{
  "aggregates": [
    {
      "candidate_pair_count": 2,
      "k": 1,
      "leaky_beats_honest_count": 17,
      "metrics": {
        "honest_audit_accuracy": {
          "count": 32,
          "maximum": 0.529296875,
          "mean": 0.50213623046875,
          "minimum": 0.466796875,
          "range": 0.0625,
          "sample_stdev": 0.015804003745083277
        },
        "honest_selection_accuracy": {
          "count": 32,
          "maximum": 0.5703125,
          "mean": 0.5157470703125,
          "minimum": 0.5,
          "range": 0.0703125,
          "sample_stdev": 0.015890027255812407
        },
        "honest_test_accuracy": {
          "count": 32,
          "maximum": 0.5703125,
          "mean": 0.4976806640625,
          "minimum": 0.44140625,
          "range": 0.12890625,
          "sample_stdev": 0.03099472625046343
        },
        "leaky_audit_accuracy": {
          "count": 32,
          "maximum": 0.529296875,
          "mean": 0.4962158203125,
          "minimum": 0.466796875,
          "range": 0.0625,
          "sample_stdev": 0.015482095211502926
        },
        "leaky_reported_test_accuracy": {
          "count": 32,
          "maximum": 0.5703125,
          "mean": 0.5242919921875,
          "minimum": 0.5,
          "range": 0.0703125,
          "sample_stdev": 0.01889681200940753
        },
        "leaky_vs_honest_test_gap": {
          "count": 32,
          "maximum": 0.1171875,
          "mean": 0.026611328125,
          "minimum": 0.0,
          "range": 0.1171875,
          "sample_stdev": 0.03401683915893226
        },
        "optimism_gap": {
          "count": 32,
          "maximum": 0.0849609375,
          "mean": 0.028076171875,
          "minimum": -0.013671875,
          "range": 0.0986328125,
          "sample_stdev": 0.025006053560812827
        }
      },
      "seed_count": 32
    },
    {
      "candidate_pair_count": 4,
      "k": 2,
      "leaky_beats_honest_count": 18,
      "metrics": {
        "honest_audit_accuracy": {
          "count": 32,
          "maximum": 0.5341796875,
          "mean": 0.501556396484375,
          "minimum": 0.4736328125,
          "range": 0.060546875,
          "sample_stdev": 0.013943380275572568
        },
        "honest_selection_accuracy": {
          "count": 32,
          "maximum": 0.5703125,
          "mean": 0.528564453125,
          "minimum": 0.50390625,
          "range": 0.06640625,
          "sample_stdev": 0.01711165622680472
        },
        "honest_test_accuracy": {
          "count": 32,
          "maximum": 0.578125,
          "mean": 0.508544921875,
          "minimum": 0.43359375,
          "range": 0.14453125,
          "sample_stdev": 0.0368640086967123
        },
        "leaky_audit_accuracy": {
          "count": 32,
          "maximum": 0.51953125,
          "mean": 0.493255615234375,
          "minimum": 0.4609375,
          "range": 0.05859375,
          "sample_stdev": 0.014115514601547809
        },
        "leaky_reported_test_accuracy": {
          "count": 32,
          "maximum": 0.578125,
          "mean": 0.5374755859375,
          "minimum": 0.50390625,
          "range": 0.07421875,
          "sample_stdev": 0.019665851087363607
        },
        "leaky_vs_honest_test_gap": {
          "count": 32,
          "maximum": 0.1328125,
          "mean": 0.0289306640625,
          "minimum": 0.0,
          "range": 0.1328125,
          "sample_stdev": 0.03365931826473632
        },
        "optimism_gap": {
          "count": 32,
          "maximum": 0.087890625,
          "mean": 0.044219970703125,
          "minimum": 0.0029296875,
          "range": 0.0849609375,
          "sample_stdev": 0.02351512155759446
        }
      },
      "seed_count": 32
    },
    {
      "candidate_pair_count": 8,
      "k": 4,
      "leaky_beats_honest_count": 24,
      "metrics": {
        "honest_audit_accuracy": {
          "count": 32,
          "maximum": 0.5302734375,
          "mean": 0.502716064453125,
          "minimum": 0.4814453125,
          "range": 0.048828125,
          "sample_stdev": 0.010935852700218124
        },
        "honest_selection_accuracy": {
          "count": 32,
          "maximum": 0.578125,
          "mean": 0.5472412109375,
          "minimum": 0.51953125,
          "range": 0.05859375,
          "sample_stdev": 0.014695091197089033
        },
        "honest_test_accuracy": {
          "count": 32,
          "maximum": 0.578125,
          "mean": 0.516845703125,
          "minimum": 0.43359375,
          "range": 0.14453125,
          "sample_stdev": 0.03341830917624865
        },
        "leaky_audit_accuracy": {
          "count": 32,
          "maximum": 0.5283203125,
          "mean": 0.497161865234375,
          "minimum": 0.4716796875,
          "range": 0.056640625,
          "sample_stdev": 0.014387501581173273
        },
        "leaky_reported_test_accuracy": {
          "count": 32,
          "maximum": 0.5859375,
          "mean": 0.5477294921875,
          "minimum": 0.51953125,
          "range": 0.06640625,
          "sample_stdev": 0.018948835851496343
        },
        "leaky_vs_honest_test_gap": {
          "count": 32,
          "maximum": 0.1328125,
          "mean": 0.0308837890625,
          "minimum": 0.0,
          "range": 0.1328125,
          "sample_stdev": 0.03271754716953525
        },
        "optimism_gap": {
          "count": 32,
          "maximum": 0.11328125,
          "mean": 0.050567626953125,
          "minimum": 0.009765625,
          "range": 0.103515625,
          "sample_stdev": 0.022022805359503857
        }
      },
      "seed_count": 32
    },
    {
      "candidate_pair_count": 16,
      "k": 8,
      "leaky_beats_honest_count": 29,
      "metrics": {
        "honest_audit_accuracy": {
          "count": 32,
          "maximum": 0.5302734375,
          "mean": 0.499114990234375,
          "minimum": 0.4736328125,
          "range": 0.056640625,
          "sample_stdev": 0.012230910049466165
        },
        "honest_selection_accuracy": {
          "count": 32,
          "maximum": 0.578125,
          "mean": 0.553466796875,
          "minimum": 0.5234375,
          "range": 0.0546875,
          "sample_stdev": 0.013906133194281101
        },
        "honest_test_accuracy": {
          "count": 32,
          "maximum": 0.578125,
          "mean": 0.5068359375,
          "minimum": 0.43359375,
          "range": 0.14453125,
          "sample_stdev": 0.03540041778190004
        },
        "leaky_audit_accuracy": {
          "count": 32,
          "maximum": 0.5283203125,
          "mean": 0.49853515625,
          "minimum": 0.4638671875,
          "range": 0.064453125,
          "sample_stdev": 0.01640864997807018
        },
        "leaky_reported_test_accuracy": {
          "count": 32,
          "maximum": 0.5859375,
          "mean": 0.5557861328125,
          "minimum": 0.53125,
          "range": 0.0546875,
          "sample_stdev": 0.0152175444341318
        },
        "leaky_vs_honest_test_gap": {
          "count": 32,
          "maximum": 0.15234375,
          "mean": 0.0489501953125,
          "minimum": 0.0,
          "range": 0.15234375,
          "sample_stdev": 0.03629273574079733
        },
        "optimism_gap": {
          "count": 32,
          "maximum": 0.11328125,
          "mean": 0.0572509765625,
          "minimum": 0.01953125,
          "range": 0.09375,
          "sample_stdev": 0.022032603316326334
        }
      },
      "seed_count": 32
    },
    {
      "candidate_pair_count": 32,
      "k": 16,
      "leaky_beats_honest_count": 30,
      "metrics": {
        "honest_audit_accuracy": {
          "count": 32,
          "maximum": 0.5166015625,
          "mean": 0.49737548828125,
          "minimum": 0.458984375,
          "range": 0.0576171875,
          "sample_stdev": 0.012351246150750164
        },
        "honest_selection_accuracy": {
          "count": 32,
          "maximum": 0.5859375,
          "mean": 0.56298828125,
          "minimum": 0.5390625,
          "range": 0.046875,
          "sample_stdev": 0.010582051925721654
        },
        "honest_test_accuracy": {
          "count": 32,
          "maximum": 0.578125,
          "mean": 0.510986328125,
          "minimum": 0.45703125,
          "range": 0.12109375,
          "sample_stdev": 0.031366929362865434
        },
        "leaky_audit_accuracy": {
          "count": 32,
          "maximum": 0.5283203125,
          "mean": 0.4990234375,
          "minimum": 0.4638671875,
          "range": 0.064453125,
          "sample_stdev": 0.016278771612453856
        },
        "leaky_reported_test_accuracy": {
          "count": 32,
          "maximum": 0.6171875,
          "mean": 0.567138671875,
          "minimum": 0.546875,
          "range": 0.0703125,
          "sample_stdev": 0.015159312206240765
        },
        "leaky_vs_honest_test_gap": {
          "count": 32,
          "maximum": 0.10546875,
          "mean": 0.05615234375,
          "minimum": 0.0,
          "range": 0.10546875,
          "sample_stdev": 0.033628457549106476
        },
        "optimism_gap": {
          "count": 32,
          "maximum": 0.1171875,
          "mean": 0.068115234375,
          "minimum": 0.03125,
          "range": 0.0859375,
          "sample_stdev": 0.022801454358867906
        }
      },
      "seed_count": 32
    },
    {
      "candidate_pair_count": 64,
      "k": 32,
      "leaky_beats_honest_count": 32,
      "metrics": {
        "honest_audit_accuracy": {
          "count": 32,
          "maximum": 0.513671875,
          "mean": 0.493988037109375,
          "minimum": 0.458984375,
          "range": 0.0546875,
          "sample_stdev": 0.012347937699517444
        },
        "honest_selection_accuracy": {
          "count": 32,
          "maximum": 0.6171875,
          "mean": 0.5714111328125,
          "minimum": 0.55078125,
          "range": 0.06640625,
          "sample_stdev": 0.014074898453192439
        },
        "honest_test_accuracy": {
          "count": 32,
          "maximum": 0.5546875,
          "mean": 0.504638671875,
          "minimum": 0.44921875,
          "range": 0.10546875,
          "sample_stdev": 0.027859788922462846
        },
        "leaky_audit_accuracy": {
          "count": 32,
          "maximum": 0.53515625,
          "mean": 0.50030517578125,
          "minimum": 0.470703125,
          "range": 0.064453125,
          "sample_stdev": 0.01625501361300045
        },
        "leaky_reported_test_accuracy": {
          "count": 32,
          "maximum": 0.6171875,
          "mean": 0.572021484375,
          "minimum": 0.55078125,
          "range": 0.06640625,
          "sample_stdev": 0.014444352857763711
        },
        "leaky_vs_honest_test_gap": {
          "count": 32,
          "maximum": 0.1171875,
          "mean": 0.0673828125,
          "minimum": 0.0078125,
          "range": 0.109375,
          "sample_stdev": 0.03142278342292985
        },
        "optimism_gap": {
          "count": 32,
          "maximum": 0.1171875,
          "mean": 0.07171630859375,
          "minimum": 0.03125,
          "range": 0.0859375,
          "sample_stdev": 0.020027309200362886
        }
      },
      "seed_count": 32
    },
    {
      "candidate_pair_count": 128,
      "k": 64,
      "leaky_beats_honest_count": 32,
      "metrics": {
        "honest_audit_accuracy": {
          "count": 32,
          "maximum": 0.521484375,
          "mean": 0.495086669921875,
          "minimum": 0.4677734375,
          "range": 0.0537109375,
          "sample_stdev": 0.012127663797067688
        },
        "honest_selection_accuracy": {
          "count": 32,
          "maximum": 0.6171875,
          "mean": 0.5792236328125,
          "minimum": 0.55859375,
          "range": 0.05859375,
          "sample_stdev": 0.012869200143544835
        },
        "honest_test_accuracy": {
          "count": 32,
          "maximum": 0.5546875,
          "mean": 0.499267578125,
          "minimum": 0.44921875,
          "range": 0.10546875,
          "sample_stdev": 0.03044317416796611
        },
        "leaky_audit_accuracy": {
          "count": 32,
          "maximum": 0.53515625,
          "mean": 0.50238037109375,
          "minimum": 0.470703125,
          "range": 0.064453125,
          "sample_stdev": 0.014936345510293247
        },
        "leaky_reported_test_accuracy": {
          "count": 32,
          "maximum": 0.6171875,
          "mean": 0.5787353515625,
          "minimum": 0.5625,
          "range": 0.0546875,
          "sample_stdev": 0.011910782740329712
        },
        "leaky_vs_honest_test_gap": {
          "count": 32,
          "maximum": 0.14453125,
          "mean": 0.0794677734375,
          "minimum": 0.01953125,
          "range": 0.125,
          "sample_stdev": 0.032297364260553345
        },
        "optimism_gap": {
          "count": 32,
          "maximum": 0.1171875,
          "mean": 0.07635498046875,
          "minimum": 0.0458984375,
          "range": 0.0712890625,
          "sample_stdev": 0.015790371836908786
        }
      },
      "seed_count": 32
    },
    {
      "candidate_pair_count": 256,
      "k": 128,
      "leaky_beats_honest_count": 32,
      "metrics": {
        "honest_audit_accuracy": {
          "count": 32,
          "maximum": 0.5341796875,
          "mean": 0.501068115234375,
          "minimum": 0.4775390625,
          "range": 0.056640625,
          "sample_stdev": 0.01500297470744906
        },
        "honest_selection_accuracy": {
          "count": 32,
          "maximum": 0.6171875,
          "mean": 0.5860595703125,
          "minimum": 0.56640625,
          "range": 0.05078125,
          "sample_stdev": 0.012049449105703757
        },
        "honest_test_accuracy": {
          "count": 32,
          "maximum": 0.546875,
          "mean": 0.4990234375,
          "minimum": 0.4375,
          "range": 0.109375,
          "sample_stdev": 0.02880765935301224
        },
        "leaky_audit_accuracy": {
          "count": 32,
          "maximum": 0.53515625,
          "mean": 0.503509521484375,
          "minimum": 0.478515625,
          "range": 0.056640625,
          "sample_stdev": 0.014812088571891025
        },
        "leaky_reported_test_accuracy": {
          "count": 32,
          "maximum": 0.6171875,
          "mean": 0.5887451171875,
          "minimum": 0.5703125,
          "range": 0.046875,
          "sample_stdev": 0.011957183541364266
        },
        "leaky_vs_honest_test_gap": {
          "count": 32,
          "maximum": 0.15625,
          "mean": 0.0897216796875,
          "minimum": 0.03125,
          "range": 0.125,
          "sample_stdev": 0.03343372506519152
        },
        "optimism_gap": {
          "count": 32,
          "maximum": 0.1171875,
          "mean": 0.085235595703125,
          "minimum": 0.052734375,
          "range": 0.064453125,
          "sample_stdev": 0.017616994671332112
        }
      },
      "seed_count": 32
    },
    {
      "candidate_pair_count": 512,
      "k": 256,
      "leaky_beats_honest_count": 32,
      "metrics": {
        "honest_audit_accuracy": {
          "count": 32,
          "maximum": 0.5263671875,
          "mean": 0.49700927734375,
          "minimum": 0.47265625,
          "range": 0.0537109375,
          "sample_stdev": 0.015092067377334706
        },
        "honest_selection_accuracy": {
          "count": 32,
          "maximum": 0.6171875,
          "mean": 0.5916748046875,
          "minimum": 0.57421875,
          "range": 0.04296875,
          "sample_stdev": 0.010546379123274118
        },
        "honest_test_accuracy": {
          "count": 32,
          "maximum": 0.546875,
          "mean": 0.5001220703125,
          "minimum": 0.4375,
          "range": 0.109375,
          "sample_stdev": 0.02759433663408379
        },
        "leaky_audit_accuracy": {
          "count": 32,
          "maximum": 0.53515625,
          "mean": 0.5050048828125,
          "minimum": 0.478515625,
          "range": 0.056640625,
          "sample_stdev": 0.014600581269402354
        },
        "leaky_reported_test_accuracy": {
          "count": 32,
          "maximum": 0.640625,
          "mean": 0.59765625,
          "minimum": 0.58203125,
          "range": 0.05859375,
          "sample_stdev": 0.011906261906267859
        },
        "leaky_vs_honest_test_gap": {
          "count": 32,
          "maximum": 0.15625,
          "mean": 0.0975341796875,
          "minimum": 0.05078125,
          "range": 0.10546875,
          "sample_stdev": 0.026853034392966022
        },
        "optimism_gap": {
          "count": 32,
          "maximum": 0.1376953125,
          "mean": 0.0926513671875,
          "minimum": 0.052734375,
          "range": 0.0849609375,
          "sample_stdev": 0.0191056676283871
        }
      },
      "seed_count": 32
    },
    {
      "candidate_pair_count": 1024,
      "k": 512,
      "leaky_beats_honest_count": 32,
      "metrics": {
        "honest_audit_accuracy": {
          "count": 32,
          "maximum": 0.52734375,
          "mean": 0.502593994140625,
          "minimum": 0.4677734375,
          "range": 0.0595703125,
          "sample_stdev": 0.014861333823845694
        },
        "honest_selection_accuracy": {
          "count": 32,
          "maximum": 0.640625,
          "mean": 0.5999755859375,
          "minimum": 0.58203125,
          "range": 0.05859375,
          "sample_stdev": 0.012739459724454784
        },
        "honest_test_accuracy": {
          "count": 32,
          "maximum": 0.56640625,
          "mean": 0.5084228515625,
          "minimum": 0.4453125,
          "range": 0.12109375,
          "sample_stdev": 0.024406815960837807
        },
        "leaky_audit_accuracy": {
          "count": 32,
          "maximum": 0.53515625,
          "mean": 0.502716064453125,
          "minimum": 0.478515625,
          "range": 0.056640625,
          "sample_stdev": 0.01311307250877283
        },
        "leaky_reported_test_accuracy": {
          "count": 32,
          "maximum": 0.640625,
          "mean": 0.603515625,
          "minimum": 0.5859375,
          "range": 0.0546875,
          "sample_stdev": 0.010640036205939793
        },
        "leaky_vs_honest_test_gap": {
          "count": 32,
          "maximum": 0.15234375,
          "mean": 0.0950927734375,
          "minimum": 0.0390625,
          "range": 0.11328125,
          "sample_stdev": 0.025808525386672398
        },
        "optimism_gap": {
          "count": 32,
          "maximum": 0.1376953125,
          "mean": 0.100799560546875,
          "minimum": 0.0625,
          "range": 0.0751953125,
          "sample_stdev": 0.017322028342403536
        }
      },
      "seed_count": 32
    },
    {
      "candidate_pair_count": 2048,
      "k": 1024,
      "leaky_beats_honest_count": 32,
      "metrics": {
        "honest_audit_accuracy": {
          "count": 32,
          "maximum": 0.529296875,
          "mean": 0.501434326171875,
          "minimum": 0.4677734375,
          "range": 0.0615234375,
          "sample_stdev": 0.016290076982528573
        },
        "honest_selection_accuracy": {
          "count": 32,
          "maximum": 0.640625,
          "mean": 0.60791015625,
          "minimum": 0.58984375,
          "range": 0.05078125,
          "sample_stdev": 0.010720687336676046
        },
        "honest_test_accuracy": {
          "count": 32,
          "maximum": 0.56640625,
          "mean": 0.5045166015625,
          "minimum": 0.4375,
          "range": 0.12890625,
          "sample_stdev": 0.029981653900146045
        },
        "leaky_audit_accuracy": {
          "count": 32,
          "maximum": 0.53125,
          "mean": 0.501922607421875,
          "minimum": 0.478515625,
          "range": 0.052734375,
          "sample_stdev": 0.013166336559852825
        },
        "leaky_reported_test_accuracy": {
          "count": 32,
          "maximum": 0.640625,
          "mean": 0.6085205078125,
          "minimum": 0.59375,
          "range": 0.046875,
          "sample_stdev": 0.009293495198794565
        },
        "leaky_vs_honest_test_gap": {
          "count": 32,
          "maximum": 0.16796875,
          "mean": 0.10400390625,
          "minimum": 0.05078125,
          "range": 0.1171875,
          "sample_stdev": 0.029893537732431107
        },
        "optimism_gap": {
          "count": 32,
          "maximum": 0.1376953125,
          "mean": 0.106597900390625,
          "minimum": 0.07421875,
          "range": 0.0634765625,
          "sample_stdev": 0.01580141823790984
        }
      },
      "seed_count": 32
    }
  ],
  "generated_by": "labs/test-set-reuse/build_lab.py",
  "interaction_boundary": {
    "audit_lock": "Pedagogical sequencing only; the retained counts are embedded in this static page and are not secret.",
    "evidence_effect": "None. Changing controls or exporting learner state does not create research evidence.",
    "execution": "No experiment module, model, seed, candidate search, or application-initiated data request runs in the browser or builder."
