.8806686613152577
        ],
        [
          0.2594014421526442,
          98.75309600816523,
          0.9875025496821381
        ]
      ],
      "logits": [
        [
          -2.491669178009033,
          3.078727960586548,
          -1.3171477317810059
        ],
        [
          -2.3323044776916504,
          2.4778294563293457,
          -1.6758965253829956
        ],
        [
          -3.1381423473358154,
          2.9509437084198,
          -1.0843682289123535
        ],
        [
          -2.8968396186828613,
          3.928727388381958,
          -1.3612570762634277
        ]
      ],
      "agreement": [
        0.9564382524581166,
        0.9826414758473851,
        0.9973943657691757
      ],
      "threshold": 0.9225,
      "exitDepth": 1,
      "explanation": "The model selects B, matching the dataset label: got into Cameron's automobile. The stopper clears the threshold at loop 1.",
      "source": {
        "dataset": "Social IQa",
        "record_id": "loopjev-eval-v1/development/id/loopjev_social_iqa/validation/1407",
        "origin_id": "loopjev_social_iqa/validation/1407",
        "partition": "LevBench development",
        "upstream_split": "validation",
        "upstream_row": 1407,
        "url": "https://huggingface.co/datasets/allenai/social_i_qa",
        "revision": "537a2ec8ec565adc0b70b70752893e59e024df26",
        "license": "CC-BY-4.0",
        "attribution": "Sap et al. (2019), Social IQa; Maarten Sap and collaborators.",
        "state_sha256": "027c906c09b585af5c90e0112d1db1221295b69b36ec112776379c5722a04b45",
        "record_group": "loopjev-eval-v1/loopjev_social_iqa/context/027c906c09b585af",
        "held_out_fold": 1
      }
    },
    "hard": {
      "label": "Hard",
      "kind": "MathQA",
      "title": "Counting blue balls",
      "heading": "How many balls are blue?",
      "options": [
        "6",
        "5",
        "4",
        "3",
        "8"
      ],
      "option_text": [
        "6",
        "5",
        "4",
        "3",
        "8"
      ],
      "option_keys": [
        "a",
        "b",
        "c",
        "d",
        "e"
      ],
      "state": {
        "problem": "in a bag of small balls 1 / 4 are green , 1 / 8 are blue , 1 / 12 are yellow and the remaining 26 white . how many balls are blue ?"
      },
      "instructions": "Which option answers the problem?",
      "dataset_label": "a",
      "probabilities": [
        [
          6.781838224864344,
          28.709399740983798,
          32.15741438204579,
          25.19233550084945,
          7.1590121512566105
        ],
        [
          24.38185572556243,
          10.733656153103418,
          22.049473029934944,
          31.118644919963025,
          11.716370171436177
        ],
        [
          14.550483527674379,
          2.1344759396220607,
          46.48020713175704,
          15.848127289977409,
          20.98670611096911
        ],
        [
          37.04200944408752,
          1.448298323772584,
          23.065493131743295,
          36.40793500541953,
          2.0362640949770836
        ]
      ],
      "logits": [
        [
          1.0889716148376465,
          2.746516227722168,
          2.8767998218536377,
          2.5963988304138184,
          1.1511435508728027
        ],
        [
          1.6802067756652832,
          0.7157226204872131,
          1.562004804611206,
          1.9670027494430542,
          0.8187038898468018
        ],
        [
          3.108205556869507,
          0.9033904075622559,
          4.4423065185546875,
          3.2063355445861816,
          3.528933525085449
        ],
        [
          3.4206666946411133,
          -0.3030266761779785,
          2.8765110969543457,
          3.4008333683013916,
          0.08836644887924194
        ]
      ],
      "agreement": [
        0.3992085457997849,
        0.5497945241247091,
        0.8736271928976953
      ],
      "threshold": 0.9225,
      "exitDepth": 4,
      "explanation": "The model selects 6 at the depth limit, matching the dataset label. White balls are 13/24 of the total: 26 ÷ (13/24) = 48; 48 ÷ 8 = 6. This explanation is editorial, not generated by Lev.",
      "source": {
        "dataset": "MathQA",
        "record_id": "loopjev-eval-v1/development/id/loopjev_mathqa/validation/883",
        "origin_id": "loopjev_mathqa/validation/883",
        "partition": "LevBench development",
        "upstream_split": "validation",
        "upstream_row": 883,
        "url": "https://huggingface.co/datasets/allenai/math_qa",
        "revision": "fafb9f7ee5b9ec4da9499f9c4177a4c91389f2d6",
        "license": "Apache-2.0",
        "attribution": "Amini et al. (2019), MathQA; AQuA-RAT, Google/DeepMind (2017).",
        "state_sha256": "2fda7c9919c3982d5d552cca839b3413d1d4eb2f91e82bad6e87aeb97a972c2b",
        "record_group": "loopjev-eval-v1/loopjev_mathqa/template/4c1d0b5bc1149dda",
        "held_out_fold": 1
      }
    }
  },
  "evidence_sha256": {
    ".runtime/experiments/phase2-s31-2682/ouro26-ls-s0/rows/eval-v1-development.json": "ba9f1640828b8a2d37f142e15a40e1b48f0b40f62e0a61da09962e0bfcbe9fdb",
    ".runtime/experiments/phase2-s31-2682/ouro26-ls-s0/rows/decision-v7-calibration.json": "e7697e34f3581cfeacebe1003a22d836a7d5028c5c62aa5a3e4ed484d7629c88",
    ".runtime/experiments/phase2-s31-2682/ouro26-ls-s0/analysis-xfit-fix1/adaptive.json": "930f4cbeb57102f5f0fde1b72be6f0d1be3a0fe00ca5fb630faffdb720fdc5d9",
    "Experiments/evals/loopjev/eval-v1/development.jsonl": "a0f8f8ad4a7c51aeb313b5ab4667ac23e9bd9e6ab184900d0a2069faa62a298c",
    "Experiments/analysis/adaptive_depth.py": "a6b97bc1afcc277a357078de64b95dd34bade9f95729f57205a3ec146afc738c"
  },
  "verification": {
    "archived_thresholds_reproduced": true,
    "archived_stop_fractions_reproduced": [
      0.09519450800915331,
      0.21556064073226544,
      0.28649885583524026,
      0.40274599542334094
    ],
    "archived_accuracy_reproduced": 0.6276887871853547,
    "locked_test_read": false
  }
}