  },
  "lesson_id": "test-set-reuse-evidence-lab-v1",
  "minimal_claim": "Under this fixed synthetic protocol, selecting the best of 1,024 noise features and polarity on the test split produced 60.85% mean reported test accuracy across 32 retained seeds. The same frozen choices averaged 50.19% on sealed audit. Mean leaky-reported-test minus honest-test gap was 10.40 percentage points.",
  "projection": {
    "candidate_feature_counts": [
      1,
      2,
      4,
      8,
      16,
      32,
      64,
      128,
      256,
      512,
      1024
    ],
    "rows": [
      {
        "candidate_pair_count": 2,
        "derived": {
          "honest_audit_accuracy": 0.521484375,
          "honest_selection_accuracy": 0.50390625,
          "honest_test_accuracy": 0.48828125,
          "leaky_audit_accuracy": 0.478515625,
          "leaky_reported_test_accuracy": 0.51171875,
          "leaky_vs_honest_test_gap": 0.0234375,
          "optimism_gap": 0.033203125
        },
        "honest": {
          "audit_correct": 534,
          "audit_total": 1024,
          "feature_index": 0,
          "polarity": -1,
          "selection_correct": 129,
          "selection_total": 256,
          "test_correct": 125,
          "test_total": 256
        },
        "k": 1,
        "leaky": {
          "audit_correct": 490,
          "audit_total": 1024,
          "feature_index": 0,
          "polarity": 1,
          "test_correct": 131,
          "test_total": 256
        },
        "seed": 1
      },
      {
        "candidate_pair_count": 4,
        "derived": {
          "honest_audit_accuracy": 0.521484375,
          "honest_selection_accuracy": 0.50390625,
          "honest_test_accuracy": 0.48828125,
          "leaky_audit_accuracy": 0.4609375,
          "leaky_reported_test_accuracy": 0.53125,
          "leaky_vs_honest_test_gap": 0.04296875,
          "optimism_gap": 0.0703125
        },
        "honest": {
          "audit_correct": 534,
          "audit_total": 1024,
          "feature_index": 0,
          "polarity": -1,
          "selection_correct": 129,
          "selection_total": 256,
          "test_correct": 125,
          "test_total": 256
        },
        "k": 2,
        "leaky": {
          "audit_correct": 472,
          "audit_total": 1024,
          "feature_index": 1,
          "polarity": -1,
          "test_correct": 136,
          "test_total": 256
        },
        "seed": 1
      },
      {
        "candidate_pair_count": 8,
        "derived": {
          "honest_audit_accuracy": 0.5087890625,
          "honest_selection_accuracy": 0.52734375,
          "honest_test_accuracy": 0.546875,
          "leaky_audit_accuracy": 0.5146484375,
          "leaky_reported_test_accuracy": 0.55859375,
          "leaky_vs_honest_test_gap": 0.01171875,
          "optimism_gap": 0.0439453125
        },
        "honest": {
          "audit_correct": 521,
          "audit_total": 1024,
          "feature_index": 3,
          "polarity": 1,
          "selection_correct": 135,
          "selection_total": 256,
          "test_correct": 140,
          "test_total": 256
        },
        "k": 4,
        "leaky": {
          "audit_correct": 527,
          "audit_total": 1024,
          "feature_index": 2,
          "polarity": -1,
          "test_correct": 143,
          "test_total": 256
        },
        "seed": 1
      },
      {
        "candidate_pair_count": 16,
        "derived": {
          "honest_audit_accuracy": 0.4736328125,
          "honest_selection_accuracy": 0.5390625,
          "honest_test_accuracy": 0.51953125,
          "leaky_audit_accuracy": 0.5146484375,
          "leaky_reported_test_accuracy": 0.55859375,
          "leaky_vs_honest_test_gap": 0.0390625,
          "optimism_gap": 0.0439453125
        },
        "honest": {
          "audit_correct": 485,
          "audit_total": 1024,
          "feature_index": 6,
          "polarity": -1,
          "selection_correct": 138,
          "selection_total": 256,
          "test_correct": 133,
          "test_total": 256
        },
        "k": 8,
        "leaky": {
          "audit_correct": 527,
          "audit_total": 1024,
          "feature_index": 2,
          "polarity": -1,
          "test_correct": 143,
          "test_total": 256
        },
        "seed": 1
      },
      {
        "candidate_pair_count": 32,
        "derived": {
          "honest_audit_accuracy": 0.458984375,
          "honest_selection_accuracy": 0.5625,
          "honest_test_accuracy": 0.5546875,
          "leaky_audit_accuracy": 0.5146484375,
          "leaky_reported_test_accuracy": 0.55859375,
          "leaky_vs_honest_test_gap": 0.00390625,
          "optimism_gap": 0.0439453125
        },
        "honest": {
          "audit_correct": 470,
          "audit_total": 1024,
          "feature_index": 9,
          "polarity": 1,
          "selection_correct": 144,
          "selection_total": 256,
          "test_correct": 142,
          "test_total": 256
        },
        "k": 16,
        "leaky": {
          "audit_correct": 527,
          "audit_total": 1024,
          "feature_index": 2,
          "polarity": -1,
          "test_correct": 143,
          "test_total": 256
        },
        "seed": 1
      },
      {
        "candidate_pair_count": 64,
        "derived": {
          "honest_audit_accuracy": 0.458984375,
          "honest_selection_accuracy": 0.5625,
          "honest_test_accuracy": 0.5546875,
          "leaky_audit_accuracy": 0.49609375,
          "leaky_reported_test_accuracy": 0.56640625,
          "leaky_vs_honest_test_gap": 0.01171875,
          "optimism_gap": 0.0703125
        },
        "honest": {
          "audit_correct": 470,
          "audit_total": 1024,
          "feature_index": 9,
          "polarity": 1,
          "selection_correct": 144,
          "selection_total": 256,
          "test_correct": 142,
          "test_total": 256
        },
        "k": 32,
        "leaky": {
          "audit_correct": 508,
          "audit_total": 1024,
          "feature_index": 25,
          "polarity": -1,
          "test_correct": 145,
          "test_total": 256
        },
        "seed": 1
      },
      {
        "candidate_pair_count": 128,
        "derived": {
          "honest_audit_accuracy": 0.513671875,
          "honest_selection_accuracy": 0.5859375,
          "honest_test_accuracy": 0.53515625,
          "leaky_audit_accuracy": 0.49609375,
          "leaky_reported_test_accuracy": 0.56640625,
          "leaky_vs_honest_test_gap": 0.03125,
          "optimism_gap": 0.0703125
        },
        "honest": {
          "audit_correct": 526,
          "audit_total": 1024,
          "feature_index": 56,
          "polarity": -1,
          "selection_correct": 150,
          "selection_total": 256,
          "test_correct": 137,
          "test_total": 256
        },
        "k": 64,
        "leaky": {
          "audit_correct": 508,
          "audit_total": 1024,
          "feature_index": 25,
          "polarity": -1,
          "test_correct": 145,
          "test_total": 256
        },
        "seed": 1
      },
      {
        "candidate_pair_count": 256,
        "derived": {
          "honest_audit_accuracy": 0.513671875,
          "honest_selection_accuracy": 0.5859375,
          "honest_test_accuracy": 0.53515625,
          "leaky_audit_accuracy": 0.5029296875,
          "leaky_reported_test_accuracy": 0.58984375,
          "leaky_vs_honest_test_gap": 0.0546875,
          "optimism_gap": 0.0869140625
        },
        "honest": {
          "audit_correct": 526,
          "audit_total": 1024,
          "feature_index": 56,
          "polarity": -1,
          "selection_correct": 150,
          "selection_total": 256,
          "test_correct": 137,
          "test_total": 256
        },
        "k": 128,
        "leaky": {
          "audit_correct": 515,
          "audit_total": 1024,
          "feature_index": 67,
          "polarity": -1,
          "test_correct": 151,
          "test_total": 256
        },
        "seed": 1
      },
      {
        "candidate_pair_count": 512,
        "derived": {
          "honest_audit_accuracy": 0.501953125,
          "honest_selection_accuracy": 0.59375,
          "honest_test_accuracy": 0.484375,
          "leaky_audit_accuracy": 0.5029296875,
          "leaky_reported_test_accuracy": 0.58984375,
          "leaky_vs_honest_test_gap": 0.10546875,
          "optimism_gap": 0.0869140625
        },
        "honest": {
          "audit_correct": 514,
          "audit_total": 1024,
          "feature_index": 189,
          "polarity": -1,
          "selection_correct": 152,
          "selection_total": 256,
          "test_correct": 124,
          "test_total": 256
        },
        "k": 256,
        "leaky": {
          "audit_correct": 515,
          "audit_total": 1024,
          "feature_index": 67,
          "polarity": -1,
          "test_correct": 151,
          "test_total": 256
        },
        "seed": 1
      },
      {
        "candidate_pair_count": 1024,
        "derived": {
          "honest_audit_accuracy": 0.515625,
          "honest_selection_accuracy": 0.60546875,
          "honest_test_accuracy": 0.4921875,
          "leaky_audit_accuracy": 0.50390625,
          "leaky_reported_test_accuracy": 0.59375,
          "leaky_vs_honest_test_gap": 0.1015625,
          "optimism_gap": 0.08984375
        },
        "honest": {
          "audit_correct": 528,
          "audit_total": 1024,
          "feature_index": 483,
          "polarity": -1,
          "selection_correct": 155,
          "selection_total": 256,
          "test_correct": 126,
          "test_total": 256
        },
        "k": 512,
        "leaky": {
          "audit_correct": 516,
          "audit_total": 1024,
          "feature_index": 507,
          "polarity": -1,
          "test_correct": 152,
          "test_total": 256
        },
        "seed": 1
      },
      {
        "candidate_pair_count": 2048,
        "derived": {
          "honest_audit_accuracy": 0.515625,
          "honest_selection_accuracy": 0.60546875,
          "honest_test_accuracy": 0.4921875,
          "leaky_audit_accuracy": 0.4794921875,
          "leaky_reported_test_accuracy": 0.60546875,
          "leaky_vs_honest_test_gap": 0.11328125,
          "optimism_gap": 0.1259765625
        },
        "honest": {
          "audit_correct": 528,
          "audit_total": 1024,
          "feature_index": 483,
          "polarity": -1,
          "selection_correct": 155,
          "selection_total": 256,
          "test_correct": 126,
          "test_total": 256
        },
        "k": 1024,
        "leaky": {
          "audit_correct": 491,
          "audit_total": 1024,
          "feature_index": 614,
          "polarity": 1,
          "test_correct": 155,
          "test_total": 256
        },
        "seed": 1
      },
      {
        "candidate_pair_count": 2,
        "derived": {
          "honest_audit_accuracy": 0.498046875,
          "honest_selection_accuracy": 0.5703125,
          "honest_test_accuracy": 0.45703125,
          "leaky_audit_accuracy": 0.501953125,
          "leaky_reported_test_accuracy": 0.54296875,
          "leaky_vs_honest_test_gap": 0.0859375,
          "optimism_gap": 0.041015625
        },
        "honest": {
          "audit_correct": 510,
          "audit_total": 1024,
          "feature_index": 0,
          "polarity": 1,
          "selection_correct": 146,
          "selection_total": 256,
          "test_correct": 117,
          "test_total": 256
        },
        "k": 1,
        "leaky": {
          "audit_correct": 514,
          "audit_total": 1024,
          "feature_index": 0,
          "polarity": -1,
          "test_correct": 139,
          "test_total": 256
        },
        "seed": 2
      },
      {
        "candidate_pair_count": 4,
        "derived": {
          "honest_audit_accuracy": 0.498046875,
          "honest_selection_accuracy": 0.5703125,
          "honest_test_accuracy": 0.45703125,
          "leaky_audit_accuracy": 0.501953125,
          "leaky_reported_test_accuracy": 0.54296875,
          "leaky_vs_honest_test_gap": 0.0859375,
          "optimism_gap": 0.041015625
        },
        "honest": {
          "audit_correct": 510,
          "audit_total": 1024,
          "feature_index": 0,
          "polarity": 1,
          "selection_correct": 146,
          "selection_total": 256,
          "test_correct": 117,
          "test_total": 256
        },
        "k": 2,
        "leaky": {
          "audit_correct": 514,
          "audit_total": 1024,
          "feature_index": 0,
          "polarity": -1,
          "test_correct": 139,
          "test_total": 256
        },
        "seed": 2
      },
      {
        "candidate_pair_count": 8,
        "derived": {
          "honest_audit_accuracy": 0.498046875,
          "honest_selection_accuracy": 0.5703125,
          "honest_test_accuracy": 0.45703125,
          "leaky_audit_accuracy": 0.501953125,
          "leaky_reported_test_accuracy": 0.54296875,
          "leaky_vs_honest_test_gap": 0.0859375,
          "optimism_gap": 0.041015625
        },
        "honest": {
          "audit_correct": 510,
          "audit_total": 1024,
          "feature_index": 0,
          "polarity": 1,
          "selection_correct": 146,
          "selection_total": 256,
          "test_correct": 117,
          "test_total": 256
        },
        "k": 4,
        "leaky": {
          "audit_correct": 514,
          "audit_total": 1024,
          "feature_index": 0,
          "polarity": -1,
          "test_correct": 139,
          "test_total": 256
        },
        "seed": 2
      },
      {
        "candidate_pair_count": 16,
        "derived": {
          "honest_audit_accuracy": 0.498046875,
          "honest_selection_accuracy": 0.5703125,
          "honest_test_accuracy": 0.45703125,
          "leaky_audit_accuracy": 0.501953125,
          "leaky_reported_test_accuracy": 0.54296875,
          "leaky_vs_honest_test_gap": 0.0859375,
          "optimism_gap": 0.041015625
        },
        "honest": {
          "audit_correct": 510,
          "audit_total": 1024,
          "feature_index": 0,
          "polarity": 1,
          "selection_correct": 146,
          "selection_total": 256,
          "test_correct": 117,
          "test_total": 256
        },
        "k": 8,
        "leaky": {
          "audit_correct": 514,
          "audit_total": 1024,
          "feature_index": 0,
          "polarity": -1,
          "test_correct": 139,
          "test_total": 256
        },
        "seed": 2
      },
      {
        "candidate_pair_count": 32,
        "derived": {
          "honest_audit_accuracy": 0.498046875,
          "honest_selection_accuracy": 0.5703125,
          "honest_test_accuracy": 0.45703125,
          "leaky_audit_accuracy": 0.5126953125,
          "leaky_reported_test_accuracy": 0.5625,
          "leaky_vs_honest_test_gap": 0.10546875,
          "optimism_gap": 0.0498046875
        },
        "honest": {
          "audit_correct": 510,
          "audit_total": 1024,
          "feature_index": 0,
          "polarity": 1,
          "selection_correct": 146,
          "selection_total": 256,
          "test_correct": 117,
          "test_total": 256
        },
        "k": 16,
        "leaky": {
          "audit_correct": 525,
          "audit_total": 1024,
          "feature_index": 10,
          "polarity": 1,
          "test_correct": 144,
          "test_total": 256
        },
        "seed": 2
      },
      {
        "candidate_pair_count": 64,
        "derived": {
          "honest_audit_accuracy": 0.498046875,
          "honest_selection_accuracy": 0.5703125,
          "honest_test_accuracy": 0.45703125,
          "leaky_audit_accuracy": 0.5126953125,
          "leaky_reported_test_accuracy": 0.5625,
          "leaky_vs_honest_test_gap": 0.10546875,
          "optimism_gap": 0.0498046875
        },
        "honest": {
          "audit_correct": 510,
          "audit_total": 1024,
          "feature_index": 0,
          "polarity": 1,
          "selection_correct": 146,
          "selection_total": 256,
          "test_correct": 117,
          "test_total": 256
        },
        "k": 32,
        "leaky": {
          "audit_correct": 525,
          "audit_total": 1024,
          "feature_index": 10,
          "polarity": 1,
          "test_correct": 144,
          "test_total": 256
        },
        "seed": 2
      },
      {
        "candidate_pair_count": 128,
        "derived": {
          "honest_audit_accuracy": 0.498046875,
          "honest_selection_accuracy": 0.5703125,
          "honest_test_accuracy": 0.45703125,
          "leaky_audit_accuracy": 0.5126953125,
          "leaky_reported_test_accuracy": 0.5859375,
          "leaky_vs_honest_test_gap": 0.12890625,
          "optimism_gap": 0.0732421875
        },
        "honest": {
          "audit_correct": 510,
          "audit_total": 1024,
          "feature_index": 0,
          "polarity": 1,
          "selection_correct": 146,
          "selection_total": 256,
          "test_correct": 117,
          "test_total": 256
        },
        "k": 64,
        "leaky": {
          "audit_correct": 525,
          "audit_total": 1024,
          "feature_index": 41,
          "polarity": -1,
          "test_correct": 150,
          "test_total": 256
        },
        "seed": 2
      },
      {
        "candidate_pair_count": 256,
        "derived": {
          "honest_audit_accuracy": 0.5146484375,
          "honest_selection_accuracy": 0.57421875,
          "honest_test_accuracy": 0.515625,
          "leaky_audit_accuracy": 0.50390625,
          "leaky_reported_test_accuracy": 0.59765625,
          "leaky_vs_honest_test_gap": 0.08203125,
          "optimism_gap": 0.09375
        },
        "honest": {
          "audit_correct": 527,
          "audit_total": 1024,
          "feature_index": 120,
          "polarity": 1,
          "selection_correct": 147,
          "selection_total": 256,
          "test_correct": 132,
          "test_total": 256
        },
        "k": 128,
        "leaky": {
          "audit_correct": 516,
          "audit_total": 1024,
          "feature_index": 87,
          "polarity": 1,
          "test_correct": 153,
          "test_total": 256
        },
        "seed": 2
      },
      {
        "candidate_pair_count": 512,
        "derived": {
          "honest_audit_accuracy": 0.5,
          "honest_selection_accuracy": 0.578125,
          "honest_test_accuracy": 0.484375,
          "leaky_audit_accuracy": 0.50390625,
          "leaky_reported_test_accuracy": 0.59765625,
          "leaky_vs_honest_test_gap": 0.11328125,
          "optimism_gap": 0.09375
        },
        "honest": {
          "audit_correct": 512,
          "audit_total": 1024,
          "feature_index": 208,
          "polarity": -1,
          "selection_correct": 148,
          "selection_total": 256,
          "test_correct": 124,
          "test_total": 256
        },
        "k": 256,
        "leaky": {
          "audit_correct": 516,
          "audit_total": 1024,
          "feature_index": 87,
          "polarity": 1,
          "test_correct": 153,
          "test_total": 256
        },
        "seed": 2
      },
      {
        "candidate_pair_count": 1024,
        "derived": {
          "honest_audit_accuracy": 0.4931640625,
          "honest_selection_accuracy": 0.5859375,
          "honest_test_accuracy": 0.49609375,
          "leaky_audit_accuracy": 0.50390625,
          "leaky_reported_test_accuracy": 0.59765625,
          "leaky_vs_honest_test_gap": 0.1015625,
          "optimism_gap": 0.09375
        },
        "honest": {
          "audit_correct": 505,
          "audit_total": 1024,
          "feature_index": 382,
          "polarity": 1,
          "selection_correct": 150,
          "selection_total": 256,
          "test_correct": 127,
          "test_total": 256
        },
        "k": 512,
        "leaky": {
          "audit_correct": 516,
          "audit_total": 1024,
          "feature_index": 87,
          "polarity": 1,
          "test_correct": 153,
          "test_total": 256
        },
        "seed": 2
      },
      {
        "candidate_pair_count": 2048,
        "derived": {
          "honest_audit_accuracy": 0.5166015625,
          "honest_selection_accuracy": 0.61328125,
          "honest_test_accuracy": 0.48828125,
          "leaky_audit_accuracy": 0.521484375,
          "leaky_reported_test_accuracy": 0.60546875,
          "leaky_vs_honest_test_gap": 0.1171875,
          "optimism_gap": 0.083984375
        },
        "honest": {
          "audit_correct": 529,
          "audit_total": 1024,
          "feature_index": 589,
          "polarity": 1,
          "selection_correct": 157,
          "selection_total": 256,
          "test_correct": 125,
          "test_total": 256
        },
        "k": 1024,
        "leaky": {
          "audit_correct": 534,
          "audit_total": 1024,
          "feature_index": 544,
          "polarity": -1,
          "test_correct": 155,
          "test_total": 256
        },
        "seed": 2
      },
      {
        "candidate_pair_count": 2,
        "derived": {
          "honest_audit_accuracy": 0.51171875,
          "honest_selection_accuracy": 0.5390625,
          "honest_test_accuracy": 0.4921875,
          "leaky_audit_accuracy": 0.48828125,
          "leaky_reported_test_accuracy": 0.5078125,
          "leaky_vs_honest_test_gap": 0.015625,
          "optimism_gap": 0.01953125
        },
        "honest": {
          "audit_correct": 524,
          "audit_total": 1024,
          "feature_index": 0,
          "polarity": -1,
          "selection_correct": 138,
          "selection_total": 256,
          "test_correct": 126,
          "test_total": 256
        },
        "k": 1,
        "leaky": {
          "audit_correct": 500,
          "audit_total": 1024,
          "feature_index": 0,
          "polarity": 1,
          "test_correct": 130,
          "test_total": 256
        },
        "seed": 3
      },
      {
        "candidate_pair_count": 4,
        "derived": {
          "honest_audit_accuracy": 0.4931640625,
          "honest_selection_accuracy": 0.55078125,
          "honest_test_accuracy": 0.43359375,
          "leaky_audit_accuracy": 0.5068359375,
          "leaky_reported_test_accuracy": 0.56640625,
          "leaky_vs_honest_test_gap": 0.1328125,
          "optimism_gap": 0.0595703125
        },
        "honest": {
          "audit_correct": 505,
          "audit_total": 1024,
          "feature_index": 1,
          "polarity": -1,
          "selection_correct": 141,
          "selection_total": 256,
          "test_correct": 111,
          "test_total": 256
        },
        "k": 2,
        "leaky": {
          "audit_correct": 519,
          "audit_total": 1024,
          "feature_index": 1,
          "polarity": 1,
          "test_correct": 145,
          "test_total": 256
        },
        "seed": 3
      },
      {
        "candidate_pair_count": 8,
        "derived": {
          "honest_audit_accuracy": 0.4931640625,
          "honest_selection_accuracy": 0.55078125,
          "honest_test_accuracy": 0.43359375,
          "leaky_audit_accuracy": 0.5068359375,
          "leaky_reported_test_accuracy": 0.56640625,
          "leaky_vs_honest_test_gap": 0.1328125,
          "optimism_gap": 0.0595703125
        },
        "honest": {
          "audit_correct": 505,
          "audit_total": 1024,
          "feature_index": 1,
          "polarity": -1,
          "selection_correct": 141,
          "selection_total": 256,
          "test_correct": 111,
          "test_total": 256
        },
        "k": 4,
        "leaky": {
          "audit_correct": 519,
          "audit_total": 1024,
          "feature_index": 1,
          "polarity": 1,
          "test_correct": 145,
          "test_total": 256
        },
        "seed": 3
      },
      {
        "candidate_pair_count": 16,
        "derived": {
          "honest_audit_accuracy": 0.4931640625,
          "honest_selection_accuracy": 0.55078125,
          "honest_test_accuracy": 0.43359375,
          "leaky_audit_accuracy": 0.486328125,
          "leaky_reported_test_accuracy": 0.5859375,
          "leaky_vs_honest_test_gap": 0.15234375,
          "optimism_gap": 0.099609375
        },
        "honest": {
          "audit_correct": 505,
          "audit_total": 1024,
          "feature_index": 1,
          "polarity": -1,
          "selection_correct": 141,
          "selection_total": 256,
          "test_correct": 111,
          "test_total": 256
        },
        "k": 8,
        "leaky": {
          "audit_correct": 498,
          "audit_total": 1024,
          "feature_index": 7,
          "polarity": 1,
          "test_correct": 150,
          "test_total": 256
        },
        "seed": 3
      },
      {
        "candidate_pair_count": 32,
        "derived": {
          "honest_audit_accuracy": 0.5048828125,
          "honest_selection_accuracy": 0.57421875,
          "honest_test_accuracy": 0.484375,
          "leaky_audit_accuracy": 0.486328125,
          "leaky_reported_test_accuracy": 0.5859375,
          "leaky_vs_honest_test_gap": 0.1015625,
          "optimism_gap": 0.099609375
        },
        "honest": {
          "audit_correct": 517,
          "audit_total": 1024,
          "feature_index": 8,
          "polarity": -1,
          "selection_correct": 147,
          "selection_total": 256,
          "test_correct": 124,
          "test_total": 256
        },
        "k": 16,
        "leaky": {
          "audit_correct": 498,
          "audit_total": 1024,
          "feature_index": 7,
          "polarity": 1,
          "test_correct": 150,
          "test_total": 256
        },
        "seed": 3
      },
      {
        "candidate_pair_count": 64,
        "derived": {
          "honest_audit_accuracy": 0.501953125,
          "honest_selection_accuracy": 0.5859375,
          "honest_test_accuracy": 0.50390625,
          "leaky_audit_accuracy": 0.486328125,
          "leaky_reported_test_accuracy": 0.5859375,
          "leaky_vs_honest_test_gap": 0.08203125,
          "optimism_gap": 0.099609375
        },
        "honest": {
          "audit_correct": 514,
          "audit_total": 1024,
          "feature_index": 16,
          "polarity": 1,
          "selection_correct": 150,
          "selection_total": 256,
          "test_correct": 129,
          "test_total": 256
        },
        "k": 32,
        "leaky": {
          "audit_correct": 498,
          "audit_total": 1024,
          "feature_index": 7,
          "polarity": 1,
          "test_correct": 150,
          "test_total": 256
        },
        "seed": 3
      },
      {
        "candidate_pair_count": 128,
        "derived": {
          "honest_audit_accuracy": 0.501953125,
          "honest_selection_accuracy": 0.5859375,
          "honest_test_accuracy": 0.50390625,
          "leaky_audit_accuracy": 0.486328125,
          "leaky_reported_test_accuracy": 0.5859375,
          "leaky_vs_honest_test_gap": 0.08203125,
          "optimism_gap": 0.099609375
        },
        "honest": {
          "audit_correct": 514,
          "audit_total": 1024,
          "feature_index": 16,
          "polarity": 1,
          "selection_correct": 150,
          "selection_total": 256,
          "test_correct": 129,
          "test_total": 256
        },
        "k": 64,
        "leaky": {
          "audit_correct": 498,
          "audit_total": 1024,
          "feature_index": 7,
          "polarity": 1,
          "test_correct": 150,
          "test_total": 256
        },
        "seed": 3
      },
      {
        "candidate_pair_count": 256,
        "derived": {
          "honest_audit_accuracy": 0.501953125,
          "honest_selection_accuracy": 0.5859375,
          "honest_test_accuracy": 0.50390625,
          "leaky_audit_accuracy": 0.486328125,
          "leaky_reported_test_accuracy": 0.5859375,
          "leaky_vs_honest_test_gap": 0.08203125,
          "optimism_gap": 0.099609375
        },
        "honest": {
          "audit_correct": 514,
          "audit_total": 1024,
          "feature_index": 16,
          "polarity": 1,
          "selection_correct": 150,
          "selection_total": 256,
          "test_correct": 129,
          "test_total": 256
        },
        "k": 128,
        "leaky": {
          "audit_correct": 498,
          "audit_total": 1024,
          "feature_index": 7,
          "polarity": 1,
          "test_correct": 150,
          "test_total": 256
        },
        "seed": 3
      },
      {
        "candidate_pair_count": 512,
        "derived": {
          "honest_audit_accuracy": 0.5029296875,
          "honest_selection_accuracy": 0.58984375,
          "honest_test_accuracy": 0.546875,
          "leaky_audit_accuracy": 0.5009765625,
          "leaky_reported_test_accuracy": 0.59765625,
          "leaky_vs_honest_test_gap": 0.05078125,
          "optimism_gap": 0.0966796875
        },
        "honest": {
          "audit_correct": 515,
          "audit_total": 1024,
          "feature_index": 145,
          "polarity": -1,
          "selection_correct": 151,
          "selection_total": 256,
          "test_correct": 140,
          "test_total": 256
        },
        "k": 256,
        "leaky": {
          "audit_correct": 513,
          "audit_total": 1024,
          "feature_index": 188,
          "polarity": -1,
          "test_correct": 153,
          "test_total": 256
        },
        "seed": 3
      },
      {
        "candidate_pair_count": 1024,
        "derived": {
          "honest_audit_accuracy": 0.5029296875,
          "honest_selection_accuracy": 0.58984375,
          "honest_test_accuracy": 0.546875,
          "leaky_audit_accuracy": 0.4892578125,
          "leaky_reported_test_accuracy": 0.61328125,
          "leaky_vs_honest_test_gap": 0.06640625,
          "optimism_gap": 0.1240234375
        },
        "honest": {
          "audit_correct": 515,
          "audit_total": 1024,
          "feature_index": 145,
          "polarity": -1,
          "selection_correct": 151,
          "selection_total": 256,
          "test_correct": 140,
          "test_total": 256
        },
        "k": 512,
        "leaky": {
          "audit_correct": 501,
          "audit_total": 1024,
          "feature_index": 429,
          "polarity": -1,
          "test_correct": 157,
          "test_total": 256
        },
        "seed": 3
      },
      {
        "candidate_pair_count": 2048,
        "derived": {
          "honest_audit_accuracy": 0.49609375,
          "honest_selection_accuracy": 0.60546875,
          "honest_test_accuracy": 0.5,
          "leaky_audit_accuracy": 0.4892578125,
          "leaky_reported_test_accuracy": 0.61328125,
          "leaky_vs_honest_test_gap": 0.11328125,
          "optimism_gap": 0.1240234375
        },
        "honest": {
          "audit_correct": 508,
          "audit_total": 1024,
          "feature_index": 775,
          "polarity": -1,
          "selection_correct": 155,
          "selection_total": 256,
          "test_correct": 128,
          "test_total": 256
        },
        "k": 1024,
        "leaky": {
          "audit_correct": 501,
          "audit_total": 1024,
          "feature_index": 429,
          "polarity": -1,
          "test_correct": 157,
          "test_total": 256
        },
        "seed": 3
      },
      {
        "candidate_pair_count": 2,
        "derived": {
          "honest_audit_accuracy": 0.4931640625,
          "honest_selection_accuracy": 0.51171875,
          "honest_test_accuracy": 0.51171875,
          "leaky_audit_accuracy": 0.4931640625,
          "leaky_reported_test_accuracy": 0.51171875,
          "leaky_vs_honest_test_gap": 0.0,
          "optimism_gap": 0.0185546875
        },
        "honest": {
          "audit_correct": 505,
          "audit_total": 1024,
          "feature_index": 0,
          "polarity": 1,
          "selection_correct": 131,
          "selection_total": 256,
          "test_correct": 131,
          "test_total": 256
        },
        "k": 1,
        "leaky": {
          "audit_correct": 505,
          "audit_total": 1024,
          "feature_index": 0,
          "polarity": 1,
          "test_correct": 131,
          "test_total": 256
        },
        "seed": 4
      },
      {
        "candidate_pair_count": 4,
        "derived": {
          "honest_audit_accuracy": 0.486328125,
          "honest_selection_accuracy": 0.52734375,
          "honest_test_accuracy": 0.52734375,
          "leaky_audit_accuracy": 0.486328125,
          "leaky_reported_test_accuracy": 0.52734375,
          "leaky_vs_honest_test_gap": 0.0,
          "optimism_gap": 0.041015625
        },
        "honest": {
          "audit_correct": 498,
          "audit_total": 1024,
          "feature_index": 1,
          "polarity": 1,
          "selection_correct": 135,
          "selection_total": 256,
          "test_correct": 135,
          "test_total": 256
        },
        "k": 2,
        "leaky": {
          "audit_correct": 498,
          "audit_total": 1024,
          "feature_index": 1,
          "polarity": 1,
          "test_correct": 135,
          "test_total": 256
        },
        "seed": 4
      },
      {
        "candidate_pair_count": 8,
        "derived": {
          "honest_audit_accuracy": 0.4931640625,
          "honest_selection_accuracy": 0.5625,
          "honest_test_accuracy": 0.4921875,
          "leaky_audit_accuracy": 0.486328125,
          "leaky_reported_test_accuracy": 0.52734375,
          "leaky_vs_honest_test_gap": 0.03515625,
          "optimism_gap": 0.041015625
        },
        "honest": {
          "audit_correct": 505,
          "audit_total": 1024,
          "feature_index": 2,
          "polarity": 1,
          "selection_correct": 144,
          "selection_total": 256,
          "test_correct": 126,
          "test_total": 256
        },
        "k": 4,
        "leaky": {
          "audit_correct": 498,
          "audit_total": 1024,
          "feature_index": 1,
          "polarity": 1,
          "test_correct": 135,
          "test_total": 256
        },
        "seed": 4
      },
      {
        "candidate_pair_count": 16,
        "derived": {
          "honest_audit_accuracy": 0.4931640625,
          "honest_selection_accuracy": 0.5625,
          "honest_test_accuracy": 0.4921875,
          "leaky_audit_accuracy": 0.4833984375,
          "leaky_reported_test_accuracy": 0.5390625,
          "leaky_vs_honest_test_gap": 0.046875,
          "optimism_gap": 0.0556640625
        },
        "honest": {
          "audit_correct": 505,
          "audit_total": 1024,
          "feature_index": 2,
          "polarity": 1,
          "selection_correct": 144,
          "selection_total": 256,
          "test_correct": 126,
          "test_total": 256
        },
        "k": 8,
        "leaky": {
          "audit_correct": 495,
          "audit_total": 1024,
          "feature_index": 7,
          "polarity": -1,
          "test_correct": 138,
          "test_total": 256
        },
        "seed": 4
      },
      {
        "candidate_pair_count": 32,
        "derived": {
          "honest_audit_accuracy": 0.4931640625,
          "honest_selection_accuracy": 0.5625,
          "honest_test_accuracy": 0.4921875,
          "leaky_audit_accuracy": 0.5,
          "leaky_reported_test_accuracy": 0.5859375,
          "leaky_vs_honest_test_gap": 0.09375,
          "optimism_gap": 0.0859375
        },
        "honest": {
          "audit_correct": 505,
          "audit_total": 1024,
          "feature_index": 2,
          "polarity": 1,
          "selection_correct": 144,
          "selection_total": 256,
          "test_correct": 126,
          "test_total": 256
        },
        "k": 16,
        "leaky": {
          "audit_correct": 512,
          "audit_total": 1024,
          "feature_index": 11,
          "polarity": -1,
          "test_correct": 150,
          "test_total": 256
        },
        "seed": 4
      },
      {
        "candidate_pair_count": 64,
        "derived": {
          "honest_audit_accuracy": 0.4951171875,
          "honest_selection_accuracy": 0.5859375,
          "honest_test_accuracy": 0.51953125,
          "leaky_audit_accuracy": 0.5,
          "leaky_reported_test_accuracy": 0.5859375,
          "leaky_vs_honest_test_gap": 0.06640625,
          "optimism_gap": 0.0859375
        },
        "honest": {
          "audit_correct": 507,
          "audit_total": 1024,
          "feature_index": 26,
          "polarity": 1,
          "selection_correct": 150,
          "selection_total": 256,
          "test_correct": 133,
          "test_total": 256
        },
        "k": 32,
        "leaky": {
          "audit_correct": 512,
          "audit_total": 1024,
          "feature_index": 11,
          "polarity": -1,
          "test_correct": 150,
          "test_total": 256
        },
        "seed": 4
      },
      {
        "candidate_pair_count": 128,
        "derived": {
          "honest_audit_accuracy": 0.4951171875,
          "honest_selection_accuracy": 0.5859375,
          "honest_test_accuracy": 0.51953125,
          "leaky_audit_accuracy": 0.5,
          "leaky_reported_test_accuracy": 0.5859375,
          "leaky_vs_honest_test_gap": 0.06640625,
          "optimism_gap": 0.0859375
        },
        "honest": {
          "audit_correct": 507,
          "audit_total": 1024,
          "feature_index": 26,
          "polarity": 1,
          "selection_correct": 150,
          "selection_total": 256,
          "test_correct": 133,
          "test_total": 256
        },
        "k": 64,
        "leaky": {
          "audit_correct": 512,
          "audit_total": 1024,
          "feature_index": 11,
          "polarity": -1,
          "test_correct": 150,
          "test_total": 256
        },
        "seed": 4
      },
      {
        "candidate_pair_count": 256,
        "derived": {
          "honest_audit_accuracy": 0.4951171875,
          "honest_selection_accuracy": 0.5859375,
          "honest_test_accuracy": 0.51953125,
          "leaky_audit_accuracy": 0.5,
          "leaky_reported_test_accuracy": 0.5859375,
          "leaky_vs_honest_test_gap": 0.06640625,
          "optimism_gap": 0.0859375
        },
        "honest": {
          "audit_correct": 507,
          "audit_total": 1024,
          "feature_index": 26,
          "polarity": 1,
          "selection_correct": 150,
          "selection_total": 256,
          "test_correct": 133,
          "test_total": 256
        },
        "k": 128,
        "leaky": {
          "audit_correct": 512,
          "audit_total": 1024,
          "feature_index": 11,
          "polarity": -1,
          "test_correct": 150,
          "test_total": 256
        },
        "seed": 4
      },
      {
        "candidate_pair_count": 512,
        "derived": {
          "honest_audit_accuracy": 0.47265625,
          "honest_selection_accuracy": 0.58984375,
          "honest_test_accuracy": 0.453125,
          "leaky_audit_accuracy": 0.5,
          "leaky_reported_test_accuracy": 0.5859375,
          "leaky_vs_honest_test_gap": 0.1328125,
          "optimism_gap": 0.0859375
        },
        "honest": {
          "audit_correct": 484,
          "audit_total": 1024,
          "feature_index": 161,
          "polarity": -1,
          "selection_correct": 151,
          "selection_total": 256,
          "test_correct": 116,
          "test_total": 256
        },
        "k": 256,
        "leaky": {
          "audit_correct": 512,
          "audit_total": 1024,
          "feature_index": 11,
          "polarity": -1,
          "test_correct": 150,
          "test_total": 256
        },
        "seed": 4
      },
      {
        "candidate_pair_count": 1024,
        "derived": {
          "honest_audit_accuracy": 0.4873046875,
          "honest_selection_accuracy": 0.59375,
          "honest_test_accuracy": 0.5078125,
          "leaky_audit_accuracy": 0.5,
          "leaky_reported_test_accuracy": 0.5859375,
          "leaky_vs_honest_test_gap": 0.078125,
          "optimism_gap": 0.0859375
        },
        "honest": {
          "audit_correct": 499,
          "audit_total": 1024,
          "feature_index": 323,
          "polarity": -1,
          "selection_correct": 152,
          "selection_total": 256,
          "test_correct": 130,
          "test_total": 256
        },
        "k": 512,
        "leaky": {
          "audit_correct": 512,
          "audit_total": 1024,
          "feature_index": 11,
          "polarity": -1,
          "test_correct": 150,
          "test_total": 256
        },
        "seed": 4
      },
      {
        "candidate_pair_count": 2048,
        "derived": {
          "honest_audit_accuracy": 0.490234375,
          "honest_selection_accuracy": 0.609375,
          "honest_test_accuracy": 0.49609375,
          "leaky_audit_accuracy": 0.4931640625,
          "leaky_reported_test_accuracy": 0.59765625,
          "leaky_vs_honest_test_gap": 0.1015625,
          "optimism_gap": 0.1044921875
        },
        "honest": {
          "audit_correct": 502,
          "audit_total": 1024,
          "feature_index": 963,
          "polarity": -1,
          "selection_correct": 156,
          "selection_total": 256,
          "test_correct": 127,
          "test_total": 256
        },
        "k": 1024,
        "leaky": {
          "audit_correct": 505,
          "audit_total": 1024,
          "feature_index": 790,
          "polarity": 1,
          "test_correct": 153,
          "test_total": 256
        },
        "seed": 4
      },
      {
        "candidate_pair_count": 2,
        "derived": {
          "honest_audit_accuracy": 0.515625,
          "honest_selection_accuracy": 0.51171875,
          "honest_test_accuracy": 0.47265625,
          "leaky_audit_accuracy": 0.484375,
          "leaky_reported_test_accuracy": 0.52734375,
          "leaky_vs_honest_test_gap": 0.0546875,
          "optimism_gap": 0.04296875
        },
        "honest": {
          "audit_correct": 528,
          "audit_total": 1024,
          "feature_index": 0,
          "polarity": -1,
          "selection_correct": 131,
          "selection_total": 256,
          "test_correct": 121,
          "test_total": 256
        },
        "k": 1,
        "leaky": {
          "audit_correct": 496,
          "audit_total": 1024,
          "feature_index": 0,
          "polarity": 1,
          "test_correct": 135,
          "test_total": 256
        },
        "seed": 5
      },
      {
        "candidate_pair_count": 4,
        "derived": {
          "honest_audit_accuracy": 0.5107421875,
          "honest_selection_accuracy": 0.55859375,
          "honest_test_accuracy": 0.578125,
          "leaky_audit_accuracy": 0.5107421875,
          "leaky_reported_test_accuracy": 0.578125,
          "leaky_vs_honest_test_gap": 0.0,
          "optimism_gap": 0.0673828125
        },
        "honest": {
          "audit_correct": 523,
          "audit_total": 1024,
          "feature_index": 1,
          "polarity": 1,
          "selection_correct": 143,
          "selection_total": 256,
          "test_correct": 148,
          "test_total": 256
        },
        "k": 2,
        "leaky": {
          "audit_correct": 523,
          "audit_total": 1024,
          "feature_index": 1,
          "polarity": 1,
          "test_correct": 148,
          "test_total": 256
        },
        "seed": 5
      },
      {
        "candidate_pair_count": 8,
        "derived": {
          "honest_audit_accuracy": 0.5107421875,
          "honest_selection_accuracy": 0.55859375,
          "honest_test_accuracy": 0.578125,
          "leaky_audit_accuracy": 0.5107421875,
          "leaky_reported_test_accuracy": 0.578125,
          "leaky_vs_honest_test_gap": 0.0,
          "optimism_gap": 0.0673828125
        },
        "honest": {
          "audit_correct": 523,
          "audit_total": 1024,
          "feature_index": 1,
          "polarity": 1,
          "selection_correct": 143,
          "selection_total": 256,
          "test_correct": 148,
          "test_total": 256
        },
        "k": 4,
        "leaky": {
          "audit_correct": 523,
          "audit_total": 1024,
          "feature_index": 1,
          "polarity": 1,
          "test_correct": 148,
          "test_total": 256
        },
        "seed": 5
      },
      {
        "candidate_pair_count": 16,
        "derived": {
          "honest_audit_accuracy": 0.5107421875,
          "honest_selection_accuracy": 0.55859375,
          "honest_test_accuracy": 0.578125,
          "leaky_audit_accuracy": 0.5107421875,
          "leaky_reported_test_accuracy": 0.578125,
          "leaky_vs_honest_test_gap": 0.0,
          "optimism_gap": 0.0673828125
        },
        "honest": {
          "audit_correct": 523,
          "audit_total": 1024,
          "feature_index": 1,
          "polarity": 1,
          "selection_correct": 143,
          "selection_total": 256,
          "test_correct": 148,
          "test_total": 256
        },
        "k": 8,
        "leaky": {
          "audit_correct": 523,
          "audit_total": 1024,
          "feature_index": 1,
          "polarity": 1,
          "test_correct": 148,
          "test_total": 256
        },
        "seed": 5
      },
      {
        "candidate_pair_count": 32,
        "derived": {
          "honest_audit_accuracy": 0.5107421875,
          "honest_selection_accuracy": 0.55859375,
          "honest_test_accuracy": 0.578125,
          "leaky_audit_accuracy": 0.5107421875,
          "leaky_reported_test_accuracy": 0.578125,
          "leaky_vs_honest_test_gap": 0.0,
          "optimism_gap": 0.0673828125
        },
        "honest": {
          "audit_correct": 523,
          "audit_total": 1024,
          "feature_index": 1,
          "polarity": 1,
          "selection_correct": 143,
          "selection_total": 256,
          "test_correct": 148,
          "test_total": 256
        },
        "k": 16,
        "leaky": {
          "audit_correct": 523,
          "audit_total": 1024,
          "feature_index": 1,
          "polarity": 1,
          "test_correct": 148,
          "test_total": 256
        },
        "seed": 5
      },
      {
        "candidate_pair_count": 64,
        "derived": {
          "honest_audit_accuracy": 0.498046875,
          "honest_selection_accuracy": 0.56640625,
          "honest_test_accuracy": 0.48828125,
          "leaky_audit_accuracy": 0.5107421875,
          "leaky_reported_test_accuracy": 0.578125,
          "leaky_vs_honest_test_gap": 0.08984375,
          "optimism_gap": 0.0673828125
        },
        "honest": {
          "audit_correct": 510,
          "audit_total": 1024,
          "feature_index": 31,
          "polarity": -1,
          "selection_correct": 145,
          "selection_total": 256,
          "test_correct": 125,
          "test_total": 256
        },
        "k": 32,
        "leaky": {
          "audit_correct": 523,
          "audit_total": 1024,
          "feature_index": 1,
          "polarity": 1,
          "test_correct": 148,
          "test_total": 256
        },
        "seed": 5
      },
      {
        "candidate_pair_count": 128,
        "derived": {
          "honest_audit_accuracy": 0.4931640625,
          "honest_selection_accuracy": 0.5703125,
          "honest_test_accuracy": 0.546875,
          "leaky_audit_accuracy": 0.5107421875,
          "leaky_reported_test_accuracy": 0.578125,
          "leaky_vs_honest_test_gap": 0.03125,
          "optimism_gap": 0.0673828125
        },
        "honest": {
          "audit_correct": 505,
          "audit_total": 1024,
          "feature_index": 33,
          "polarity": 1,
          "selection_correct": 146,
          "selection_total": 256,
          "test_correct": 140,
          "test_total": 256
        },
        "k": 64,
        "leaky": {
          "audit_correct": 523,
          "audit_total": 1024,
          "feature_index": 1,
          "polarity": 1,
          "test_correct": 148,
          "test_total": 256
        },
        "seed": 5
      },
      {
        "candidate_pair_count": 256,
        "derived": {
          "honest_audit_accuracy": 0.4931640625,
          "honest_selection_accuracy": 0.5703125,
          "honest_test_accuracy": 0.546875,
          "leaky_audit_accuracy": 0.5107421875,
          "leaky_reported_test_accuracy": 0.578125,
          "leaky_vs_honest_test_gap": 0.03125,
          "optimism_gap": 0.0673828125
        },
        "honest": {
          "audit_correct": 505,
          "audit_total": 1024,
          "feature_index": 33,
          "polarity": 1,
          "selection_correct": 146,
          "selection_total": 256,
          "test_correct": 140,
          "test_total": 256
        },
        "k": 128,
        "leaky": {
          "audit_correct": 523,
          "audit_total": 1024,
          "feature_index": 1,
          "polarity": 1,
          "test_correct": 148,
          "test_total": 256
        },
        "seed": 5
      },
      {
        "candidate_pair_count": 512,
        "derived": {
          "honest_audit_accuracy": 0.4775390625,
          "honest_selection_accuracy": 0.57421875,
          "honest_test_accuracy": 0.47265625,
          "leaky_audit_accuracy": 0.501953125,
          "leaky_reported_test_accuracy": 0.6015625,
          "leaky_vs_honest_test_gap": 0.12890625,
          "optimism_gap": 0.099609375
        },
        "honest": {
          "audit_correct": 489,
          "audit_total": 1024,
          "feature_index": 135,
          "polarity": -1,
          "selection_correct": 147,
          "selection_total": 256,
          "test_correct": 121,
          "test_total": 256
        },
        "k": 256,
        "leaky": {
          "audit_correct": 514,
          "audit_total": 1024,
          "feature_index": 154,
          "polarity": 1,
          "test_correct": 154,
          "test_total": 256
        },
        "seed": 5
      },
      {
        "candidate_pair_count": 1024,
        "derived": {
          "honest_audit_accuracy": 0.48828125,
          "honest_selection_accuracy": 0.59765625,
          "honest_test_accuracy": 0.515625,
          "leaky_audit_accuracy": 0.501953125,
          "leaky_reported_test_accuracy": 0.6015625,
          "leaky_vs_honest_test_gap": 0.0859375,
          "optimism_gap": 0.099609375
        },
        "honest": {
          "audit_correct": 500,
          "audit_total": 1024,
          "feature_index": 320,
          "polarity": -1,
          "selection_correct": 153,
          "selection_total": 256,
          "test_correct": 132,
          "test_total": 256
        },
        "k": 512,
        "leaky": {
          "audit_correct": 514,
          "audit_total": 1024,
          "feature_index": 154,
          "polarity": 1,
          "test_correct": 154,
          "test_total": 256
        },
        "seed": 5
      },
      {
        "candidate_pair_count": 2048,
        "derived": {
          "honest_audit_accuracy": 0.48828125,
          "honest_selection_accuracy": 0.59765625,
          "honest_test_accuracy": 0.515625,
          "leaky_audit_accuracy": 0.501953125,
          "leaky_reported_test_accuracy": 0.6015625,
          "leaky_vs_honest_test_gap": 0.0859375,
          "optimism_gap": 0.099609375
        },
        "honest": {
          "audit_correct": 500,
          "audit_total": 1024,
          "feature_index": 320,
          "polarity": -1,
          "selection_correct": 153,
          "selection_total": 256,
          "test_correct": 132,
          "test_total": 256
        },
        "k": 1024,
        "leaky": {
          "audit_correct": 514,
          "audit_total": 1024,
          "feature_index": 154,
          "polarity": 1,
          "test_correct": 154,
          "test_total": 256
        },
        "seed": 5
      },
      {
        "candidate_pair_count": 2,
        "derived": {
          "honest_audit_accuracy": 0.5,
          "honest_selection_accuracy": 0.55078125,
          "honest_test_accuracy": 0.48046875,
          "leaky_audit_accuracy": 0.5,
          "leaky_reported_test_accuracy": 0.51953125,
          "leaky_vs_honest_test_gap": 0.0390625,
          "optimism_gap": 0.01953125
        },
        "honest": {
          "audit_correct": 512,
          "audit_total": 1024,
          "feature_index": 0,
          "polarity": 1,
          "selection_correct": 141,
          "selection_total": 256,
          "test_correct": 123,
          "test_total": 256
        },
        "k": 1,
        "leaky": {
          "audit_correct": 512,
          "audit_total": 1024,
          "feature_index": 0,
          "polarity": -1,
          "test_correct": 133,
          "test_total": 256
        },
        "seed": 6
      },
      {
        "candidate_pair_count": 4,
        "derived": {
          "honest_audit_accuracy": 0.5,
          "honest_selection_accuracy": 0.55078125,
          "honest_test_accuracy": 0.48046875,
          "leaky_audit_accuracy": 0.5,
          "leaky_reported_test_accuracy": 0.51953125,
          "leaky_vs_honest_test_gap": 0.0390625,
          "optimism_gap": 0.01953125
        },
        "honest": {
          "audit_correct": 512,
          "audit_total": 1024,
          "feature_index": 0,
          "polarity": 1,
          "selection_correct": 141,
          "selection_total": 256,
          "test_correct": 123,
          "test_total": 256
        },
        "k": 2,
        "leaky": {
          "audit_correct": 512,
          "audit_total": 1024,
          "feature_index": 0,
          "polarity": -1,
          "test_correct": 133,
          "test_total": 256
        },
        "seed": 6
      },
      {
        "candidate_pair_count": 8,
        "derived": {
          "honest_audit_accuracy": 0.5,
          "honest_selection_accuracy": 0.55078125,
          "honest_test_accuracy": 0.48046875,
          "leaky_audit_accuracy": 0.5,
          "leaky_reported_test_accuracy": 0.51953125,
          "leaky_vs_honest_test_gap": 0.0390625,
          "optimism_gap": 0.01953125
        },
        "honest": {
          "audit_correct": 512,
          "audit_total": 1024,
          "feature_index": 0,
          "polarity": 1,
          "selection_correct": 141,
          "selection_total": 256,
          "test_correct": 123,
          "test_total": 256
        },
        "k": 4,
        "leaky": {
          "audit_correct": 512,
          "audit_total": 1024,
          "feature_index": 0,
          "polarity": -1,
          "test_correct": 133,
          "test_total": 256
        },
        "seed": 6
      },
      {
        "candidate_pair_count": 16,
        "derived": {
          "honest_audit_accuracy": 0.5,
          "honest_selection_accuracy": 0.55078125,
          "honest_test_accuracy": 0.48046875,
          "leaky_audit_accuracy": 0.505859375,
          "leaky_reported_test_accuracy": 0.578125,
          "leaky_vs_honest_test_gap": 0.09765625,
          "optimism_gap": 0.072265625
        },
        "honest": {
          "audit_correct": 512,
          "audit_total": 1024,
          "feature_index": 0,
          "polarity": 1,
          "selection_correct": 141,
          "selection_total": 256,
          "test_correct": 123,
          "test_total": 256
        },
        "k": 8,
        "leaky": {
          "audit_correct": 518,
          "audit_total": 1024,
          "feature_index": 4,
          "polarity": -1,
          "test_correct": 148,
          "test_total": 256
        },
        "seed": 6
      },
      {
        "candidate_pair_count": 32,
        "derived": {
          "honest_audit_accuracy": 0.5,
          "honest_selection_accuracy": 0.55078125,
          "honest_test_accuracy": 0.48046875,
          "leaky_audit_accuracy": 0.505859375,
          "leaky_reported_test_accuracy": 0.578125,
          "leaky_vs_honest_test_gap": 0.09765625,
          "optimism_gap": 0.072265625
        },
        "honest": {
          "audit_correct": 512,
          "audit_total": 1024,
          "feature_index": 0,
          "polarity": 1,
          "selection_correct": 141,
          "selection_total": 256,
          "test_correct": 123,
          "test_total": 256
        },
        "k": 16,
        "leaky": {
          "audit_correct": 518,
          "audit_total": 1024,
          "feature_index": 4,
          "polarity": -1,
          "test_correct": 148,
          "test_total": 256
        },
        "seed": 6
      },
      {
        "candidate_pair_count": 64,
        "derived": {
          "honest_audit_accuracy": 0.48828125,
          "honest_selection_accuracy": 0.56640625,
          "honest_test_accuracy": 0.515625,
          "leaky_audit_accuracy": 0.505859375,
          "leaky_reported_test_accuracy": 0.578125,
          "leaky_vs_honest_test_gap": 0.0625,
          "optimism_gap": 0.072265625
        },
        "honest": {
          "audit_correct": 500,
          "audit_total": 1024,
          "feature_index": 19,
          "polarity": 1,
          "selection_correct": 145,
          "selection_total": 256,
          "test_correct": 132,
          "test_total": 256
        },
        "k": 32,
        "leaky": {
          "audit_correct": 518,
          "audit_total": 1024,
          "feature_index": 4,
          "polarity": -1,
          "test_correct": 148,
          "test_total": 256
        },
        "seed": 6
      },
      {
        "candidate_pair_count": 128,
        "derived": {
          "honest_audit_accuracy": 0.4990234375,
          "honest_selection_accuracy": 0.578125,
          "honest_test_accuracy": 0.4921875,
          "leaky_audit_accuracy": 0.505859375,
          "leaky_reported_test_accuracy": 0.578125,
          "leaky_vs_honest_test_gap": 0.0859375,
          "optimism_gap": 0.072265625
        },
        "honest": {
          "audit_correct": 511,
          "audit_total": 1024,
          "feature_index": 62,
          "polarity": -1,
          "selection_correct": 148,
          "selection_total": 256,
          "test_correct": 126,
          "test_total": 256
        },
        "k": 64,
        "leaky": {
          "audit_correct": 518,
          "audit_total": 1024,
          "feature_index": 4,
          "polarity": -1,
          "test_correct": 148,
          "test_total": 256
        },
        "seed": 6
      },
      {
        "candidate_pair_count": 256,
        "derived": {
          "honest_audit_accuracy": 0.5048828125,
          "honest_selection_accuracy": 0.609375,
          "honest_test_accuracy": 0.51953125,
          "leaky_audit_accuracy": 0.505859375,
          "leaky_reported_test_accuracy": 0.578125,
          "leaky_vs_honest_test_gap": 0.05859375,
          "optimism_gap": 0.072265625
        },
        "honest": {
          "audit_correct": 517,
          "audit_total": 1024,
          "feature_index": 102,
          "polarity": -1,
          "selection_correct": 156,
          "selection_total": 256,
          "test_correct": 133,
          "test_total": 256
        },
        "k": 128,
        "leaky": {
          "audit_correct": 518,
          "audit_total": 1024,
          "feature_index": 4,
          "polarity": -1,
          "test_correct": 148,
          "test_total": 256
        },
        "seed": 6
      },
      {
        "candidate_pair_count": 512,
        "derived": {
          "honest_audit_accuracy": 0.5048828125,
          "honest_selection_accuracy": 0.609375,
          "honest_test_accuracy": 0.51953125,
          "leaky_audit_accuracy": 0.478515625,
          "leaky_reported_test_accuracy": 0.6015625,
          "leaky_vs_honest_test_gap": 0.08203125,
          "optimism_gap": 0.123046875
        },
        "honest": {
          "audit_correct": 517,
          "audit_total": 1024,
          "feature_index": 102,
          "polarity": -1,
          "selection_correct": 156,
          "selection_total": 256,
          "test_correct": 133,
          "test_total": 256
        },
        "k": 256,
        "leaky": {
          "audit_correct": 490,
          "audit_total": 1024,
          "feature_index": 151,
          "polarity": 1,
          "test_correct": 154,
          "test_total": 256
        },
        "seed": 6
      },
      {
        "candidate_pair_count": 1024,
        "derived": {
          "honest_audit_accuracy": 0.5048828125,
          "honest_selection_accuracy": 0.609375,
          "honest_test_accuracy": 0.51953125,
          "leaky_audit_accuracy": 0.478515625,
          "leaky_reported_test_accuracy": 0.6015625,
          "leaky_vs_honest_test_gap": 0.08203125,
          "optimism_gap": 0.123046875
        },
        "honest": {
          "audit_correct": 517,
          "audit_total": 1024,
          "feature_index": 102,
          "polarity": -1,
          "selection_correct": 156,
          "selection_total": 256,
          "test_correct": 133,
          "test_total": 256
        },
        "k": 512,
        "leaky": {
          "audit_correct": 490,
          "audit_total": 1024,
          "feature_index": 151,
          "polarity": 1,
          "test_correct": 154,
          "test_total": 256
        },
        "seed": 6
      },
      {
        "candidate_pair_count": 2048,
        "derived": {
          "honest_audit_accuracy": 0.5048828125,
          "honest_selection_accuracy": 0.609375,
          "honest_test_accuracy": 0.51953125,
          "leaky_audit_accuracy": 0.478515625,
          "leaky_reported_test_accuracy": 0.6015625,
          "leaky_vs_honest_test_gap": 0.08203125,
          "optimism_gap": 0.123046875
        },
        "honest": {
          "audit_correct": 517,
          "audit_total": 1024,
          "feature_index": 102,
          "polarity": -1,
          "selection_correct": 156,
          "selection_total": 256,
          "test_correct": 133,
          "test_total": 256
        },
        "k": 1024,
        "leaky": {
          "audit_correct": 490,
          "audit_total": 1024,
          "feature_index": 151,
          "polarity": 1,
          "test_correct": 154,
          "test_total": 256
        },
        "seed": 6
      },
      {
        "candidate_pair_count": 2,
        "derived": {
          "honest_audit_accuracy": 0.5263671875,
          "honest_selection_accuracy": 0.50390625,
          "honest_test_accuracy": 0.4765625,
          "leaky_audit_accuracy": 0.4736328125,
          "leaky_reported_test_accuracy": 0.5234375,
          "leaky_vs_honest_test_gap": 0.046875,
          "optimism_gap": 0.0498046875
        },
        "honest": {
          "audit_correct": 539,
          "audit_total": 1024,
          "feature_index": 0,
          "polarity": 1,
          "selection_correct": 129,
          "selection_total": 256,
          "test_correct": 122,
          "test_total": 256
        },
        "k": 1,
        "leaky": {
          "audit_correct": 485,
          "audit_total": 1024,
          "feature_index": 0,
          "polarity": -1,
          "test_correct": 134,
          "test_total": 256
        },
        "seed": 7
      },
      {
        "candidate_pair_count": 4,
        "derived": {
          "honest_audit_accuracy": 0.5166015625,
          "honest_selection_accuracy": 0.55078125,
          "honest_test_accuracy": 0.51171875,
          "leaky_audit_accuracy": 0.4736328125,
          "leaky_reported_test_accuracy": 0.5234375,
          "leaky_vs_honest_test_gap": 0.01171875,
          "optimism_gap": 0.0498046875
        },
        "honest": {
          "audit_correct": 529,
          "audit_total": 1024,
          "feature_index": 1,
          "polarity": 1,
          "selection_correct": 141,
          "selection_total": 256,
          "test_correct": 131,
          "test_total": 256
        },
        "k": 2,
        "leaky": {
          "audit_correct": 485,
          "audit_total": 1024,
          "feature_index": 0,
          "polarity": -1,
          "test_correct": 134,
          "test_total": 256
        },
        "seed": 7
      },
      {
        "candidate_pair_count": 8,
        "derived": {
          "honest_audit_accuracy": 0.5166015625,
          "honest_selection_accuracy": 0.55078125,
          "honest_test_accuracy": 0.51171875,
          "leaky_audit_accuracy": 0.5234375,
          "leaky_reported_test_accuracy": 0.5546875,
          "leaky_vs_honest_test_gap": 0.04296875,
          "optimism_gap": 0.03125
        },
        "honest": {
          "audit_correct": 529,
          "audit_total": 1024,
          "feature_index": 1,
          "polarity": 1,
          "selection_correct": 141,
          "selection_total": 256,
          "test_correct": 131,
          "test_total": 256
        },
        "k": 4,
        "leaky": {
          "audit_correct": 536,
          "audit_total": 1024,
          "feature_index": 3,
          "polarity": 1,
          "test_correct": 142,
          "test_total": 256
        },
        "seed": 7
      },
      {
        "candidate_pair_count": 16,
        "derived": {
          "honest_audit_accuracy": 0.5166015625,
          "honest_selection_accuracy": 0.55078125,
          "honest_test_accuracy": 0.51171875,
          "leaky_audit_accuracy": 0.5234375,
          "leaky_reported_test_accuracy": 0.5546875,
          "leaky_vs_honest_test_gap": 0.04296875,
          "optimism_gap": 0.03125
        },
        "honest": {
          "audit_correct": 529,
          "audit_total": 1024,
          "feature_index": 1,
          "polarity": 1,
          "selection_correct": 141,
          "selection_total": 256,
          "test_correct": 131,
          "test_total": 256
        },
        "k": 8,
        "leaky": {
          "audit_correct": 536,
          "audit_total": 1024,
          "feature_index": 3,
          "polarity": 1,
          "test_correct": 142,
          "test_total": 256
        },
        "seed": 7
      },
      {
        "candidate_pair_count": 32,
        "derived": {
          "honest_audit_accuracy": 0.4990234375,
          "honest_selection_accuracy": 0.578125,
          "honest_test_accuracy": 0.51171875,
          "leaky_audit_accuracy": 0.5234375,
          "leaky_reported_test_accuracy": 0.5546875,
          "leaky_vs_honest_test_gap": 0.04296875,
          "optimism_gap": 0.03125
        },
        "honest": {
          "audit_correct": 511,
          "audit_total": 1024,
          "feature_index": 12,
          "polarity": -1,
          "selection_correct": 148,
          "selection_total": 256,
          "test_correct": 131,
          "test_total": 256
        },
        "k": 16,
        "leaky": {
          "audit_correct": 536,
          "audit_total": 1024,
          "feature_index": 3,
          "polarity": 1,
          "test_correct": 142,
          "test_total": 256
        },
        "seed": 7
      },
      {
        "candidate_pair_count": 64,
        "derived": {
          "honest_audit_accuracy": 0.498046875,
          "honest_selection_accuracy": 0.6171875,
          "honest_test_accuracy": 0.52734375,
          "leaky_audit_accuracy": 0.5234375,
          "leaky_reported_test_accuracy": 0.5546875,
          "leaky_vs_honest_test_gap": 0.02734375,
          "optimism_gap": 0.03125
        },
        "honest": {
          "audit_correct": 510,
          "audit_total": 1024,
          "feature_index": 25,
          "polarity": -1,
          "selection_correct": 158,
          "selection_total": 256,
          "test_correct": 135,
          "test_total": 256
        },
        "k": 32,
        "leaky": {
          "audit_correct": 536,
          "audit_total": 1024,
          "feature_index": 3,
          "polarity": 1,
          "test_correct": 142,
          "test_total": 256
        },
        "seed": 7
      },
      {
        "candidate_pair_count": 128,
        "derived": {
          "honest_audit_accuracy": 0.498046875,
          "honest_selection_accuracy": 0.6171875,
          "honest_test_accuracy": 0.52734375,
          "leaky_audit_accuracy": 0.4912109375,
          "leaky_reported_test_accuracy": 0.5859375,
          "leaky_vs_honest_test_gap": 0.05859375,
          "optimism_gap": 0.0947265625
        },
        "honest": {
          "audit_correct": 510,
          "audit_total": 1024,
          "feature_index": 25,
          "polarity": -1,
          "selection_correct": 158,
          "selection_total": 256,
          "test_correct": 135,
          "test_total": 256
        },
        "k": 64,
        "leaky": {
          "audit_correct": 503,
          "audit_total": 1024,
          "feature_index": 58,
          "polarity": 1,
          "test_correct": 150,
          "test_total": 256
        },
        "seed": 7
      },
      {
        "candidate_pair_count": 256,
        "derived": {
          "honest_audit_accuracy": 0.498046875,
          "honest_selection_accuracy": 0.6171875,
          "honest_test_accuracy": 0.52734375,
          "leaky_audit_accuracy": 0.4912109375,
          "leaky_reported_test_accuracy": 0.5859375,
          "leaky_vs_honest_test_gap": 0.05859375,
          "optimism_gap": 0.0947265625
        },
        "honest": {
          "audit_correct": 510,
          "audit_total": 1024,
          "feature_index": 25,
          "polarity": -1,
          "selection_correct": 158,
          "selection_total": 256,
          "test_correct": 135,
          "test_total": 256
        },
        "k": 128,
        "leaky": {
          "audit_correct": 503,
          "audit_total": 1024,
          "feature_index": 58,
          "polarity": 1,
          "test_correct": 150,
          "test_total": 256
        },
        "seed": 7
      },
      {
        "candidate_pair_count": 512,
        "derived": {
          "honest_audit_accuracy": 0.498046875,
          "honest_selection_accuracy": 0.6171875,
          "honest_test_accuracy": 0.52734375,
          "leaky_audit_accuracy": 0.5087890625,
          "leaky_reported_test_accuracy": 0.60546875,
          "leaky_vs_honest_test_gap": 0.078125,
          "optimism_gap": 0.0966796875
        },
        "honest": {
          "audit_correct": 510,
          "audit_total": 1024,
          "feature_index": 25,
          "polarity": -1,
          "selection_correct": 158,
          "selection_total": 256,
          "test_correct": 135,
          "test_total": 256
        },
        "k": 256,
        "leaky": {
          "audit_correct": 521,
          "audit_total": 1024,
          "feature_index": 239,
          "polarity": 1,
          "test_correct": 155,
          "test_total": 256
        },
        "seed": 7
      },
      {
        "candidate_pair_count": 1024,
        "derived": {
          "honest_audit_accuracy": 0.498046875,
          "honest_selection_accuracy": 0.6171875,
          "honest_test_accuracy": 0.52734375,
          "leaky_audit_accuracy": 0.5087890625,
          "leaky_reported_test_accuracy": 0.60546875,
          "leaky_vs_honest_test_gap": 0.078125,
          "optimism_gap": 0.0966796875
        },
        "honest": {
          "audit_correct": 510,
          "audit_total": 1024,
          "feature_index": 25,
          "polarity": -1,
          "selection_correct": 158,
          "selection_total": 256,
          "test_correct": 135,
          "test_total": 256
        },
        "k": 512,
        "leaky": {
          "audit_correct": 521,
          "audit_total": 1024,
          "feature_index": 239,
          "polarity": 1,
          "test_correct": 155,
          "test_total": 256
        },
        "seed": 7
      },
      {
        "candidate_pair_count": 2048,
        "derived": {
          "honest_audit_accuracy": 0.498046875,
          "honest_selection_accuracy": 0.6171875,
          "honest_test_accuracy": 0.52734375,
          "leaky_audit_accuracy": 0.5087890625,
          "leaky_reported_test_accuracy": 0.60546875,
          "leaky_vs_honest_test_gap": 0.078125,
          "optimism_gap": 0.0966796875
        },
        "honest": {
          "audit_correct": 510,
          "audit_total": 1024,
          "feature_index": 25,
          "polarity": -1,
          "selection_correct": 158,
          "selection_total": 256,
          "test_correct": 135,
          "test_total": 256
        },
        "k": 1024,
        "leaky": {
          "audit_correct": 521,
          "audit_total": 1024,
          "feature_index": 239,
          "polarity": 1,
          "test_correct": 155,
          "test_total": 256
        },
        "seed": 7
      },
      {
        "candidate_pair_count": 2,
        "derived": {
          "honest_audit_accuracy": 0.509765625,
          "honest_selection_accuracy": 0.52734375,
          "honest_test_accuracy": 0.45703125,
          "leaky_audit_accuracy": 0.490234375,
          "leaky_reported_test_accuracy": 0.54296875,
          "leaky_vs_honest_test_gap": 0.0859375,
          "optimism_gap": 0.052734375
        },
        "honest": {
          "audit_correct": 522,
          "audit_total": 1024,
          "feature_index": 0,
          "polarity": 1,
          "selection_correct": 135,
          "selection_total": 256,
          "test_correct": 117,
          "test_total": 256
        },
        "k": 1,
        "leaky": {
          "audit_correct": 502,
          "audit_total": 1024,
          "feature_index": 0,
          "polarity": -1,
          "test_correct": 139,
          "test_total": 256
        },
        "seed": 8
      },
      {
        "candidate_pair_count": 4,
        "derived": {
          "honest_audit_accuracy": 0.4814453125,
          "honest_selection_accuracy": 0.53125,
          "honest_test_accuracy": 0.546875,
          "leaky_audit_accuracy": 0.4814453125,
          "leaky_reported_test_accuracy": 0.546875,
          "leaky_vs_honest_test_gap": 0.0,
          "optimism_gap": 0.0654296875
        },
        "honest": {
          "audit_correct": 493,
          "audit_total": 1024,
          "feature_index": 1,
          "polarity": 1,
          "selection_correct": 136,
          "selection_total": 256,
          "test_correct": 140,
          "test_total": 256
        },
        "k": 2,
        "leaky": {
          "audit_correct": 493,
          "audit_total": 1024,
          "feature_index": 1,
          "polarity": 1,
          "test_correct": 140,
          "test_total": 256
        },
        "seed": 8
      },
      {
        "candidate_pair_count": 8,
        "derived": {
          "honest_audit_accuracy": 0.4814453125,
          "honest_selection_accuracy": 0.53125,
          "honest_test_accuracy": 0.546875,
          "leaky_audit_accuracy": 0.4814453125,
          "leaky_reported_test_accuracy": 0.546875,
          "leaky_vs_honest_test_gap": 0.0,
          "optimism_gap": 0.0654296875
        },
        "honest": {
          "audit_correct": 493,
          "audit_total": 1024,
          "feature_index": 1,
          "polarity": 1,
          "selection_correct": 136,
          "selection_total": 256,
          "test_correct": 140,
          "test_total": 256
        },
        "k": 4,
        "leaky": {
          "audit_correct": 493,
          "audit_total": 1024,
          "feature_index": 1,
          "polarity": 1,
          "test_correct": 140,
          "test_total": 256
        },
        "seed": 8
      },
      {
        "candidate_pair_count": 16,
        "derived": {
          "honest_audit_accuracy": 0.4814453125,
          "honest_selection_accuracy": 0.53125,
          "honest_test_accuracy": 0.546875,
          "leaky_audit_accuracy": 0.4814453125,
          "leaky_reported_test_accuracy": 0.546875,
          "leaky_vs_honest_test_gap": 0.0,
          "optimism_gap": 0.0654296875
        },
        "honest": {
          "audit_correct": 493,
          "audit_total": 1024,
          "feature_index": 1,
          "polarity": 1,
          "selection_correct": 136,
          "selection_total": 256,
          "test_correct": 140,
          "test_total": 256
        },
        "k": 8,
        "leaky": {
          "audit_correct": 493,
          "audit_total": 1024,
          "feature_index": 1,
          "polarity": 1,
          "test_correct": 140,
          "test_total": 256
        },
        "seed": 8
      },
      {
        "candidate_pair_count": 32,
        "derived": {
          "honest_audit_accuracy": 0.5048828125,
          "honest_selection_accuracy": 0.5390625,
          "honest_test_accuracy": 0.53125,
          "leaky_audit_accuracy": 0.5,
          "leaky_reported_test_accuracy": 0.6171875,
          "leaky_vs_honest_test_gap": 0.0859375,
          "optimism_gap": 0.1171875
        },
        "honest": {
          "audit_correct": 517,
          "audit_total": 1024,
          "feature_index": 9,
          "polarity": -1,
          "selection_correct": 138,
          "selection_total": 256,
          "test_correct": 136,
          "test_total": 256
        },
        "k": 16,
        "leaky": {
          "audit_correct": 512,
          "audit_total": 1024,
          "feature_index": 12,
          "polarity": -1,
          "test_correct": 158,
          "test_total": 256
        },
        "seed": 8
      },
      {
        "candidate_pair_count": 64,
        "derived": {
          "honest_audit_accuracy": 0.498046875,
          "honest_selection_accuracy": 0.57421875,
          "honest_test_accuracy": 0.5234375,
          "leaky_audit_accuracy": 0.5,
          "leaky_reported_test_accuracy": 0.6171875,
          "leaky_vs_honest_test_gap": 0.09375,
          "optimism_gap": 0.1171875
        },
        "honest": {
          "audit_correct": 510,
          "audit_total": 1024,
          "feature_index": 29,
          "polarity": -1,
          "selection_correct": 147,
          "selection_total": 256,
          "test_correct": 134,
          "test_total": 256
        },
        "k": 32,
        "leaky": {
          "audit_correct": 512,
          "audit_total": 1024,
          "feature_index": 12,
          "polarity": -1,
          "test_correct": 158,
          "test_total": 256
        },
        "seed": 8
      },
      {
        "candidate_pair_count": 128,
        "derived": {
          "honest_audit_accuracy": 0.521484375,
          "honest_selection_accuracy": 0.58984375,
          "honest_test_accuracy": 0.51953125,
          "leaky_audit_accuracy": 0.5,
          "leaky_reported_test_accuracy": 0.6171875,
          "leaky_vs_honest_test_gap": 0.09765625,
          "optimism_gap": 0.1171875
        },
        "honest": {
          "audit_correct": 534,
          "audit_total": 1024,
          "feature_index": 48,
          "polarity": -1,
          "selection_correct": 151,
          "selection_total": 256,
          "test_correct": 133,
          "test_total": 256
        },
        "k": 64,
        "leaky": {
          "audit_correct": 512,
          "audit_total": 1024,
          "feature_index": 12,
          "polarity": -1,
          "test_correct": 158,
          "test_total": 256
        },
        "seed": 8
      },
      {
        "candidate_pair_count": 256,
        "derived": {
          "honest_audit_accuracy": 0.521484375,
          "honest_selection_accuracy": 0.58984375,
          "honest_test_accuracy": 0.51953125,
          "leaky_audit_accuracy": 0.5,
          "leaky_reported_test_accuracy": 0.6171875,
          "leaky_vs_honest_test_gap": 0.09765625,
          "optimism_gap": 0.1171875
        },
        "honest": {
          "audit_correct": 534,
          "audit_total": 1024,
          "feature_index": 48,
          "polarity": -1,
          "selection_correct": 151,
          "selection_total": 256,
          "test_correct": 133,
          "test_total": 256
        },
        "k": 128,
        "leaky": {
          "audit_correct": 512,
          "audit_total": 1024,
          "feature_index": 12,
          "polarity": -1,
          "test_correct": 158,
          "test_total": 256
        },
        "seed": 8
      },
      {
        "candidate_pair_count": 512,
        "derived": {
          "honest_audit_accuracy": 0.521484375,
          "honest_selection_accuracy": 0.58984375,
          "honest_test_accuracy": 0.51953125,
          "leaky_audit_accuracy": 0.5,
          "leaky_reported_test_accuracy": 0.6171875,
          "leaky_vs_honest_test_gap": 0.09765625,
          "optimism_gap": 0.1171875
        },
        "honest": {
          "audit_correct": 534,
          "audit_total": 1024,
          "feature_index": 48,
          "polarity": -1,
          "selection_correct": 151,
          "selection_total": 256,
          "test_correct": 133,
          "test_total": 256
        },
        "k": 256,
        "leaky": {
          "audit_correct": 512,
          "audit_total": 1024,
          "feature_index": 12,
          "polarity": -1,
          "test_correct": 158,
          "test_total": 256
        },
        "seed": 8
      },
      {
        "candidate_pair_count": 1024,
        "derived": {
          "honest_audit_accuracy": 0.52734375,
          "honest_selection_accuracy": 0.61328125,
          "honest_test_accuracy": 0.55078125,
          "leaky_audit_accuracy": 0.5,
          "leaky_reported_test_accuracy": 0.6171875,
          "leaky_vs_honest_test_gap": 0.06640625,
          "optimism_gap": 0.1171875
        },
        "honest": {
          "audit_correct": 540,
          "audit_total": 1024,
          "feature_index": 320,
          "polarity": -1,
          "selection_correct": 157,
          "selection_total": 256,
          "test_correct": 141,
          "test_total": 256
        },
        "k": 512,
        "leaky": {
          "audit_correct": 512,
          "audit_total": 1024,
          "feature_index": 12,
          "polarity": -1,
          "test_correct": 158,
          "test_total": 256
        },
        "seed": 8
      },
      {
        "candidate_pair_count": 2048,
        "derived": {
          "honest_audit_accuracy": 0.52734375,
          "honest_selection_accuracy": 0.61328125,
          "honest_test_accuracy": 0.55078125,
          "leaky_audit_accuracy": 0.5,
          "leaky_reported_test_accuracy": 0.6171875,
          "leaky_vs_honest_test_gap": 0.06640625,
          "optimism_gap": 0.1171875
        },
        "honest": {
          "audit_correct": 540,
          "audit_total": 1024,
          "feature_index": 320,
          "polarity": -1,
          "selection_correct": 157,
          "selection_total": 256,
          "test_correct": 141,
          "test_total": 256
        },
        "k": 1024,
        "leaky": {
          "audit_correct": 512,
          "audit_total": 1024,
          "feature_index": 12,
          "polarity": -1,
          "test_correct": 158,
          "test_total": 256
        },
        "seed": 8
      },
      {
        "candidate_pair_count": 2,
        "derived": {
          "honest_audit_accuracy": 0.48046875,
          "honest_selection_accuracy": 0.515625,
          "honest_test_accuracy": 0.5,
          "leaky_audit_accuracy": 0.48046875,
          "leaky_reported_test_accuracy": 0.5,
          "leaky_vs_honest_test_gap": 0.0,
          "optimism_gap": 0.01953125
        },
        "honest": {
          "audit_correct": 492,
          "audit_total": 1024,
          "feature_index": 0,
          "polarity": 1,
          "selection_correct": 132,
          "selection_total": 256,
          "test_correct": 128,
          "test_total": 256
        },
        "k": 1,
        "leaky": {
          "audit_correct": 492,
          "audit_total": 1024,
          "feature_index": 0,
          "polarity": 1,
          "test_correct": 128,
          "test_total": 256
        },
        "seed": 9
      },
      {
        "candidate_pair_count": 4,
        "derived": {
          "honest_audit_accuracy": 0.48828125,
          "honest_selection_accuracy": 0.51953125,
          "honest_test_accuracy": 0.4765625,
          "leaky_audit_accuracy": 0.51171875,
          "leaky_reported_test_accuracy": 0.5234375,
          "leaky_vs_honest_test_gap": 0.046875,
          "optimism_gap": 0.01171875
        },
        "honest": {
          "audit_correct": 500,
          "audit_total": 1024,
          "feature_index": 1,
          "polarity": -1,
          "selection_correct": 133,
          "selection_total": 256,
          "test_correct": 122,
          "test_total": 256
        },
        "k": 2,
        "leaky": {
          "audit_correct": 524,
          "audit_total": 1024,
          "feature_index": 1,
          "polarity": 1,
          "test_correct": 134,
          "test_total": 256
        },
        "seed": 9
      },
      {
        "candidate_pair_count": 8,
        "derived": {
          "honest_audit_accuracy": 0.49609375,
          "honest_selection_accuracy": 0.546875,
          "honest_test_accuracy": 0.5,
          "leaky_audit_accuracy": 0.51171875,
          "leaky_reported_test_accuracy": 0.5234375,
          "leaky_vs_honest_test_gap": 0.0234375,
          "optimism_gap": 0.01171875
        },
        "honest": {
          "audit_correct": 508,
          "audit_total": 1024,
          "feature_index": 3,
          "polarity": -1,
          "selection_correct": 140,
          "selection_total": 256,
          "test_correct": 128,
          "test_total": 256
        },
        "k": 4,
        "leaky": {
          "audit_correct": 524,
          "audit_total": 1024,
          "feature_index": 1,
          "polarity": 1,
          "test_correct": 134,
          "test_total": 256
        },
        "seed": 9
      },
      {
        "candidate_pair_count": 16,
        "derived": {
          "honest_audit_accuracy": 0.49609375,
          "honest_selection_accuracy": 0.546875,
          "honest_test_accuracy": 0.5,
          "leaky_audit_accuracy": 0.509765625,
          "leaky_reported_test_accuracy": 0.54296875,
          "leaky_vs_honest_test_gap": 0.04296875,
          "optimism_gap": 0.033203125
        },
        "honest": {
          "audit_correct": 508,
          "audit_total": 1024,
          "feature_index": 3,
          "polarity": -1,
          "selection_correct": 140,
          "selection_total": 256,
          "test_correct": 128,
          "test_total": 256
        },
        "k": 8,
        "leaky": {
          "audit_correct": 522,
          "audit_total": 1024,
          "feature_index": 5,
          "polarity": 1,
          "test_correct": 139,
          "test_total": 256
        },
        "seed": 9
      },
      {
        "candidate_pair_count": 32,
        "derived": {
          "honest_audit_accuracy": 0.498046875,
          "honest_selection_accuracy": 0.5625,
          "honest_test_accuracy": 0.4765625,
          "leaky_audit_accuracy": 0.490234375,
          "leaky_reported_test_accuracy": 0.5625,
          "leaky_vs_honest_test_gap": 0.0859375,
          "optimism_gap": 0.072265625
        },
        "honest": {
          "audit_correct": 510,
          "audit_total": 1024,
          "feature_index": 11,
          "polarity": -1,
          "selection_correct": 144,
          "selection_total": 256,
          "test_correct": 122,
          "test_total": 256
        },
        "k": 16,
        "leaky": {
          "audit_correct": 502,
          "audit_total": 1024,
          "feature_index": 12,
          "polarity": -1,
          "test_correct": 144,
          "test_total": 256
        },
        "seed": 9
      },
      {
        "candidate_pair_count": 64,
        "derived": {
          "honest_audit_accuracy": 0.4677734375,
          "honest_selection_accuracy": 0.57421875,
          "honest_test_accuracy": 0.51171875,
          "leaky_audit_accuracy": 0.4931640625,
          "leaky_reported_test_accuracy": 0.56640625,
          "leaky_vs_honest_test_gap": 0.0546875,
          "optimism_gap": 0.0732421875
        },
        "honest": {
          "audit_correct": 479,
          "audit_total": 1024,
          "feature_index": 17,
          "polarity": 1,
          "selection_correct": 147,
          "selection_total": 256,
          "test_correct": 131,
          "test_total": 256
        },
        "k": 32,
        "leaky": {
          "audit_correct": 505,
          "audit_total": 1024,
          "feature_index": 31,
          "polarity": -1,
          "test_correct": 145,
          "test_total": 256
        },
        "seed": 9
      },
      {
        "candidate_pair_count": 128,
        "derived": {
          "honest_audit_accuracy": 0.4677734375,
          "honest_selection_accuracy": 0.57421875,
          "honest_test_accuracy": 0.51171875,
          "leaky_audit_accuracy": 0.5166015625,
          "leaky_reported_test_accuracy": 0.58203125,
          "leaky_vs_honest_test_gap": 0.0703125,
          "optimism_gap": 0.0654296875
        },
        "honest": {
          "audit_correct": 479,
          "audit_total": 1024,
          "feature_index": 17,
          "polarity": 1,
          "selection_correct": 147,
          "selection_total": 256,
          "test_correct": 131,
          "test_total": 256
        },
        "k": 64,
        "leaky": {
          "audit_correct": 529,
          "audit_total": 1024,
          "feature_index": 63,
          "polarity": -1,
          "test_correct": 149,
          "test_total": 256
        },
        "seed": 9
      },
      {
        "candidate_pair_count": 256,
        "derived": {
          "honest_audit_accuracy": 0.5263671875,
          "honest_selection_accuracy": 0.61328125,
          "honest_test_accuracy": 0.51171875,
          "leaky_audit_accuracy": 0.5166015625,
          "leaky_reported_test_accuracy": 0.58203125,
          "leaky_vs_honest_test_gap": 0.0703125,
          "optimism_gap": 0.0654296875
        },
        "honest": {
          "audit_correct": 539,
          "audit_total": 1024,
          "feature_index": 84,
          "polarity": 1,
          "selection_correct": 157,
          "selection_total": 256,
          "test_correct": 131,
          "test_total": 256
        },
        "k": 128,
        "leaky": {
          "audit_correct": 529,
          "audit_total": 1024,
          "feature_index": 63,
          "polarity": -1,
          "test_correct": 149,
          "test_total": 256
        },
        "seed": 9
      },
      {
        "candidate_pair_count": 512,
        "derived": {
          "honest_audit_accuracy": 0.5263671875,
          "honest_selection_accuracy": 0.61328125,
          "honest_test_accuracy": 0.51171875,
          "leaky_audit_accuracy": 0.5146484375,
          "leaky_reported_test_accuracy": 0.609375,
          "leaky_vs_honest_test_gap": 0.09765625,
          "optimism_gap": 0.0947265625
        },
        "honest": {
          "audit_correct": 539,
          "audit_total": 1024,
          "feature_index": 84,
          "polarity": 1,
          "selection_correct": 157,
          "selection_total": 256,
          "test_correct": 131,
          "test_total": 256
        },
        "k": 256,
        "leaky": {
          "audit_correct": 527,
          "audit_total": 1024,
          "feature_index": 129,
          "polarity": 1,
          "test_correct": 156,
          "test_total": 256
        },
        "seed": 9
      },
      {
        "candidate_pair_count": 1024,
        "derived": {
          "honest_audit_accuracy": 0.5263671875,
          "honest_selection_accuracy": 0.61328125,
          "honest_test_accuracy": 0.51171875,
          "leaky_audit_accuracy": 0.5146484375,
          "leaky_reported_test_accuracy": 0.609375,
          "leaky_vs_honest_test_gap": 0.09765625,
          "optimism_gap": 0.0947265625
        },
        "honest": {
          "audit_correct": 539,
          "audit_total": 1024,
          "feature_index": 84,
          "polarity": 1,
          "selection_correct": 157,
          "selection_total": 256,
          "test_correct": 131,
          "test_total": 256
        },
        "k": 512,
        "leaky": {
          "audit_correct": 527,
          "audit_total": 1024,
          "feature_index": 129,
          "polarity": 1,
          "test_correct": 156,
          "test_total": 256
        },
        "seed": 9
      },
      {
        "candidate_pair_count": 2048,
        "derived": {
          "honest_audit_accuracy": 0.5263671875,
          "honest_selection_accuracy": 0.61328125,
          "honest_test_accuracy": 0.51171875,
          "leaky_audit_accuracy": 0.5146484375,
          "leaky_reported_test_accuracy": 0.609375,
          "leaky_vs_honest_test_gap": 0.09765625,
          "optimism_gap": 0.0947265625
        },
        "honest": {
          "audit_correct": 539,
          "audit_total": 1024,
          "feature_index": 84,
          "polarity": 1,
          "selection_correct": 157,
          "selection_total": 256,
          "test_correct": 131,
          "test_total": 256
        },
        "k": 1024,
        "leaky": {
          "audit_correct": 527,
          "audit_total": 1024,
          "feature_index": 129,
          "polarity": 1,
          "test_correct": 156,
          "test_total": 256
        },
        "seed": 9
      },
      {
        "candidate_pair_count": 2,
        "derived": {
          "honest_audit_accuracy": 0.4970703125,
          "honest_selection_accuracy": 0.5234375,
          "honest_test_accuracy": 0.484375,
          "leaky_audit_accuracy": 0.5029296875,
          "leaky_reported_test_accuracy": 0.515625,
          "leaky_vs_honest_test_gap": 0.03125,
          "optimism_gap": 0.0126953125
        },
        "honest": {
          "audit_correct": 509,
          "audit_total": 1024,
          "feature_index": 0,
          "polarity": -1,
          "selection_correct": 134,
          "selection_total": 256,
          "test_correct": 124,
          "test_total": 256
        },
        "k": 1,
        "leaky": {
          "audit_correct": 515,
          "audit_total": 1024,
          "feature_index": 0,
          "polarity": 1,
          "test_correct": 132,
          "test_total": 256
        },
        "seed": 10
      },
      {
        "candidate_pair_count": 4,
        "derived": {
          "honest_audit_accuracy": 0.4970703125,
          "honest_selection_accuracy": 0.5234375,
          "honest_test_accuracy": 0.484375,
          "leaky_audit_accuracy": 0.478515625,
          "leaky_reported_test_accuracy": 0.51953125,
          "leaky_vs_honest_test_gap": 0.03515625,
          "optimism_gap": 0.041015625
        },
        "honest": {
          "audit_correct": 509,
          "audit_total": 1024,
          "feature_index": 0,
          "polarity": -1,
          "selection_correct": 134,
          "selection_total": 256,
          "test_correct": 124,
          "test_total": 256
        },
        "k": 2,
        "leaky": {
          "audit_correct": 490,
          "audit_total": 1024,
          "feature_index": 1,
          "polarity": -1,
          "test_correct": 133,
          "test_total": 256
        },
        "seed": 10
      },
      {
        "candidate_pair_count": 8,
        "derived": {
          "honest_audit_accuracy": 0.494140625,
          "honest_selection_accuracy": 0.5625,
          "honest_test_accuracy": 0.515625,
          "leaky_audit_accuracy": 0.47265625,
          "leaky_reported_test_accuracy": 0.5859375,
          "leaky_vs_honest_test_gap": 0.0703125,
          "optimism_gap": 0.11328125
        },
        "honest": {
          "audit_correct": 506,
          "audit_total": 1024,
          "feature_index": 3,
          "polarity": -1,
          "selection_correct": 144,
          "selection_total": 256,
          "test_correct": 132,
          "test_total": 256
        },
        "k": 4,
        "leaky": {
          "audit_correct": 484,
          "audit_total": 1024,
          "feature_index": 2,
          "polarity": -1,
          "test_correct": 150,
          "test_total": 256
        },
        "seed": 10
      },
      {
        "candidate_pair_count": 16,
        "derived": {
          "honest_audit_accuracy": 0.494140625,
          "honest_selection_accuracy": 0.5625,
          "honest_test_accuracy": 0.515625,
          "leaky_audit_accuracy": 0.47265625,
          "leaky_reported_test_accuracy": 0.5859375,
          "leaky_vs_honest_test_gap": 0.0703125,
          "optimism_gap": 0.11328125
        },
        "honest": {
          "audit_correct": 506,
          "audit_total": 1024,
          "feature_index": 3,
          "polarity": -1,
          "selection_correct": 144,
          "selection_total": 256,
          "test_correct": 132,
          "test_total": 256
        },
        "k": 8,
        "leaky": {
          "audit_correct": 484,
          "audit_total": 1024,
          "feature_index": 2,
          "polarity": -1,
          "test_correct": 150,
          "test_total": 256
        },
        "seed": 10
      },
      {
        "candidate_pair_count": 32,
        "derived": {
          "honest_audit_accuracy": 0.494140625,
          "honest_selection_accuracy": 0.5625,
          "honest_test_accuracy": 0.515625,
          "leaky_audit_accuracy": 0.47265625,
          "leaky_reported_test_accuracy": 0.5859375,
          "leaky_vs_honest_test_gap": 0.0703125,
          "optimism_gap": 0.11328125
        },
        "honest": {
          "audit_correct": 506,
          "audit_total": 1024,
          "feature_index": 3,
          "polarity": -1,
          "selection_correct": 144,
          "selection_total": 256,
          "test_correct": 132,
          "test_total": 256
        },
        "k": 16,
        "leaky": {
          "audit_correct": 484,
          "audit_total": 1024,
          "feature_index": 2,
          "polarity": -1,
          "test_correct": 150,
          "test_total": 256
        },
        "seed": 10
      },
      {
        "candidate_pair_count": 64,
        "derived": {
          "honest_audit_accuracy": 0.494140625,
          "honest_selection_accuracy": 0.5625,
          "honest_test_accuracy": 0.515625,
          "leaky_audit_accuracy": 0.50390625,
          "leaky_reported_test_accuracy": 0.59375,
          "leaky_vs_honest_test_gap": 0.078125,
          "optimism_gap": 0.08984375
        },
        "honest": {
          "audit_correct": 506,
          "audit_total": 1024,
          "feature_index": 3,
          "polarity": -1,
          "selection_correct": 144,
          "selection_total": 256,
          "test_correct": 132,
          "test_total": 256
        },
        "k": 32,
        "leaky": {
          "audit_correct": 516,
          "audit_total": 1024,
          "feature_index": 17,
          "polarity": 1,
          "test_correct": 152,
          "test_total": 256
        },
        "seed": 10
      },
      {
        "candidate_pair_count": 128,
        "derived": {
          "honest_audit_accuracy": 0.4833984375,
          "honest_selection_accuracy": 0.5703125,
          "honest_test_accuracy": 0.515625,
          "leaky_audit_accuracy": 0.50390625,
          "leaky_reported_test_accuracy": 0.59375,
          "leaky_vs_honest_test_gap": 0.078125,
          "optimism_gap": 0.08984375
        },
        "honest": {
          "audit_correct": 495,
          "audit_total": 1024,
          "feature_index": 56,
          "polarity": -1,
          "selection_correct": 146,
          "selection_total": 256,
          "test_correct": 132,
          "test_total": 256
        },
        "k": 64,
        "leaky": {
          "audit_correct": 516,
          "audit_total": 1024,
          "feature_index": 17,
          "polarity": 1,
          "test_correct": 152,
          "test_total": 256
        },
        "seed": 10
      },
      {
        "candidate_pair_count": 256,
        "derived": {
          "honest_audit_accuracy": 0.4814453125,
          "honest_selection_accuracy": 0.578125,
          "honest_test_accuracy": 0.484375,
          "leaky_audit_accuracy": 0.53125,
          "leaky_reported_test_accuracy": 0.609375,
          "leaky_vs_honest_test_gap": 0.125,
          "optimism_gap": 0.078125
        },
        "honest": {
          "audit_correct": 493,
          "audit_total": 1024,
          "feature_index": 90,
          "polarity": 1,
          "selection_correct": 148,
          "selection_total": 256,
          "test_correct": 124,
          "test_total": 256
        },
        "k": 128,
        "leaky": {
          "audit_correct": 544,
          "audit_total": 1024,
          "feature_index": 121,
          "polarity": -1,
          "test_correct": 156,
          "test_total": 256
        },
        "seed": 10
      },
      {
        "candidate_pair_count": 512,
        "derived": {
          "honest_audit_accuracy": 0.4775390625,
          "honest_selection_accuracy": 0.6015625,
          "honest_test_accuracy": 0.5234375,
          "leaky_audit_accuracy": 0.53125,
          "leaky_reported_test_accuracy": 0.609375,
          "leaky_vs_honest_test_gap": 0.0859375,
          "optimism_gap": 0.078125
        },
        "honest": {
          "audit_correct": 489,
          "audit_total": 1024,
          "feature_index": 182,
          "polarity": 1,
          "selection_correct": 154,
          "selection_total": 256,
          "test_correct": 134,
          "test_total": 256
        },
        "k": 256,
        "leaky": {
          "audit_correct": 544,
          "audit_total": 1024,
          "feature_index": 121,
          "polarity": -1,
          "test_correct": 156,
          "test_total": 256
        },
        "seed": 10
      },
      {
        "candidate_pair_count": 1024,
        "derived": {
          "honest_audit_accuracy": 0.5087890625,
          "honest_selection_accuracy": 0.640625,
          "honest_test_accuracy": 0.50390625,
          "leaky_audit_accuracy": 0.53125,
          "leaky_reported_test_accuracy": 0.609375,
          "leaky_vs_honest_test_gap": 0.10546875,
          "optimism_gap": 0.078125
        },
        "honest": {
          "audit_correct": 521,
          "audit_total": 1024,
          "feature_index": 416,
          "polarity": -1,
          "selection_correct": 164,
          "selection_total": 256,
          "test_correct": 129,
          "test_total": 256
        },
        "k": 512,
        "leaky": {
          "audit_correct": 544,
          "audit_total": 1024,
          "feature_index": 121,
          "polarity": -1,
          "test_correct": 156,
          "test_total": 256
        },
        "seed": 10
      },
      {
        "candidate_pair_count": 2048,
        "derived": {
          "honest_audit_accuracy": 0.5087890625,
          "honest_selection_accuracy": 0.640625,
          "honest_test_accuracy": 0.50390625,
          "leaky_audit_accuracy": 0.53125,
          "leaky_reported_test_accuracy": 0.609375,
          "leaky_vs_honest_test_gap": 0.10546875,
          "optimism_gap": 0.078125
        },
        "honest": {
          "audit_correct": 521,
          "audit_total": 1024,
          "feature_index": 416,
          "polarity": -1,
          "selection_correct": 164,
          "selection_total": 256,
          "test_correct": 129,
          "test_total": 256
        },
        "k": 1024,
        "leaky": {
          "audit_correct": 544,
          "audit_total": 1024,
          "feature_index": 121,
          "polarity": -1,
          "test_correct": 156,
          "test_total": 256
        },
        "seed": 10
      },
      {
        "candidate_pair_count": 2,
        "derived": {
          "honest_audit_accuracy": 0.4853515625,
          "honest_selection_accuracy": 0.50390625,
          "honest_test_accuracy": 0.48828125,
          "leaky_audit_accuracy": 0.5146484375,
          "leaky_reported_test_accuracy": 0.51171875,
          "leaky_vs_honest_test_gap": 0.0234375,
          "optimism_gap": -0.0029296875
        },
        "honest": {
          "audit_correct": 497,
          "audit_total": 1024,
          "feature_index": 0,
          "polarity": 1,
          "selection_correct": 129,
          "selection_total": 256,
          "test_correct": 125,
          "test_total": 256
        },
        "k": 1,
        "leaky": {
          "audit_correct": 527,
          "audit_total": 1024,
          "feature_index": 0,
          "polarity": -1,
          "test_correct": 131,
          "test_total": 256
        },
        "seed": 11
      },
      {
        "candidate_pair_count": 4,
        "derived": {
          "honest_audit_accuracy": 0.529296875,
          "honest_selection_accuracy": 0.53125,
          "honest_test_accuracy": 0.46875,
          "leaky_audit_accuracy": 0.470703125,
          "leaky_reported_test_accuracy": 0.53125,
          "leaky_vs_honest_test_gap": 0.0625,
          "optimism_gap": 0.060546875
        },
        "honest": {
          "audit_correct": 542,
          "audit_total": 1024,
          "feature_index": 1,
          "polarity": -1,
          "selection_correct": 136,
          "selection_total": 256,
          "test_correct": 120,
          "test_total": 256
        },
        "k": 2,
        "leaky": {
          "audit_correct": 482,
          "audit_total": 1024,
          "feature_index": 1,
          "polarity": 1,
          "test_correct": 136,
          "test_total": 256
        },
        "seed": 11
      },
      {
        "candidate_pair_count": 8,
        "derived": {
          "honest_audit_accuracy": 0.5302734375,
          "honest_selection_accuracy": 0.54296875,
          "honest_test_accuracy": 0.51953125,
          "leaky_audit_accuracy": 0.4951171875,
          "leaky_reported_test_accuracy": 0.53515625,
          "leaky_vs_honest_test_gap": 0.015625,
          "optimism_gap": 0.0400390625
        },
        "honest": {
          "audit_correct": 543,
          "audit_total": 1024,
          "feature_index": 3,
          "polarity": -1,
          "selection_correct": 139,
          "selection_total": 256,
          "test_correct": 133,
          "test_total": 256
        },
        "k": 4,
        "leaky": {
          "audit_correct": 507,
          "audit_total": 1024,
          "feature_index": 2,
          "polarity": 1,
          "test_correct": 137,
          "test_total": 256
        },
        "seed": 11
      },
      {
        "candidate_pair_count": 16,
        "derived": {
          "honest_audit_accuracy": 0.5302734375,
          "honest_selection_accuracy": 0.54296875,
          "honest_test_accuracy": 0.51953125,
          "leaky_audit_accuracy": 0.4990234375,
          "leaky_reported_test_accuracy": 0.5390625,
          "leaky_vs_honest_test_gap": 0.01953125,
          "optimism_gap": 0.0400390625
        },
        "honest": {
          "audit_correct": 543,
          "audit_total": 1024,
          "feature_index": 3,
          "polarity": -1,
          "selection_correct": 139,
          "selection_total": 256,
          "test_correct": 133,
          "test_total": 256
        },
        "k": 8,
        "leaky": {
          "audit_correct": 511,
          "audit_total": 1024,
          "feature_index": 7,
          "polarity": 1,
          "test_correct": 138,
          "test_total": 256
        },
        "seed": 11
      },
      {
        "candidate_pair_count": 32,
        "derived": {
          "honest_audit_accuracy": 0.505859375,
          "honest_selection_accuracy": 0.55078125,
          "honest_test_accuracy": 0.4921875,
          "leaky_audit_accuracy": 0.51171875,
          "leaky_reported_test_accuracy": 0.546875,
          "leaky_vs_honest_test_gap": 0.0546875,
          "optimism_gap": 0.03515625
        },
        "honest": {
          "audit_correct": 518,
          "audit_total": 1024,
          "feature_index": 13,
          "polarity": 1,
          "selection_correct": 141,
          "selection_total": 256,
          "test_correct": 126,
          "test_total": 256
        },
        "k": 16,
        "leaky": {
          "audit_correct": 524,
          "audit_total": 1024,
          "feature_index": 8,
          "polarity": 1,
          "test_correct": 140,
          "test_total": 256
        },
        "seed": 11
      },
      {
        "candidate_pair_count": 64,
        "derived": {
          "honest_audit_accuracy": 0.505859375,
          "honest_selection_accuracy": 0.55078125,
          "honest_test_accuracy": 0.4921875,
          "leaky_audit_accuracy": 0.4814453125,
          "leaky_reported_test_accuracy": 0.5625,
          "leaky_vs_honest_test_gap": 0.0703125,
          "optimism_gap": 0.0810546875
        },
        "honest": {
          "audit_correct": 518,
          "audit_total": 1024,
          "feature_index": 13,
          "polarity": 1,
          "selection_correct": 141,
          "selection_total": 256,
          "test_correct": 126,
          "test_total": 256
        },
        "k": 32,
        "leaky": {
          "audit_correct": 493,
          "audit_total": 1024,
          "feature_index": 20,
          "polarity": 1,
          "test_correct": 144,
          "test_total": 256
        },
        "seed": 11
      },
      {
        "candidate_pair_count": 128,
        "derived": {
          "honest_audit_accuracy": 0.4951171875,
          "honest_selection_accuracy": 0.55859375,
          "honest_test_accuracy": 0.4765625,
          "leaky_audit_accuracy": 0.494140625,
          "leaky_reported_test_accuracy": 0.56640625,
          "leaky_vs_honest_test_gap": 0.08984375,
          "optimism_gap": 0.072265625
        },
        "honest": {
          "audit_correct": 507,
          "audit_total": 1024,
          "feature_index": 51,
          "polarity": 1,
          "selection_correct": 143,
          "selection_total": 256,
          "test_correct": 122,
          "test_total": 256
        },
        "k": 64,
        "leaky": {
          "audit_correct": 506,
          "audit_total": 1024,
          "feature_index": 37,
          "polarity": -1,
          "test_correct": 145,
          "test_total": 256
        },
        "seed": 11
      },
      {
        "candidate_pair_count": 256,
        "derived": {
          "honest_audit_accuracy": 0.4892578125,
          "honest_selection_accuracy": 0.58203125,
          "honest_test_accuracy": 0.50390625,
          "leaky_audit_accuracy": 0.4951171875,
          "leaky_reported_test_accuracy": 0.5703125,
          "leaky_vs_honest_test_gap": 0.06640625,
          "optimism_gap": 0.0751953125
        },
        "honest": {
          "audit_correct": 501,
          "audit_total": 1024,
          "feature_index": 112,
          "polarity": 1,
          "selection_correct": 149,
          "selection_total": 256,
          "test_correct": 129,
          "test_total": 256
        },
        "k": 128,
        "leaky": {
          "audit_correct": 507,
          "audit_total": 1024,
          "feature_index": 85,
          "polarity": 1,
          "test_correct": 146,
          "test_total": 256
        },
        "seed": 11
      },
      {
        "candidate_pair_count": 512,
        "derived": {
          "honest_audit_accuracy": 0.5029296875,
          "honest_selection_accuracy": 0.60546875,
          "honest_test_accuracy": 0.5234375,
          "leaky_audit_accuracy": 0.515625,
          "leaky_reported_test_accuracy": 0.5859375,
          "leaky_vs_honest_test_gap": 0.0625,
          "optimism_gap": 0.0703125
        },
        "honest": {
          "audit_correct": 515,
          "audit_total": 1024,
          "feature_index": 249,
          "polarity": -1,
          "selection_correct": 155,
          "selection_total": 256,
          "test_correct": 134,
          "test_total": 256
        },
        "k": 256,
        "leaky": {
          "audit_correct": 528,
          "audit_total": 1024,
          "feature_index": 244,
          "polarity": -1,
          "test_correct": 150,
          "test_total": 256
        },
        "seed": 11
      },
      {
        "candidate_pair_count": 1024,
        "derived": {
          "honest_audit_accuracy": 0.5263671875,
          "honest_selection_accuracy": 0.62109375,
          "honest_test_accuracy": 0.51953125,
          "leaky_audit_accuracy": 0.5068359375,
          "leaky_reported_test_accuracy": 0.59375,
          "leaky_vs_honest_test_gap": 0.07421875,
          "optimism_gap": 0.0869140625
        },
        "honest": {
          "audit_correct": 539,
          "audit_total": 1024,
          "feature_index": 398,
          "polarity": 1,
          "selection_correct": 159,
          "selection_total": 256,
          "test_correct": 133,
          "test_total": 256
        },
        "k": 512,
        "leaky": {
          "audit_correct": 519,
          "audit_total": 1024,
          "feature_index": 275,
          "polarity": -1,
          "test_correct": 152,
          "test_total": 256
        },
        "seed": 11
      },
      {
        "candidate_pair_count": 2048,
        "derived": {
          "honest_audit_accuracy": 0.517578125,
          "honest_selection_accuracy": 0.6328125,
          "honest_test_accuracy": 0.53125,
          "leaky_audit_accuracy": 0.5068359375,
          "leaky_reported_test_accuracy": 0.59375,
          "leaky_vs_honest_test_gap": 0.0625,
          "optimism_gap": 0.0869140625
        },
        "honest": {
          "audit_correct": 530,
          "audit_total": 1024,
          "feature_index": 585,
          "polarity": 1,
          "selection_correct": 162,
          "selection_total": 256,
          "test_correct": 136,
          "test_total": 256
        },
        "k": 1024,
        "leaky": {
          "audit_correct": 519,
          "audit_total": 1024,
          "feature_index": 275,
          "polarity": -1,
          "test_correct": 152,
          "test_total": 256
        },
        "seed": 11
      },
      {
        "candidate_pair_count": 2,
        "derived": {
          "honest_audit_accuracy": 0.4853515625,
          "honest_selection_accuracy": 0.5234375,
          "honest_test_accuracy": 0.49609375,
          "leaky_audit_accuracy": 0.5146484375,
          "leaky_reported_test_accuracy": 0.50390625,
          "leaky_vs_honest_test_gap": 0.0078125,
          "optimism_gap": -0.0107421875
        },
        "honest": {
          "audit_correct": 497,
          "audit_total": 1024,
          "feature_index": 0,
          "polarity": 1,
          "selection_correct": 134,
          "selection_total": 256,
          "test_correct": 127,
          "test_total": 256
        },
        "k": 1,
        "leaky": {
          "audit_correct": 527,
          "audit_total": 1024,
          "feature_index": 0,
          "polarity": -1,
          "test_correct": 129,
          "test_total": 256
        },
        "seed": 12
      },
      {
        "candidate_pair_count": 4,
        "derived": {
          "honest_audit_accuracy": 0.5009765625,
          "honest_selection_accuracy": 0.54296875,
          "honest_test_accuracy": 0.53125,
          "leaky_audit_accuracy": 0.5009765625,
          "leaky_reported_test_accuracy": 0.53125,
          "leaky_vs_honest_test_gap": 0.0,
          "optimism_gap": 0.0302734375
        },
        "honest": {
          "audit_correct": 513,
          "audit_total": 1024,
          "feature_index": 1,
          "polarity": 1,
          "selection_correct": 139,
          "selection_total": 256,
          "test_correct": 136,
          "test_total": 256
        },
        "k": 2,
        "leaky": {
          "audit_correct": 513,
          "audit_total": 1024,
          "feature_index": 1,
          "polarity": 1,
          "test_correct": 136,
          "test_total": 256
        },
        "seed": 12
      },
      {
        "candidate_pair_count": 8,
        "derived": {
          "honest_audit_accuracy": 0.505859375,
          "honest_selection_accuracy": 0.55859375,
          "honest_test_accuracy": 0.5234375,
          "leaky_audit_accuracy": 0.5009765625,
          "leaky_reported_test_accuracy": 0.53125,
          "leaky_vs_honest_test_gap": 0.0078125,
          "optimism_gap": 0.0302734375
        },
        "honest": {
          "audit_correct": 518,
          "audit_total": 1024,
          "feature_index": 3,
          "polarity": -1,
          "selection_correct": 143,
          "selection_total": 256,
          "test_correct": 134,
          "test_total": 256
        },
        "k": 4,
        "leaky": {
          "audit_correct": 513,
          "audit_total": 1024,
          "feature_index": 1,
          "polarity": 1,
          "test_correct": 136,
          "test_total": 256
        },
        "seed": 12
      },
      {
        "candidate_pair_count": 16,
        "derived": {
          "honest_audit_accuracy": 0.505859375,
          "honest_selection_accuracy": 0.55859375,
          "honest_test_accuracy": 0.5234375,
          "leaky_audit_accuracy": 0.50390625,
          "leaky_reported_test_accuracy": 0.546875,
          "leaky_vs_honest_test_gap": 0.0234375,
          "optimism_gap": 0.04296875
        },
        "honest": {
          "audit_correct": 518,
          "audit_total": 1024,
          "feature_index": 3,
          "polarity": -1,
          "selection_correct": 143,
          "selection_total": 256,
          "test_correct": 134,
          "test_total": 256
        },
        "k": 8,
        "leaky": {
          "audit_correct": 516,
          "audit_total": 1024,
          "feature_index": 7,
          "polarity": -1,
          "test_correct": 140,
          "test_total": 256
        },
        "seed": 12
      },
      {
        "candidate_pair_count": 32,
        "derived": {
          "honest_audit_accuracy": 0.505859375,
          "honest_selection_accuracy": 0.55859375,
          "honest_test_accuracy": 0.5234375,
          "leaky_audit_accuracy": 0.5009765625,
          "leaky_reported_test_accuracy": 0.56640625,
          "leaky_vs_honest_test_gap": 0.04296875,
          "optimism_gap": 0.0654296875
        },
        "honest": {
          "audit_correct": 518,
          "audit_total": 1024,
          "feature_index": 3,
          "polarity": -1,
          "selection_correct": 143,
          "selection_total": 256,
          "test_correct": 134,
          "test_total": 256
        },
        "k": 16,
        "leaky": {
          "audit_correct": 513,
          "audit_total": 1024,
          "feature_index": 9,
          "polarity": -1,
          "test_correct": 145,
          "test_total": 256
        },
        "seed": 12
      },
      {
        "candidate_pair_count": 64,
        "derived": {
          "honest_audit_accuracy": 0.501953125,
          "honest_selection_accuracy": 0.58203125,
          "honest_test_accuracy": 0.49609375,
          "leaky_audit_accuracy": 0.5009765625,
          "leaky_reported_test_accuracy": 0.56640625,
          "leaky_vs_honest_test_gap": 0.0703125,
          "optimism_gap": 0.0654296875
        },
        "honest": {
          "audit_correct": 514,
          "audit_total": 1024,
          "feature_index": 30,
          "polarity": -1,
          "selection_correct": 149,
          "selection_total": 256,
          "test_correct": 127,
          "test_total": 256
        },
        "k": 32,
        "leaky": {
          "audit_correct": 513,
          "audit_total": 1024,
          "feature_index": 9,
          "polarity": -1,
          "test_correct": 145,
          "test_total": 256
        },
        "seed": 12
      },
      {
        "candidate_pair_count": 128,
        "derived": {
          "honest_audit_accuracy": 0.501953125,
          "honest_selection_accuracy": 0.58203125,
          "honest_test_accuracy": 0.49609375,
          "leaky_audit_accuracy": 0.5009765625,
          "leaky_reported_test_accuracy": 0.56640625,
          "leaky_vs_honest_test_gap": 0.0703125,
          "optimism_gap": 0.0654296875
        },
        "honest": {
          "audit_correct": 514,
          "audit_total": 1024,
          "feature_index": 30,
          "polarity": -1,
          "selection_correct": 149,
          "selection_total": 256,
          "test_correct": 127,
          "test_total": 256
        },
        "k": 64,
        "leaky": {
          "audit_correct": 513,
          "audit_total": 1024,
          "feature_index": 9,
          "polarity": -1,
          "test_correct": 145,
          "test_total": 256
        },
        "seed": 12
      },
      {
        "candidate_pair_count": 256,
        "derived": {
          "honest_audit_accuracy": 0.50390625,
          "honest_selection_accuracy": 0.5859375,
          "honest_test_accuracy": 0.5,
          "leaky_audit_accuracy": 0.5078125,
          "leaky_reported_test_accuracy": 0.5859375,
          "leaky_vs_honest_test_gap": 0.0859375,
          "optimism_gap": 0.078125
        },
        "honest": {
          "audit_correct": 516,
          "audit_total": 1024,
          "feature_index": 112,
          "polarity": -1,
          "selection_correct": 150,
          "selection_total": 256,
          "test_correct": 128,
          "test_total": 256
        },
        "k": 128,
        "leaky": {
          "audit_correct": 520,
          "audit_total": 1024,
          "feature_index": 115,
          "polarity": 1,
          "test_correct": 150,
          "test_total": 256
        },
        "seed": 12
      },
      {
        "candidate_pair_count": 512,
        "derived": {
          "honest_audit_accuracy": 0.5029296875,
          "honest_selection_accuracy": 0.58984375,
          "honest_test_accuracy": 0.49609375,
          "leaky_audit_accuracy": 0.5078125,
          "leaky_reported_test_accuracy": 0.5859375,
          "leaky_vs_honest_test_gap": 0.08984375,
          "optimism_gap": 0.078125
        },
        "honest": {
          "audit_correct": 515,
          "audit_total": 1024,
          "feature_index": 250,
          "polarity": 1,
          "selection_correct": 151,
          "selection_total": 256,
          "test_correct": 127,
          "test_total": 256
        },
        "k": 256,
        "leaky": {
          "audit_correct": 520,
          "audit_total": 1024,
          "feature_index": 115,
          "polarity": 1,
          "test_correct": 150,
          "test_total": 256
        },
        "seed": 12
      },
      {
        "candidate_pair_count": 1024,
        "derived": {
          "honest_audit_accuracy": 0.4677734375,
          "honest_selection_accuracy": 0.609375,
          "honest_test_accuracy": 0.51953125,
          "leaky_audit_accuracy": 0.5078125,
          "leaky_reported_test_accuracy": 0.5859375,
          "leaky_vs_honest_test_gap": 0.06640625,
          "optimism_gap": 0.078125
        },
        "honest": {
          "audit_correct": 479,
          "audit_total": 1024,
          "feature_index": 369,
          "polarity": 1,
          "selection_correct": 156,
          "selection_total": 256,
          "test_correct": 133,
          "test_total": 256
        },
        "k": 512,
        "leaky": {
          "audit_correct": 520,
          "audit_total": 1024,
          "feature_index": 115,
          "polarity": 1,
          "test_correct": 150,
          "test_total": 256
        },
        "seed": 12
      },
      {
        "candidate_pair_count": 2048,
        "derived": {
          "honest_audit_accuracy": 0.4677734375,
          "honest_selection_accuracy": 0.609375,
          "honest_test_accuracy": 0.51953125,
          "leaky_audit_accuracy": 0.5224609375,
          "leaky_reported_test_accuracy": 0.6015625,
          "leaky_vs_honest_test_gap": 0.08203125,
          "optimism_gap": 0.0791015625
        },
        "honest": {
          "audit_correct": 479,
          "audit_total": 1024,
          "feature_index": 369,
          "polarity": 1,
          "selection_correct": 156,
          "selection_total": 256,
          "test_correct": 133,
          "test_total": 256
        },
        "k": 1024,
        "leaky": {
          "audit_correct": 535,
          "audit_total": 1024,
          "feature_index": 847,
          "polarity": 1,
          "test_correct": 154,
          "test_total": 256
        },
        "seed": 12
      },
      {
        "candidate_pair_count": 2,
        "derived": {
          "honest_audit_accuracy": 0.5078125,
          "honest_selection_accuracy": 0.5234375,
          "honest_test_accuracy": 0.5234375,
          "leaky_audit_accuracy": 0.5078125,
          "leaky_reported_test_accuracy": 0.5234375,
          "leaky_vs_honest_test_gap": 0.0,
          "optimism_gap": 0.015625
        },
        "honest": {
          "audit_correct": 520,
          "audit_total": 1024,
          "feature_index": 0,
          "polarity": -1,
          "selection_correct": 134,
          "selection_total": 256,
          "test_correct": 134,
          "test_total": 256
        },
        "k": 1,
        "leaky": {
          "audit_correct": 520,
          "audit_total": 1024,
          "feature_index": 0,
          "polarity": -1,
          "test_correct": 134,
          "test_total": 256
        },
        "seed": 13
      },
      {
        "candidate_pair_count": 4,
        "derived": {
          "honest_audit_accuracy": 0.5078125,
          "honest_selection_accuracy": 0.5234375,
          "honest_test_accuracy": 0.5234375,
          "leaky_audit_accuracy": 0.4873046875,
          "leaky_reported_test_accuracy": 0.52734375,
          "leaky_vs_honest_test_gap": 0.00390625,
          "optimism_gap": 0.0400390625
        },
        "honest": {
          "audit_correct": 520,
          "audit_total": 1024,
          "feature_index": 0,
          "polarity": -1,
          "selection_correct": 134,
          "selection_total": 256,
          "test_correct": 134,
          "test_total": 256
        },
        "k": 2,
        "leaky": {
          "audit_correct": 499,
          "audit_total": 1024,
          "feature_index": 1,
          "polarity": 1,
          "test_correct": 135,
          "test_total": 256
        },
        "seed": 13
      },
      {
        "candidate_pair_count": 8,
        "derived": {
          "honest_audit_accuracy": 0.5078125,
          "honest_selection_accuracy": 0.5234375,
          "honest_test_accuracy": 0.5234375,
          "leaky_audit_accuracy": 0.4873046875,
          "leaky_reported_test_accuracy": 0.52734375,
          "leaky_vs_honest_test_gap": 0.00390625,
          "optimism_gap": 0.0400390625
        },
        "honest": {
          "audit_correct": 520,
          "audit_total": 1024,
          "feature_index": 0,
          "polarity": -1,
          "selection_correct": 134,
          "selection_total": 256,
          "test_correct": 134,
          "test_total": 256
        },
        "k": 4,
        "leaky": {
          "audit_correct": 499,
          "audit_total": 1024,
          "feature_index": 1,
          "polarity": 1,
          "test_correct": 135,
          "test_total": 256
        },
        "seed": 13
      },
      {
        "candidate_pair_count": 16,
        "derived": {
          "honest_audit_accuracy": 0.5078125,
          "honest_selection_accuracy": 0.5234375,
          "honest_test_accuracy": 0.5234375,
          "leaky_audit_accuracy": 0.4970703125,
          "leaky_reported_test_accuracy": 0.53125,
          "leaky_vs_honest_test_gap": 0.0078125,
          "optimism_gap": 0.0341796875
        },
        "honest": {
          "audit_correct": 520,
          "audit_total": 1024,
          "feature_index": 0,
          "polarity": -1,
          "selection_correct": 134,
          "selection_total": 256,
          "test_correct": 134,
          "test_total": 256
        },
        "k": 8,
        "leaky": {
          "audit_correct": 509,
          "audit_total": 1024,
          "feature_index": 4,
          "polarity": 1,
          "test_correct": 136,
          "test_total": 256
        },
        "seed": 13
      },
      {
        "candidate_pair_count": 32,
        "derived": {
          "honest_audit_accuracy": 0.5166015625,
          "honest_selection_accuracy": 0.55078125,
          "honest_test_accuracy": 0.55859375,
          "leaky_audit_accuracy": 0.4892578125,
          "leaky_reported_test_accuracy": 0.57421875,
          "leaky_vs_honest_test_gap": 0.015625,
          "optimism_gap": 0.0849609375
        },
        "honest": {
          "audit_correct": 529,
          "audit_total": 1024,
          "feature_index": 8,
          "polarity": -1,
          "selection_correct": 141,
          "selection_total": 256,
          "test_correct": 143,
          "test_total": 256
        },
        "k": 16,
        "leaky": {
          "audit_correct": 501,
          "audit_total": 1024,
          "feature_index": 10,
          "polarity": 1,
          "test_correct": 147,
          "test_total": 256
        },
        "seed": 13
      },
      {
        "candidate_pair_count": 64,
        "derived": {
          "honest_audit_accuracy": 0.50390625,
          "honest_selection_accuracy": 0.5546875,
          "honest_test_accuracy": 0.47265625,
          "leaky_audit_accuracy": 0.4892578125,
          "leaky_reported_test_accuracy": 0.57421875,
          "leaky_vs_honest_test_gap": 0.1015625,
          "optimism_gap": 0.0849609375
        },
        "honest": {
          "audit_correct": 516,
          "audit_total": 1024,
          "feature_index": 25,
          "polarity": -1,
          "selection_correct": 142,
          "selection_total": 256,
          "test_correct": 121,
          "test_total": 256
        },
        "k": 32,
        "leaky": {
          "audit_correct": 501,
          "audit_total": 1024,
          "feature_index": 10,
          "polarity": 1,
          "test_correct": 147,
          "test_total": 256
        },
        "seed": 13
      },
      {
        "candidate_pair_count": 128,
        "derived": {
          "honest_audit_accuracy": 0.4833984375,
          "honest_selection_accuracy": 0.5703125,
          "honest_test_accuracy": 0.50390625,
          "leaky_audit_accuracy": 0.4970703125,
          "leaky_reported_test_accuracy": 0.58984375,
          "leaky_vs_honest_test_gap": 0.0859375,
          "optimism_gap": 0.0927734375
        },
        "honest": {
          "audit_correct": 495,
          "audit_total": 1024,
          "feature_index": 53,
          "polarity": -1,
          "selection_correct": 146,
          "selection_total": 256,
          "test_correct": 129,
          "test_total": 256
        },
        "k": 64,
        "leaky": {
          "audit_correct": 509,
          "audit_total": 1024,
          "feature_index": 59,
          "polarity": 1,
          "test_correct": 151,
          "test_total": 256
        },
        "seed": 13
      },
      {
        "candidate_pair_count": 256,
        "derived": {
          "honest_audit_accuracy": 0.517578125,
          "honest_selection_accuracy": 0.58203125,
          "honest_test_accuracy": 0.5078125,
          "leaky_audit_accuracy": 0.4970703125,
          "leaky_reported_test_accuracy": 0.58984375,
          "leaky_vs_honest_test_gap": 0.08203125,
          "optimism_gap": 0.0927734375
        },
        "honest": {
          "audit_correct": 530,
          "audit_total": 1024,
          "feature_index": 96,
          "polarity": 1,
          "selection_correct": 149,
          "selection_total": 256,
          "test_correct": 130,
          "test_total": 256
        },
        "k": 128,
        "leaky": {
          "audit_correct": 509,
          "audit_total": 1024,
          "feature_index": 59,
          "polarity": 1,
          "test_correct": 151,
          "test_total": 256
        },
        "seed": 13
      },
      {
        "candidate_pair_count": 512,
        "derived": {
          "honest_audit_accuracy": 0.517578125,
          "honest_selection_accuracy": 0.58203125,
          "honest_test_accuracy": 0.5078125,
          "leaky_audit_accuracy": 0.4970703125,
          "leaky_reported_test_accuracy": 0.58984375,
          "leaky_vs_honest_test_gap": 0.08203125,
          "optimism_gap": 0.0927734375
        },
        "honest": {
          "audit_correct": 530,
          "audit_total": 1024,
          "feature_index": 96,
          "polarity": 1,
          "selection_correct": 149,
          "selection_total": 256,
          "test_correct": 130,
          "test_total": 256
        },
        "k": 256,
        "leaky": {
          "audit_correct": 509,
          "audit_total": 1024,
          "feature_index": 59,
          "polarity": 1,
          "test_correct": 151,
          "test_total": 256
        },
        "seed": 13
      },
      {
        "candidate_pair_count": 1024,
        "derived": {
          "honest_audit_accuracy": 0.517578125,
          "honest_selection_accuracy": 0.58203125,
          "honest_test_accuracy": 0.5078125,
          "leaky_audit_accuracy": 0.4873046875,
          "leaky_reported_test_accuracy": 0.6171875,
          "leaky_vs_honest_test_gap": 0.109375,
          "optimism_gap": 0.1298828125
        },
        "honest": {
          "audit_correct": 530,
          "audit_total": 1024,
          "feature_index": 96,
          "polarity": 1,
          "selection_correct": 149,
          "selection_total": 256,
          "test_correct": 130,
          "test_total": 256
        },
        "k": 512,
        "leaky": {
          "audit_correct": 499,
          "audit_total": 1024,
          "feature_index": 345,
          "polarity": -1,
          "test_correct": 158,
          "test_total": 256
        },
        "seed": 13
      },
      {
        "candidate_pair_count": 2048,
        "derived": {
          "honest_audit_accuracy": 0.5078125,
          "honest_selection_accuracy": 0.6015625,
          "honest_test_accuracy": 0.453125,
          "leaky_audit_accuracy": 0.4873046875,
          "leaky_reported_test_accuracy": 0.6171875,
          "leaky_vs_honest_test_gap": 0.1640625,
          "optimism_gap": 0.1298828125
        },
        "honest": {
          "audit_correct": 520,
          "audit_total": 1024,
          "feature_index": 797,
          "polarity": -1,
          "selection_correct": 154,
          "selection_total": 256,
          "test_correct": 116,
          "test_total": 256
        },
        "k": 1024,
        "leaky": {
          "audit_correct": 499,
          "audit_total": 1024,
          "feature_index": 345,
          "polarity": -1,
          "test_correct": 158,
          "test_total": 256
        },
        "seed": 13
      },
      {
        "candidate_pair_count": 2,
        "derived": {
          "honest_audit_accuracy": 0.5068359375,
          "honest_selection_accuracy": 0.50390625,
          "honest_test_accuracy": 0.5703125,
          "leaky_audit_accuracy": 0.5068359375,
          "leaky_reported_test_accuracy": 0.5703125,
          "leaky_vs_honest_test_gap": 0.0,
          "optimism_gap": 0.0634765625
        },
        "honest": {
          "audit_correct": 519,
          "audit_total": 1024,
          "feature_index": 0,
          "polarity": -1,
          "selection_correct": 129,
          "selection_total": 256,
          "test_correct": 146,
          "test_total": 256
        },
        "k": 1,
        "leaky": {
          "audit_correct": 519,
          "audit_total": 1024,
          "feature_index": 0,
          "polarity": -1,
          "test_correct": 146,
          "test_total": 256
        },
        "seed": 14
      },
      {
        "candidate_pair_count": 4,
        "derived": {
          "honest_audit_accuracy": 0.5341796875,
          "honest_selection_accuracy": 0.51953125,
          "honest_test_accuracy": 0.5,
          "leaky_audit_accuracy": 0.5068359375,
          "leaky_reported_test_accuracy": 0.5703125,
          "leaky_vs_honest_test_gap": 0.0703125,
          "optimism_gap": 0.0634765625
        },
        "honest": {
          "audit_correct": 547,
          "audit_total": 1024,
          "feature_index": 1,
          "polarity": 1,
          "selection_correct": 133,
          "selection_total": 256,
          "test_correct": 128,
          "test_total": 256
        },
        "k": 2,
        "leaky": {
          "audit_correct": 519,
          "audit_total": 1024,
          "feature_index": 0,
          "polarity": -1,
          "test_correct": 146,
          "test_total": 256
        },
        "seed": 14
      },
      {
        "candidate_pair_count": 8,
        "derived": {
          "honest_audit_accuracy": 0.4990234375,
          "honest_selection_accuracy": 0.578125,
          "honest_test_accuracy": 0.54296875,
          "leaky_audit_accuracy": 0.5068359375,
          "leaky_reported_test_accuracy": 0.5703125,
          "leaky_vs_honest_test_gap": 0.02734375,
          "optimism_gap": 0.0634765625
        },
        "honest": {
          "audit_correct": 511,
          "audit_total": 1024,
          "feature_index": 3,
          "polarity": 1,
          "selection_correct": 148,
          "selection_total": 256,
          "test_correct": 139,
          "test_total": 256
        },
        "k": 4,
        "leaky": {
          "audit_correct": 519,
          "audit_total": 1024,
          "feature_index": 0,
          "polarity": -1,
          "test_correct": 146,
          "test_total": 256
        },
        "seed": 14
      },
      {
        "candidate_pair_count": 16,
        "derived": {
          "honest_audit_accuracy": 0.4990234375,
          "honest_selection_accuracy": 0.578125,
          "honest_test_accuracy": 0.54296875,
          "leaky_audit_accuracy": 0.5068359375,
          "leaky_reported_test_accuracy": 0.5703125,
          "leaky_vs_honest_test_gap": 0.02734375,
          "optimism_gap": 0.0634765625
        },
        "honest": {
          "audit_correct": 511,
          "audit_total": 1024,
          "feature_index": 3,
          "polarity": 1,
          "selection_correct": 148,
          "selection_total": 256,
          "test_correct": 139,
          "test_total": 256
        },
        "k": 8,
        "leaky": {
          "audit_correct": 519,
          "audit_total": 1024,
          "feature_index": 0,
          "polarity": -1,
          "test_correct": 146,
          "test_total": 256
        },
        "seed": 14
      },
      {
        "candidate_pair_count": 32,
        "derived": {
          "honest_audit_accuracy": 0.4990234375,
          "honest_selection_accuracy": 0.578125,
          "honest_test_accuracy": 0.54296875,
          "leaky_audit_accuracy": 0.5068359375,
          "leaky_reported_test_accuracy": 0.5703125,
          "leaky_vs_honest_test_gap": 0.02734375,
          "optimism_gap": 0.0634765625
        },
        "honest": {
          "audit_correct": 511,
          "audit_total": 1024,
          "feature_index": 3,
          "polarity": 1,
          "selection_correct": 148,
          "selection_total": 256,
          "test_correct": 139,
          "test_total": 256
        },
        "k": 16,
        "leaky": {
          "audit_correct": 519,
          "audit_total": 1024,
          "feature_index": 0,
          "polarity": -1,
          "test_correct": 146,
          "test_total": 256
        },
        "seed": 14
      },
      {
        "candidate_pair_count": 64,
        "derived": {
          "honest_audit_accuracy": 0.4990234375,
          "honest_selection_accuracy": 0.578125,
          "honest_test_accuracy": 0.54296875,
          "leaky_audit_accuracy": 0.5068359375,
          "leaky_reported_test_accuracy": 0.5703125,
          "leaky_vs_honest_test_gap": 0.02734375,
          "optimism_gap": 0.0634765625
        },
        "honest": {
          "audit_correct": 511,
          "audit_total": 1024,
          "feature_index": 3,
          "polarity": 1,
          "selection_correct": 148,
          "selection_total": 256,
          "test_correct": 139,
          "test_total": 256
        },
        "k": 32,
        "leaky": {
          "audit_correct": 519,
          "audit_total": 1024,
          "feature_index": 0,
          "polarity": -1,
          "test_correct": 146,
          "test_total": 256
        },
        "seed": 14
      },
      {
        "candidate_pair_count": 128,
        "derived": {
          "honest_audit_accuracy": 0.4990234375,
          "honest_selection_accuracy": 0.578125,
          "honest_test_accuracy": 0.54296875,
          "leaky_audit_accuracy": 0.5068359375,
          "leaky_reported_test_accuracy": 0.5703125,
          "leaky_vs_honest_test_gap": 0.02734375,
          "optimism_gap": 0.0634765625
        },
        "honest": {
          "audit_correct": 511,
          "audit_total": 1024,
          "feature_index": 3,
          "polarity": 1,
          "selection_correct": 148,
          "selection_total": 256,
          "test_correct": 139,
          "test_total": 256
        },
        "k": 64,
        "leaky": {
          "audit_correct": 519,
          "audit_total": 1024,
          "feature_index": 0,
          "polarity": -1,
          "test_correct": 146,
          "test_total": 256
        },
        "seed": 14
      },
      {
        "candidate_pair_count": 256,
        "derived": {
          "honest_audit_accuracy": 0.51953125,
          "honest_selection_accuracy": 0.59375,
          "honest_test_accuracy": 0.51953125,
          "leaky_audit_accuracy": 0.515625,
          "leaky_reported_test_accuracy": 0.609375,
          "leaky_vs_honest_test_gap": 0.08984375,
          "optimism_gap": 0.09375
        },
        "honest": {
          "audit_correct": 532,
          "audit_total": 1024,
          "feature_index": 72,
          "polarity": -1,
          "selection_correct": 152,
          "selection_total": 256,
          "test_correct": 133,
          "test_total": 256
        },
        "k": 128,
        "leaky": {
          "audit_correct": 528,
          "audit_total": 1024,
          "feature_index": 67,
          "polarity": -1,
          "test_correct": 156,
          "test_total": 256
        },
        "seed": 14
      },
      {
        "candidate_pair_count": 512,
        "derived": {
          "honest_audit_accuracy": 0.51953125,
          "honest_selection_accuracy": 0.59375,
          "honest_test_accuracy": 0.51953125,
          "leaky_audit_accuracy": 0.515625,
          "leaky_reported_test_accuracy": 0.609375,
          "leaky_vs_honest_test_gap": 0.08984375,
          "optimism_gap": 0.09375
        },
        "honest": {
          "audit_correct": 532,
          "audit_total": 1024,
          "feature_index": 72,
          "polarity": -1,
          "selection_correct": 152,
          "selection_total": 256,
          "test_correct": 133,
          "test_total": 256
        },
        "k": 256,
        "leaky": {
          "audit_correct": 528,
          "audit_total": 1024,
          "feature_index": 67,
          "polarity": -1,
          "test_correct": 156,
          "test_total": 256
        },
        "seed": 14
      },
      {
        "candidate_pair_count": 1024,
        "derived": {
          "honest_audit_accuracy": 0.4853515625,
          "honest_selection_accuracy": 0.59765625,
          "honest_test_accuracy": 0.4765625,
          "leaky_audit_accuracy": 0.515625,
          "leaky_reported_test_accuracy": 0.609375,
          "leaky_vs_honest_test_gap": 0.1328125,
          "optimism_gap": 0.09375
        },
        "honest": {
          "audit_correct": 497,
          "audit_total": 1024,
          "feature_index": 416,
          "polarity": 1,
          "selection_correct": 153,
          "selection_total": 256,
          "test_correct": 122,
          "test_total": 256
        },
        "k": 512,
        "leaky": {
          "audit_correct": 528,
          "audit_total": 1024,
          "feature_index": 67,
          "polarity": -1,
          "test_correct": 156,
          "test_total": 256
        },
        "seed": 14
      },
      {
        "candidate_pair_count": 2048,
        "derived": {
          "honest_audit_accuracy": 0.46875,
          "honest_selection_accuracy": 0.60546875,
          "honest_test_accuracy": 0.515625,
          "leaky_audit_accuracy": 0.515625,
          "leaky_reported_test_accuracy": 0.609375,
          "leaky_vs_honest_test_gap": 0.09375,
          "optimism_gap": 0.09375
        },
        "honest": {
          "audit_correct": 480,
          "audit_total": 1024,
          "feature_index": 601,
          "polarity": -1,
          "selection_correct": 155,
          "selection_total": 256,
          "test_correct": 132,
          "test_total": 256
        },
        "k": 1024,
        "leaky": {
          "audit_correct": 528,
          "audit_total": 1024,
          "feature_index": 67,
          "polarity": -1,
          "test_correct": 156,
          "test_total": 256
        },
        "seed": 14
      },
      {
        "candidate_pair_count": 2,
        "derived": {
          "honest_audit_accuracy": 0.517578125,
          "honest_selection_accuracy": 0.5234375,
          "honest_test_accuracy": 0.515625,
          "leaky_audit_accuracy": 0.517578125,
          "leaky_reported_test_accuracy": 0.515625,
          "leaky_vs_honest_test_gap": 0.0,
          "optimism_gap": -0.001953125
        },
        "honest": {
          "audit_correct": 530,
          "audit_total": 1024,
          "feature_index": 0,
          "polarity": -1,
          "selection_correct": 134,
          "selection_total": 256,
          "test_correct": 132,
          "test_total": 256
        },
        "k": 1,
        "leaky": {
          "audit_correct": 530,
          "audit_total": 1024,
          "feature_index": 0,
          "polarity": -1,
          "test_correct": 132,
          "test_total": 256
        },
        "seed": 15
      },
      {
        "candidate_pair_count": 4,
        "derived": {
          "honest_audit_accuracy": 0.50390625,
          "honest_selection_accuracy": 0.53125,
          "honest_test_accuracy": 0.51953125,
          "leaky_audit_accuracy": 0.50390625,
          "leaky_reported_test_accuracy": 0.51953125,
          "leaky_vs_honest_test_gap": 0.0,
          "optimism_gap": 0.015625
        },
        "honest": {
          "audit_correct": 516,
          "audit_total": 1024,
          "feature_index": 1,
          "polarity": 1,
          "selection_correct": 136,
          "selection_total": 256,
          "test_correct": 133,
          "test_total": 256
        },
        "k": 2,
        "leaky": {
          "audit_correct": 516,
          "audit_total": 1024,
          "feature_index": 1,
          "polarity": 1,
          "test_correct": 133,
          "test_total": 256
        },
        "seed": 15
      },
      {
        "candidate_pair_count": 8,
        "derived": {
          "honest_audit_accuracy": 0.50390625,
          "honest_selection_accuracy": 0.53125,
          "honest_test_accuracy": 0.51953125,
          "leaky_audit_accuracy": 0.4951171875,
          "leaky_reported_test_accuracy": 0.546875,
          "leaky_vs_honest_test_gap": 0.02734375,
          "optimism_gap": 0.0517578125
        },
        "honest": {
          "audit_correct": 516,
          "audit_total": 1024,
          "feature_index": 1,
          "polarity": 1,
          "selection_correct": 136,
          "selection_total": 256,
          "test_correct": 133,
          "test_total": 256
        },
        "k": 4,
        "leaky": {
          "audit_correct": 507,
          "audit_total": 1024,
          "feature_index": 3,
          "polarity": -1,
          "test_correct": 140,
          "test_total": 256
        },
        "seed": 15
      },
      {
        "candidate_pair_count": 16,
        "derived": {
          "honest_audit_accuracy": 0.5,
          "honest_selection_accuracy": 0.5546875,
          "honest_test_accuracy": 0.4921875,
          "leaky_audit_accuracy": 0.505859375,
          "leaky_reported_test_accuracy": 0.55078125,
          "leaky_vs_honest_test_gap": 0.05859375,
          "optimism_gap": 0.044921875
        },
        "honest": {
          "audit_correct": 512,
          "audit_total": 1024,
          "feature_index": 7,
          "polarity": 1,
          "selection_correct": 142,
          "selection_total": 256,
          "test_correct": 126,
          "test_total": 256
        },
        "k": 8,
        "leaky": {
          "audit_correct": 518,
          "audit_total": 1024,
          "feature_index": 4,
          "polarity": -1,
          "test_correct": 141,
          "test_total": 256
        },
        "seed": 15
      },
      {
        "candidate_pair_count": 32,
        "derived": {
          "honest_audit_accuracy": 0.5,
          "honest_selection_accuracy": 0.5546875,
          "honest_test_accuracy": 0.4921875,
          "leaky_audit_accuracy": 0.4736328125,
          "leaky_reported_test_accuracy": 0.56640625,
          "leaky_vs_honest_test_gap": 0.07421875,
          "optimism_gap": 0.0927734375
        },
        "honest": {
          "audit_correct": 512,
          "audit_total": 1024,
          "feature_index": 7,
          "polarity": 1,
          "selection_correct": 142,
          "selection_total": 256,
          "test_correct": 126,
          "test_total": 256
        },
        "k": 16,
        "leaky": {
          "audit_correct": 485,
          "audit_total": 1024,
          "feature_index": 9,
          "polarity": 1,
          "test_correct": 145,
          "test_total": 256
        },
        "seed": 15
      },
      {
        "candidate_pair_count": 64,
        "derived": {
          "honest_audit_accuracy": 0.490234375,
          "honest_selection_accuracy": 0.55859375,
          "honest_test_accuracy": 0.44921875,
          "leaky_audit_accuracy": 0.4736328125,
          "leaky_reported_test_accuracy": 0.56640625,
          "leaky_vs_honest_test_gap": 0.1171875,
          "optimism_gap": 0.0927734375
        },
        "honest": {
          "audit_correct": 502,
          "audit_total": 1024,
          "feature_index": 21,
          "polarity": 1,
          "selection_correct": 143,
          "selection_total": 256,
          "test_correct": 115,
          "test_total": 256
        },
        "k": 32,
        "leaky": {
          "audit_correct": 485,
          "audit_total": 1024,
          "feature_index": 9,
          "polarity": 1,
          "test_correct": 145,
          "test_total": 256
        },
        "seed": 15
      },
      {
        "candidate_pair_count": 128,
        "derived": {
          "honest_audit_accuracy": 0.490234375,
          "honest_selection_accuracy": 0.55859375,
          "honest_test_accuracy": 0.44921875,
          "leaky_audit_accuracy": 0.4736328125,
          "leaky_reported_test_accuracy": 0.56640625,
          "leaky_vs_honest_test_gap": 0.1171875,
          "optimism_gap": 0.0927734375
        },
        "honest": {
          "audit_correct": 502,
          "audit_total": 1024,
          "feature_index": 21,
          "polarity": 1,
          "selection_correct": 143,
          "selection_total": 256,
          "test_correct": 115,
          "test_total": 256
        },
        "k": 64,
        "leaky": {
          "audit_correct": 485,
          "audit_total": 1024,
          "feature_index": 9,
          "polarity": 1,
          "test_correct": 145,
          "test_total": 256
        },
        "seed": 15
      },
      {
        "candidate_pair_count": 256,
        "derived": {
          "honest_audit_accuracy": 0.51171875,
          "honest_selection_accuracy": 0.57421875,
          "honest_test_accuracy": 0.5078125,
          "leaky_audit_accuracy": 0.4990234375,
          "leaky_reported_test_accuracy": 0.58203125,
          "leaky_vs_honest_test_gap": 0.07421875,
          "optimism_gap": 0.0830078125
        },
        "honest": {
          "audit_correct": 524,
          "audit_total": 1024,
          "feature_index": 64,
          "polarity": 1,
          "selection_correct": 147,
          "selection_total": 256,
          "test_correct": 130,
          "test_total": 256
        },
        "k": 128,
        "leaky": {
          "audit_correct": 511,
          "audit_total": 1024,
          "feature_index": 89,
          "polarity": 1,
          "test_correct": 149,
          "test_total": 256
        },
        "seed": 15
      },
      {
        "candidate_pair_count": 512,
        "derived": {
          "honest_audit_accuracy": 0.4873046875,
          "honest_selection_accuracy": 0.58984375,
          "honest_test_accuracy": 0.50390625,
          "leaky_audit_accuracy": 0.501953125,
          "leaky_reported_test_accuracy": 0.5859375,
          "leaky_vs_honest_test_gap": 0.08203125,
          "optimism_gap": 0.083984375
        },
        "honest": {
          "audit_correct": 499,
          "audit_total": 1024,
          "feature_index": 164,
          "polarity": 1,
          "selection_correct": 151,
          "selection_total": 256,
          "test_correct": 129,
          "test_total": 256
        },
        "k": 256,
        "leaky": {
          "audit_correct": 514,
          "audit_total": 1024,
          "feature_index": 130,
          "polarity": -1,
          "test_correct": 150,
          "test_total": 256
        },
        "seed": 15
      },
      {
        "candidate_pair_count": 1024,
        "derived": {
          "honest_audit_accuracy": 0.4873046875,
          "honest_selection_accuracy": 0.58984375,
          "honest_test_accuracy": 0.50390625,
          "leaky_audit_accuracy": 0.50390625,
          "leaky_reported_test_accuracy": 0.59375,
          "leaky_vs_honest_test_gap": 0.08984375,
          "optimism_gap": 0.08984375
        },
        "honest": {
          "audit_correct": 499,
          "audit_total": 1024,
          "feature_index": 164,
          "polarity": 1,
          "selection_correct": 151,
          "selection_total": 256,
          "test_correct": 129,
          "test_total": 256
        },
        "k": 512,
        "leaky": {
          "audit_correct": 516,
          "audit_total": 1024,
          "feature_index": 507,
          "polarity": 1,
          "test_correct": 152,
          "test_total": 256
        },
        "seed": 15
      },
      {
        "candidate_pair_count": 2048,
        "derived": {
          "honest_audit_accuracy": 0.529296875,
          "honest_selection_accuracy": 0.6171875,
          "honest_test_accuracy": 0.5234375,
          "leaky_audit_accuracy": 0.490234375,
          "leaky_reported_test_accuracy": 0.60546875,
          "leaky_vs_honest_test_gap": 0.08203125,
          "optimism_gap": 0.115234375
        },
        "honest": {
          "audit_correct": 542,
          "audit_total": 1024,
          "feature_index": 806,
          "polarity": 1,
          "selection_correct": 158,
          "selection_total": 256,
          "test_correct": 134,
          "test_total": 256
        },
        "k": 1024,
        "leaky": {
          "audit_correct": 502,
          "audit_total": 1024,
          "feature_index": 995,
          "polarity": 1,
          "test_correct": 155,
          "test_total": 256
        },
        "seed": 15
      },
      {
        "candidate_pair_count": 2,
        "derived": {
          "honest_audit_accuracy": 0.4970703125,
          "honest_selection_accuracy": 0.52734375,
          "honest_test_accuracy": 0.44140625,
          "leaky_audit_accuracy": 0.5029296875,
          "leaky_reported_test_accuracy": 0.55859375,
          "leaky_vs_honest_test_gap": 0.1171875,
          "optimism_gap": 0.0556640625
        },
        "honest": {
          "audit_correct": 509,
          "audit_total": 1024,
          "feature_index": 0,
          "polarity": 1,
          "selection_correct": 135,
          "selection_total": 256,
          "test_correct": 113,
          "test_total": 256
        },
        "k": 1,
        "leaky": {
          "audit_correct": 515,
          "audit_total": 1024,
          "feature_index": 0,
          "polarity": -1,
          "test_correct": 143,
          "test_total": 256
        },
        "seed": 16
      },
      {
        "candidate_pair_count": 4,
        "derived": {
          "honest_audit_accuracy": 0.48828125,
          "honest_selection_accuracy": 0.53515625,
          "honest_test_accuracy": 0.53515625,
          "leaky_audit_accuracy": 0.5029296875,
          "leaky_reported_test_accuracy": 0.55859375,
          "leaky_vs_honest_test_gap": 0.0234375,
          "optimism_gap": 0.0556640625
        },
        "honest": {
          "audit_correct": 500,
          "audit_total": 1024,
          "feature_index": 1,
          "polarity": 1,
          "selection_correct": 137,
          "selection_total": 256,
          "test_correct": 137,
          "test_total": 256
        },
        "k": 2,
        "leaky": {
          "audit_correct": 515,
          "audit_total": 1024,
          "feature_index": 0,
          "polarity": -1,
          "test_correct": 143,
          "test_total": 256
        },
        "seed": 16
      },
      {
        "candidate_pair_count": 8,
        "derived": {
          "honest_audit_accuracy": 0.48828125,
          "honest_selection_accuracy": 0.53515625,
          "honest_test_accuracy": 0.53515625,
          "leaky_audit_accuracy": 0.5283203125,
          "leaky_reported_test_accuracy": 0.578125,
          "leaky_vs_honest_test_gap": 0.04296875,
          "optimism_gap": 0.0498046875
        },
        "honest": {
          "audit_correct": 500,
          "audit_total": 1024,
          "feature_index": 1,
          "polarity": 1,
          "selection_correct": 137,
          "selection_total": 256,
          "test_correct": 137,
          "test_total": 256
        },
        "k": 4,
        "leaky": {
          "audit_correct": 541,
          "audit_total": 1024,
          "feature_index": 3,
          "polarity": -1,
          "test_correct": 148,
          "test_total": 256
        },
        "seed": 16
      },
      {
        "candidate_pair_count": 16,
        "derived": {
          "honest_audit_accuracy": 0.48828125,
          "honest_selection_accuracy": 0.53515625,
          "honest_test_accuracy": 0.53515625,
          "leaky_audit_accuracy": 0.5283203125,
          "leaky_reported_test_accuracy": 0.578125,
          "leaky_vs_honest_test_gap": 0.04296875,
          "optimism_gap": 0.0498046875
        },
        "honest": {
          "audit_correct": 500,
          "audit_total": 1024,
          "feature_index": 1,
          "polarity": 1,
          "selection_correct": 137,
          "selection_total": 256,
          "test_correct": 137,
          "test_total": 256
        },
        "k": 8,
        "leaky": {
          "audit_correct": 541,
          "audit_total": 1024,
          "feature_index": 3,
          "polarity": -1,
          "test_correct": 148,
          "test_total": 256
        },
        "seed": 16
      },
      {
        "candidate_pair_count": 32,
        "derived": {
          "honest_audit_accuracy": 0.5126953125,
          "honest_selection_accuracy": 0.5546875,
          "honest_test_accuracy": 0.55078125,
          "leaky_audit_accuracy": 0.5283203125,
          "leaky_reported_test_accuracy": 0.578125,
          "leaky_vs_honest_test_gap": 0.02734375,
          "optimism_gap": 0.0498046875
        },
        "honest": {
          "audit_correct": 525,
          "audit_total": 1024,
          "feature_index": 15,
          "polarity": -1,
          "selection_correct": 142,
          "selection_total": 256,
          "test_correct": 141,
          "test_total": 256
        },
        "k": 16,
        "leaky": {
          "audit_correct": 541,
          "audit_total": 1024,
          "feature_index": 3,
          "polarity": -1,
          "test_correct": 148,
          "test_total": 256
        },
        "seed": 16
      },
      {
        "candidate_pair_count": 64,
        "derived": {
          "honest_audit_accuracy": 0.513671875,
          "honest_selection_accuracy": 0.59375,
          "honest_test_accuracy": 0.4765625,
          "leaky_audit_accuracy": 0.4931640625,
          "leaky_reported_test_accuracy": 0.58984375,
          "leaky_vs_honest_test_gap": 0.11328125,
          "optimism_gap": 0.0966796875
        },
        "honest": {
          "audit_correct": 526,
          "audit_total": 1024,
          "feature_index": 18,
          "polarity": 1,
          "selection_correct": 152,
          "selection_total": 256,
          "test_correct": 122,
          "test_total": 256
        },
        "k": 32,
        "leaky": {
          "audit_correct": 505,
          "audit_total": 1024,
          "feature_index": 25,
          "polarity": 1,
          "test_correct": 151,
          "test_total": 256
        },
        "seed": 16
      },
      {
        "candidate_pair_count": 128,
        "derived": {
          "honest_audit_accuracy": 0.513671875,
          "honest_selection_accuracy": 0.59375,
          "honest_test_accuracy": 0.4765625,
          "leaky_audit_accuracy": 0.4931640625,
          "leaky_reported_test_accuracy": 0.58984375,
          "leaky_vs_honest_test_gap": 0.11328125,
          "optimism_gap": 0.0966796875
        },
        "honest": {
          "audit_correct": 526,
          "audit_total": 1024,
          "feature_index": 18,
          "polarity": 1,
          "selection_correct": 152,
          "selection_total": 256,
          "test_correct": 122,
          "test_total": 256
        },
        "k": 64,
        "leaky": {
          "audit_correct": 505,
          "audit_total": 1024,
          "feature_index": 25,
          "polarity": 1,
          "test_correct": 151,
          "test_total": 256
        },
        "seed": 16
      },
      {
        "candidate_pair_count": 256,
        "derived": {
          "honest_audit_accuracy": 0.513671875,
          "honest_selection_accuracy": 0.59375,
          "honest_test_accuracy": 0.4765625,
          "leaky_audit_accuracy": 0.4931640625,
          "leaky_reported_test_accuracy": 0.58984375,
          "leaky_vs_honest_test_gap": 0.11328125,
          "optimism_gap": 0.0966796875
        },
        "honest": {
          "audit_correct": 526,
          "audit_total": 1024,
          "feature_index": 18,
          "polarity": 1,
          "selection_correct": 152,
          "selection_total": 256,
          "test_correct": 122,
          "test_total": 256
        },
        "k": 128,
        "leaky": {
          "audit_correct": 505,
          "audit_total": 1024,
          "feature_index": 25,
          "polarity": 1,
          "test_correct": 151,
          "test_total": 256
        },
        "seed": 16
      },
      {
        "candidate_pair_count": 512,
        "derived": {
          "honest_audit_accuracy": 0.513671875,
          "honest_selection_accuracy": 0.59375,
          "honest_test_accuracy": 0.4765625,
          "leaky_audit_accuracy": 0.4931640625,
          "leaky_reported_test_accuracy": 0.58984375,
          "leaky_vs_honest_test_gap": 0.11328125,
          "optimism_gap": 0.0966796875
        },
        "honest": {
          "audit_correct": 526,
          "audit_total": 1024,
          "feature_index": 18,
          "polarity": 1,
          "selection_correct": 152,
          "selection_total": 256,
          "test_correct": 122,
          "test_total": 256
        },
        "k": 256,
        "leaky": {
          "audit_correct": 505,
          "audit_total": 1024,
          "feature_index": 25,
          "polarity": 1,
          "test_correct": 151,
          "test_total": 256
        },
        "seed": 16
      },
      {
        "candidate_pair_count": 1024,
        "derived": {
          "honest_audit_accuracy": 0.513671875,
          "honest_selection_accuracy": 0.59375,
          "honest_test_accuracy": 0.4765625,
          "leaky_audit_accuracy": 0.5029296875,
          "leaky_reported_test_accuracy": 0.6015625,
          "leaky_vs_honest_test_gap": 0.125,
          "optimism_gap": 0.0986328125
        },
        "honest": {
          "audit_correct": 526,
          "audit_total": 1024,
          "feature_index": 18,
          "polarity": 1,
          "selection_correct": 152,
          "selection_total": 256,
          "test_correct": 122,
          "test_total": 256
        },
        "k": 512,
        "leaky": {
          "audit_correct": 515,
          "audit_total": 1024,
          "feature_index": 509,
          "polarity": -1,
          "test_correct": 154,
          "test_total": 256
        },
        "seed": 16
      },
      {
        "candidate_pair_count": 2048,
        "derived": {
          "honest_audit_accuracy": 0.513671875,
          "honest_selection_accuracy": 0.59375,
          "honest_test_accuracy": 0.4765625,
          "leaky_audit_accuracy": 0.5146484375,
          "leaky_reported_test_accuracy": 0.60546875,
          "leaky_vs_honest_test_gap": 0.12890625,
          "optimism_gap": 0.0908203125
        },
        "honest": {
          "audit_correct": 526,
          "audit_total": 1024,
          "feature_index": 18,
          "polarity": 1,
          "selection_correct": 152,
          "selection_total": 256,
          "test_correct": 122,
          "test_total": 256
        },
        "k": 1024,
        "leaky": {
          "audit_correct": 527,
          "audit_total": 1024,
          "feature_index": 879,
          "polarity": -1,
          "test_correct": 155,
          "test_total": 256
        },
        "seed": 16
      },
      {
        "candidate_pair_count": 2,
        "derived": {
          "honest_audit_accuracy": 0.5078125,
          "honest_selection_accuracy": 0.5,
          "honest_test_accuracy": 0.50390625,
          "leaky_audit_accuracy": 0.5078125,
          "leaky_reported_test_accuracy": 0.50390625,
          "leaky_vs_honest_test_gap": 0.0,
          "optimism_gap": -0.00390625
        },
        "honest": {
          "audit_correct": 520,
          "audit_total": 1024,
          "feature_index": 0,
          "polarity": 1,
          "selection_correct": 128,
          "selection_total": 256,
          "test_correct": 129,
          "test_total": 256
        },
        "k": 1,
        "leaky": {
          "audit_correct": 520,
          "audit_total": 1024,
          "feature_index": 0,
          "polarity": 1,
          "test_correct": 129,
          "test_total": 256
        },
        "seed": 17
      },
      {
        "candidate_pair_count": 4,
        "derived": {
          "honest_audit_accuracy": 0.5048828125,
          "honest_selection_accuracy": 0.55078125,
          "honest_test_accuracy": 0.5078125,
          "leaky_audit_accuracy": 0.5048828125,
          "leaky_reported_test_accuracy": 0.5078125,
          "leaky_vs_honest_test_gap": 0.0,
          "optimism_gap": 0.0029296875
        },
        "honest": {
          "audit_correct": 517,
          "audit_total": 1024,
          "feature_index": 1,
          "polarity": -1,
          "selection_correct": 141,
          "selection_total": 256,
          "test_correct": 130,
          "test_total": 256
        },
        "k": 2,
        "leaky": {
          "audit_correct": 517,
          "audit_total": 1024,
          "feature_index": 1,
          "polarity": -1,
          "test_correct": 130,
          "test_total": 256
        },
        "seed": 17
      },
      {
        "candidate_pair_count": 8,
        "derived": {
          "honest_audit_accuracy": 0.513671875,
          "honest_selection_accuracy": 0.5546875,
          "honest_test_accuracy": 0.5234375,
          "leaky_audit_accuracy": 0.513671875,
          "leaky_reported_test_accuracy": 0.5234375,
          "leaky_vs_honest_test_gap": 0.0,
          "optimism_gap": 0.009765625
        },
        "honest": {
          "audit_correct": 526,
          "audit_total": 1024,
          "feature_index": 2,
          "polarity": 1,
          "selection_correct": 142,
          "selection_total": 256,
          "test_correct": 134,
          "test_total": 256
        },
        "k": 4,
        "leaky": {
          "audit_correct": 526,
          "audit_total": 1024,
          "feature_index": 2,
          "polarity": 1,
          "test_correct": 134,
          "test_total": 256
        },
        "seed": 17
      },
      {
        "candidate_pair_count": 16,
        "derived": {
          "honest_audit_accuracy": 0.513671875,
          "honest_selection_accuracy": 0.5546875,
          "honest_test_accuracy": 0.5234375,
          "leaky_audit_accuracy": 0.4638671875,
          "leaky_reported_test_accuracy": 0.546875,
          "leaky_vs_honest_test_gap": 0.0234375,
          "optimism_gap": 0.0830078125
        },
        "honest": {
          "audit_correct": 526,
          "audit_total": 1024,
          "feature_index": 2,
          "polarity": 1,
          "selection_correct": 142,
          "selection_total": 256,
          "test_correct": 134,
          "test_total": 256
        },
        "k": 8,
        "leaky": {
          "audit_correct": 475,
          "audit_total": 1024,
          "feature_index": 6,
          "polarity": -1,
          "test_correct": 140,
          "test_total": 256
        },
        "seed": 17
      },
      {
        "candidate_pair_count": 32,
        "derived": {
          "honest_audit_accuracy": 0.513671875,
          "honest_selection_accuracy": 0.5546875,
          "honest_test_accuracy": 0.5234375,
          "leaky_audit_accuracy": 0.4638671875,
          "leaky_reported_test_accuracy": 0.546875,
          "leaky_vs_honest_test_gap": 0.0234375,
          "optimism_gap": 0.0830078125
        },
        "honest": {
          "audit_correct": 526,
          "audit_total": 1024,
          "feature_index": 2,
          "polarity": 1,
          "selection_correct": 142,
          "selection_total": 256,
          "test_correct": 134,
          "test_total": 256
        },
        "k": 16,
        "leaky": {
          "audit_correct": 475,
          "audit_total": 1024,
          "feature_index": 6,
          "polarity": -1,
          "test_correct": 140,
          "test_total": 256
        },
        "seed": 17
      },
      {
        "candidate_pair_count": 64,
        "derived": {
          "honest_audit_accuracy": 0.48046875,
          "honest_selection_accuracy": 0.55859375,
          "honest_test_accuracy": 0.54296875,
          "leaky_audit_accuracy": 0.478515625,
          "leaky_reported_test_accuracy": 0.5625,
          "leaky_vs_honest_test_gap": 0.01953125,
          "optimism_gap": 0.083984375
        },
        "honest": {
          "audit_correct": 492,
          "audit_total": 1024,
          "feature_index": 22,
          "polarity": 1,
          "selection_correct": 143,
          "selection_total": 256,
          "test_correct": 139,
          "test_total": 256
        },
        "k": 32,
        "leaky": {
          "audit_correct": 490,
          "audit_total": 1024,
          "feature_index": 19,
          "polarity": -1,
          "test_correct": 144,
          "test_total": 256
        },
        "seed": 17
      },
      {
        "candidate_pair_count": 128,
        "derived": {
          "honest_audit_accuracy": 0.49609375,
          "honest_selection_accuracy": 0.5859375,
          "honest_test_accuracy": 0.45703125,
          "leaky_audit_accuracy": 0.505859375,
          "leaky_reported_test_accuracy": 0.5703125,
          "leaky_vs_honest_test_gap": 0.11328125,
          "optimism_gap": 0.064453125
        },
        "honest": {
          "audit_correct": 508,
          "audit_total": 1024,
          "feature_index": 52,
          "polarity": -1,
          "selection_correct": 150,
          "selection_total": 256,
          "test_correct": 117,
          "test_total": 256
        },
        "k": 64,
        "leaky": {
          "audit_correct": 518,
          "audit_total": 1024,
          "feature_index": 58,
          "polarity": 1,
          "test_correct": 146,
          "test_total": 256
        },
        "seed": 17
      },
      {
        "candidate_pair_count": 256,
        "derived": {
          "honest_audit_accuracy": 0.49609375,
          "honest_selection_accuracy": 0.5859375,
          "honest_test_accuracy": 0.45703125,
          "leaky_audit_accuracy": 0.478515625,
          "leaky_reported_test_accuracy": 0.58203125,
          "leaky_vs_honest_test_gap": 0.125,
          "optimism_gap": 0.103515625
        },
        "honest": {
          "audit_correct": 508,
          "audit_total": 1024,
          "feature_index": 52,
          "polarity": -1,
          "selection_correct": 150,
          "selection_total": 256,
          "test_correct": 117,
          "test_total": 256
        },
        "k": 128,
        "leaky": {
          "audit_correct": 490,
          "audit_total": 1024,
          "feature_index": 121,
          "polarity": 1,
          "test_correct": 149,
          "test_total": 256
        },
        "seed": 17
      },
      {
        "candidate_pair_count": 512,
        "derived": {
          "honest_audit_accuracy": 0.4765625,
          "honest_selection_accuracy": 0.60546875,
          "honest_test_accuracy": 0.51953125,
          "leaky_audit_accuracy": 0.521484375,
          "leaky_reported_test_accuracy": 0.59765625,
          "leaky_vs_honest_test_gap": 0.078125,
          "optimism_gap": 0.076171875
        },
        "honest": {
          "audit_correct": 488,
          "audit_total": 1024,
          "feature_index": 194,
          "polarity": -1,
          "selection_correct": 155,
          "selection_total": 256,
          "test_correct": 133,
          "test_total": 256
        },
        "k": 256,
        "leaky": {
          "audit_correct": 534,
          "audit_total": 1024,
          "feature_index": 216,
          "polarity": 1,
          "test_correct": 153,
          "test_total": 256
        },
        "seed": 17
      },
      {
        "candidate_pair_count": 1024,
        "derived": {
          "honest_audit_accuracy": 0.4765625,
          "honest_selection_accuracy": 0.60546875,
          "honest_test_accuracy": 0.51953125,
          "leaky_audit_accuracy": 0.521484375,
          "leaky_reported_test_accuracy": 0.59765625,
          "leaky_vs_honest_test_gap": 0.078125,
          "optimism_gap": 0.076171875
        },
        "honest": {
          "audit_correct": 488,
          "audit_total": 1024,
          "feature_index": 194,
          "polarity": -1,
          "selection_correct": 155,
          "selection_total": 256,
          "test_correct": 133,
          "test_total": 256
        },
        "k": 512,
        "leaky": {
          "audit_correct": 534,
          "audit_total": 1024,
          "feature_index": 216,
          "polarity": 1,
          "test_correct": 153,
          "test_total": 256
        },
        "seed": 17
      },
      {
        "candidate_pair_count": 2048,
        "derived": {
          "honest_audit_accuracy": 0.4931640625,
          "honest_selection_accuracy": 0.609375,
          "honest_test_accuracy": 0.453125,
          "leaky_audit_accuracy": 0.49609375,
          "leaky_reported_test_accuracy": 0.609375,
          "leaky_vs_honest_test_gap": 0.15625,
          "optimism_gap": 0.11328125
        },
        "honest": {
          "audit_correct": 505,
          "audit_total": 1024,
          "feature_index": 749,
          "polarity": 1,
          "selection_correct": 156,
          "selection_total": 256,
          "test_correct": 116,
          "test_total": 256
        },
        "k": 1024,
        "leaky": {
          "audit_correct": 508,
          "audit_total": 1024,
          "feature_index": 520,
          "polarity": -1,
          "test_correct": 156,
          "test_total": 256
        },
        "seed": 17
      },
      {
        "candidate_pair_count": 2,
        "derived": {
          "honest_audit_accuracy": 0.515625,
          "honest_selection_accuracy": 0.515625,
          "honest_test_accuracy": 0.4765625,
          "leaky_audit_accuracy": 0.484375,
          "leaky_reported_test_accuracy": 0.5234375,
          "leaky_vs_honest_test_gap": 0.046875,
          "optimism_gap": 0.0390625
        },
        "honest": {
          "audit_correct": 528,
          "audit_total": 1024,
          "feature_index": 0,
          "polarity": -1,
          "selection_correct": 132,
          "selection_total": 256,
          "test_correct": 122,
          "test_total": 256
        },
        "k": 1,
        "leaky": {
          "audit_correct": 496,
          "audit_total": 1024,
          "feature_index": 0,
          "polarity": 1,
          "test_correct": 134,
          "test_total": 256
        },
        "seed": 18
      },
      {
        "candidate_pair_count": 4,
        "derived": {
          "honest_audit_accuracy": 0.515625,
          "honest_selection_accuracy": 0.515625,
          "honest_test_accuracy": 0.4765625,
          "leaky_audit_accuracy": 0.4921875,
          "leaky_reported_test_accuracy": 0.53515625,
          "leaky_vs_honest_test_gap": 0.05859375,
          "optimism_gap": 0.04296875
        },
        "honest": {
          "audit_correct": 528,
          "audit_total": 1024,
          "feature_index": 0,
          "polarity": -1,
          "selection_correct": 132,
          "selection_total": 256,
          "test_correct": 122,
          "test_total": 256
        },
        "k": 2,
        "leaky": {
          "audit_correct": 504,
          "audit_total": 1024,
          "feature_index": 1,
          "polarity": -1,
          "test_correct": 137,
          "test_total": 256
        },
        "seed": 18
      },
      {
        "candidate_pair_count": 8,
        "derived": {
          "honest_audit_accuracy": 0.4970703125,
          "honest_selection_accuracy": 0.56640625,
          "honest_test_accuracy": 0.47265625,
          "leaky_audit_accuracy": 0.4921875,
          "leaky_reported_test_accuracy": 0.53515625,
          "leaky_vs_honest_test_gap": 0.0625,
          "optimism_gap": 0.04296875
        },
        "honest": {
          "audit_correct": 509,
          "audit_total": 1024,
          "feature_index": 3,
          "polarity": -1,
          "selection_correct": 145,
          "selection_total": 256,
          "test_correct": 121,
          "test_total": 256
        },
        "k": 4,
        "leaky": {
          "audit_correct": 504,
          "audit_total": 1024,
          "feature_index": 1,
          "polarity": -1,
          "test_correct": 137,
          "test_total": 256
        },
        "seed": 18
      },
      {
        "candidate_pair_count": 16,
        "derived": {
          "honest_audit_accuracy": 0.4970703125,
          "honest_selection_accuracy": 0.56640625,
          "honest_test_accuracy": 0.47265625,
          "leaky_audit_accuracy": 0.521484375,
          "leaky_reported_test_accuracy": 0.546875,
          "leaky_vs_honest_test_gap": 0.07421875,
          "optimism_gap": 0.025390625
        },
        "honest": {
          "audit_correct": 509,
          "audit_total": 1024,
          "feature_index": 3,
          "polarity": -1,
          "selection_correct": 145,
          "selection_total": 256,
          "test_correct": 121,
          "test_total": 256
        },
        "k": 8,
        "leaky": {
          "audit_correct": 534,
          "audit_total": 1024,
          "feature_index": 6,
          "polarity": 1,
          "test_correct": 140,
          "test_total": 256
        },
        "seed": 18
      },
      {
        "candidate_pair_count": 32,
        "derived": {
          "honest_audit_accuracy": 0.4970703125,
          "honest_selection_accuracy": 0.56640625,
          "honest_test_accuracy": 0.47265625,
          "leaky_audit_accuracy": 0.470703125,
          "leaky_reported_test_accuracy": 0.5625,
          "leaky_vs_honest_test_gap": 0.08984375,
          "optimism_gap": 0.091796875
        },
        "honest": {
          "audit_correct": 509,
          "audit_total": 1024,
          "feature_index": 3,
          "polarity": -1,
          "selection_correct": 145,
          "selection_total": 256,
          "test_correct": 121,
          "test_total": 256
        },
        "k": 16,
        "leaky": {
          "audit_correct": 482,
          "audit_total": 1024,
          "feature_index": 8,
          "polarity": -1,
          "test_correct": 144,
          "test_total": 256
        },
        "seed": 18
      },
      {
        "candidate_pair_count": 64,
        "derived": {
          "honest_audit_accuracy": 0.4970703125,
          "honest_selection_accuracy": 0.56640625,
          "honest_test_accuracy": 0.47265625,
          "leaky_audit_accuracy": 0.470703125,
          "leaky_reported_test_accuracy": 0.5625,
          "leaky_vs_honest_test_gap": 0.08984375,
          "optimism_gap": 0.091796875
        },
        "honest": {
          "audit_correct": 509,
          "audit_total": 1024,
          "feature_index": 3,
          "polarity": -1,
          "selection_correct": 145,
          "selection_total": 256,
          "test_correct": 121,
          "test_total": 256
        },
        "k": 32,
        "leaky": {
          "audit_correct": 482,
          "audit_total": 1024,
          "feature_index": 8,
          "polarity": -1,
          "test_correct": 144,
          "test_total": 256
        },
        "seed": 18
      },
      {
        "candidate_pair_count": 128,
        "derived": {
          "honest_audit_accuracy": 0.4970703125,
          "honest_selection_accuracy": 0.56640625,
          "honest_test_accuracy": 0.47265625,
          "leaky_audit_accuracy": 0.470703125,
          "leaky_reported_test_accuracy": 0.5625,
          "leaky_vs_honest_test_gap": 0.08984375,
          "optimism_gap": 0.091796875
        },
        "honest": {
          "audit_correct": 509,
          "audit_total": 1024,
          "feature_index": 3,
          "polarity": -1,
          "selection_correct": 145,
          "selection_total": 256,
          "test_correct": 121,
          "test_total": 256
        },
        "k": 64,
        "leaky": {
          "audit_correct": 482,
          "audit_total": 1024,
          "feature_index": 8,
          "polarity": -1,
          "test_correct": 144,
          "test_total": 256
        },
        "seed": 18
      },
      {
        "candidate_pair_count": 256,
        "derived": {
          "honest_audit_accuracy": 0.4970703125,
          "honest_selection_accuracy": 0.56640625,
          "honest_test_accuracy": 0.47265625,
          "leaky_audit_accuracy": 0.4921875,
          "leaky_reported_test_accuracy": 0.60546875,
          "leaky_vs_honest_test_gap": 0.1328125,
          "optimism_gap": 0.11328125
        },
        "honest": {
          "audit_correct": 509,
          "audit_total": 1024,
          "feature_index": 3,
          "polarity": -1,
          "selection_correct": 145,
          "selection_total": 256,
          "test_correct": 121,
          "test_total": 256
        },
        "k": 128,
        "leaky": {
          "audit_correct": 504,
          "audit_total": 1024,
          "feature_index": 83,
          "polarity": 1,
          "test_correct": 155,
          "test_total": 256
        },
        "seed": 18
      },
      {
        "candidate_pair_count": 512,
        "derived": {
          "honest_audit_accuracy": 0.4931640625,
          "honest_selection_accuracy": 0.58203125,
          "honest_test_accuracy": 0.5234375,
          "leaky_audit_accuracy": 0.4921875,
          "leaky_reported_test_accuracy": 0.60546875,
          "leaky_vs_honest_test_gap": 0.08203125,
          "optimism_gap": 0.11328125
        },
        "honest": {
          "audit_correct": 505,
          "audit_total": 1024,
          "feature_index": 197,
          "polarity": -1,
          "selection_correct": 149,
          "selection_total": 256,
          "test_correct": 134,
          "test_total": 256
        },
        "k": 256,
        "leaky": {
          "audit_correct": 504,
          "audit_total": 1024,
          "feature_index": 83,
          "polarity": 1,
          "test_correct": 155,
          "test_total": 256
        },
        "seed": 18
      },
      {
        "candidate_pair_count": 1024,
        "derived": {
          "honest_audit_accuracy": 0.4970703125,
          "honest_selection_accuracy": 0.6171875,
          "honest_test_accuracy": 0.5,
          "leaky_audit_accuracy": 0.5029296875,
          "leaky_reported_test_accuracy": 0.61328125,
          "leaky_vs_honest_test_gap": 0.11328125,
          "optimism_gap": 0.1103515625
        },
        "honest": {
          "audit_correct": 509,
          "audit_total": 1024,
          "feature_index": 489,
          "polarity": -1,
          "selection_correct": 158,
          "selection_total": 256,
          "test_correct": 128,
          "test_total": 256
        },
        "k": 512,
        "leaky": {
          "audit_correct": 515,
          "audit_total": 1024,
          "feature_index": 406,
          "polarity": 1,
          "test_correct": 157,
          "test_total": 256
        },
        "seed": 18
      },
      {
        "candidate_pair_count": 2048,
        "derived": {
          "honest_audit_accuracy": 0.4970703125,
          "honest_selection_accuracy": 0.6171875,
          "honest_test_accuracy": 0.5,
          "leaky_audit_accuracy": 0.5029296875,
          "leaky_reported_test_accuracy": 0.61328125,
          "leaky_vs_honest_test_gap": 0.11328125,
          "optimism_gap": 0.1103515625
        },
        "honest": {
          "audit_correct": 509,
          "audit_total": 1024,
          "feature_index": 489,
          "polarity": -1,
          "selection_correct": 158,
          "selection_total": 256,
          "test_correct": 128,
          "test_total": 256
        },
        "k": 1024,
        "leaky": {
          "audit_correct": 515,
          "audit_total": 1024,
          "feature_index": 406,
          "polarity": 1,
          "test_correct": 157,
          "test_total": 256
        },
        "seed": 18
      },
      {
        "candidate_pair_count": 2,
        "derived": {
          "honest_audit_accuracy": 0.48046875,
          "honest_selection_accuracy": 0.5078125,
          "honest_test_accuracy": 0.4765625,
          "leaky_audit_accuracy": 0.51953125,
          "leaky_reported_test_accuracy": 0.5234375,
          "leaky_vs_honest_test_gap": 0.046875,
          "optimism_gap": 0.00390625
        },
        "honest": {
          "audit_correct": 492,
          "audit_total": 1024,
          "feature_index": 0,
          "polarity": -1,
          "selection_correct": 130,
          "selection_total": 256,
          "test_correct": 122,
          "test_total": 256
        },
        "k": 1,
        "leaky": {
          "audit_correct": 532,
          "audit_total": 1024,
          "feature_index": 0,
          "polarity": 1,
          "test_correct": 134,
          "test_total": 256
        },
        "seed": 19
      },
      {
        "candidate_pair_count": 4,
        "derived": {
          "honest_audit_accuracy": 0.48046875,
          "honest_selection_accuracy": 0.5078125,
          "honest_test_accuracy": 0.4765625,
          "leaky_audit_accuracy": 0.51953125,
          "leaky_reported_test_accuracy": 0.5234375,
          "leaky_vs_honest_test_gap": 0.046875,
          "optimism_gap": 0.00390625
        },
        "honest": {
          "audit_correct": 492,
          "audit_total": 1024,
          "feature_index": 0,
          "polarity": -1,
          "selection_correct": 130,
          "selection_total": 256,
          "test_correct": 122,
          "test_total": 256
        },
        "k": 2,
        "leaky": {
          "audit_correct": 532,
          "audit_total": 1024,
          "feature_index": 0,
          "polarity": 1,
          "test_correct": 134,
          "test_total": 256
        },
        "seed": 19
      },
      {
        "candidate_pair_count": 8,
        "derived": {
          "honest_audit_accuracy": 0.4931640625,
          "honest_selection_accuracy": 0.55078125,
          "honest_test_accuracy": 0.53125,
          "leaky_audit_accuracy": 0.4931640625,
          "leaky_reported_test_accuracy": 0.53125,
          "leaky_vs_honest_test_gap": 0.0,
          "optimism_gap": 0.0380859375
        },
        "honest": {
          "audit_correct": 505,
          "audit_total": 1024,
          "feature_index": 3,
          "polarity": -1,
          "selection_correct": 141,
          "selection_total": 256,
          "test_correct": 136,
          "test_total": 256
        },
        "k": 4,
        "leaky": {
          "audit_correct": 505,
          "audit_total": 1024,
          "feature_index": 3,
          "polarity": -1,
          "test_correct": 136,
          "test_total": 256
        },
        "seed": 19
      },
      {
        "candidate_pair_count": 16,
        "derived": {
          "honest_audit_accuracy": 0.48046875,
          "honest_selection_accuracy": 0.5703125,
          "honest_test_accuracy": 0.4609375,
          "leaky_audit_accuracy": 0.51953125,
          "leaky_reported_test_accuracy": 0.5390625,
          "leaky_vs_honest_test_gap": 0.078125,
          "optimism_gap": 0.01953125
        },
        "honest": {
          "audit_correct": 492,
          "audit_total": 1024,
          "feature_index": 6,
          "polarity": 1,
          "selection_correct": 146,
          "selection_total": 256,
          "test_correct": 118,
          "test_total": 256
        },
        "k": 8,
        "leaky": {
          "audit_correct": 532,
          "audit_total": 1024,
          "feature_index": 6,
          "polarity": -1,
          "test_correct": 138,
          "test_total": 256
        },
        "seed": 19
      },
      {
        "candidate_pair_count": 32,
        "derived": {
          "honest_audit_accuracy": 0.4990234375,
          "honest_selection_accuracy": 0.5859375,
          "honest_test_accuracy": 0.5546875,
          "leaky_audit_accuracy": 0.4990234375,
          "leaky_reported_test_accuracy": 0.5546875,
          "leaky_vs_honest_test_gap": 0.0,
          "optimism_gap": 0.0556640625
        },
        "honest": {
          "audit_correct": 511,
          "audit_total": 1024,
          "feature_index": 10,
          "polarity": -1,
          "selection_correct": 150,
          "selection_total": 256,
          "test_correct": 142,
          "test_total": 256
        },
        "k": 16,
        "leaky": {
          "audit_correct": 511,
          "audit_total": 1024,
          "feature_index": 10,
          "polarity": -1,
          "test_correct": 142,
          "test_total": 256
        },
        "seed": 19
      },
      {
        "candidate_pair_count": 64,
        "derived": {
          "honest_audit_accuracy": 0.4990234375,
          "honest_selection_accuracy": 0.5859375,
          "honest_test_accuracy": 0.5546875,
          "leaky_audit_accuracy": 0.517578125,
          "leaky_reported_test_accuracy": 0.5625,
          "leaky_vs_honest_test_gap": 0.0078125,
          "optimism_gap": 0.044921875
        },
        "honest": {
          "audit_correct": 511,
          "audit_total": 1024,
          "feature_index": 10,
          "polarity": -1,
          "selection_correct": 150,
          "selection_total": 256,
          "test_correct": 142,
          "test_total": 256
        },
        "k": 32,
        "leaky": {
          "audit_correct": 530,
          "audit_total": 1024,
          "feature_index": 25,
          "polarity": 1,
          "test_correct": 144,
          "test_total": 256
        },
        "seed": 19
      },
      {
        "candidate_pair_count": 128,
        "derived": {
          "honest_audit_accuracy": 0.4990234375,
          "honest_selection_accuracy": 0.5859375,
          "honest_test_accuracy": 0.5546875,
          "leaky_audit_accuracy": 0.49609375,
          "leaky_reported_test_accuracy": 0.57421875,
          "leaky_vs_honest_test_gap": 0.01953125,
          "optimism_gap": 0.078125
        },
        "honest": {
          "audit_correct": 511,
          "audit_total": 1024,
          "feature_index": 10,
          "polarity": -1,
          "selection_correct": 150,
          "selection_total": 256,
          "test_correct": 142,
          "test_total": 256
        },
        "k": 64,
        "leaky": {
          "audit_correct": 508,
          "audit_total": 1024,
          "feature_index": 49,
          "polarity": 1,
          "test_correct": 147,
          "test_total": 256
        },
        "seed": 19
      },
      {
        "candidate_pair_count": 256,
        "derived": {
          "honest_audit_accuracy": 0.517578125,
          "honest_selection_accuracy": 0.59375,
          "honest_test_accuracy": 0.4921875,
          "leaky_audit_accuracy": 0.5029296875,
          "leaky_reported_test_accuracy": 0.58203125,
          "leaky_vs_honest_test_gap": 0.08984375,
          "optimism_gap": 0.0791015625
        },
        "honest": {
          "audit_correct": 530,
          "audit_total": 1024,
          "feature_index": 80,
          "polarity": 1,
          "selection_correct": 152,
          "selection_total": 256,
          "test_correct": 126,
          "test_total": 256
        },
        "k": 128,
        "leaky": {
          "audit_correct": 515,
          "audit_total": 1024,
          "feature_index": 104,
          "polarity": -1,
          "test_correct": 149,
          "test_total": 256
        },
        "seed": 19
      },
      {
        "candidate_pair_count": 512,
        "derived": {
          "honest_audit_accuracy": 0.517578125,
          "honest_selection_accuracy": 0.59375,
          "honest_test_accuracy": 0.4921875,
          "leaky_audit_accuracy": 0.5029296875,
          "leaky_reported_test_accuracy": 0.640625,
          "leaky_vs_honest_test_gap": 0.1484375,
          "optimism_gap": 0.1376953125
        },
        "honest": {
          "audit_correct": 530,
          "audit_total": 1024,
          "feature_index": 80,
          "polarity": 1,
          "selection_correct": 152,
          "selection_total": 256,
          "test_correct": 126,
          "test_total": 256
        },
        "k": 256,
        "leaky": {
          "audit_correct": 515,
          "audit_total": 1024,
          "feature_index": 190,
          "polarity": 1,
          "test_correct": 164,
          "test_total": 256
        },
        "seed": 19
      },
      {
        "candidate_pair_count": 1024,
        "derived": {
          "honest_audit_accuracy": 0.517578125,
          "honest_selection_accuracy": 0.59375,
          "honest_test_accuracy": 0.4921875,
          "leaky_audit_accuracy": 0.5029296875,
          "leaky_reported_test_accuracy": 0.640625,
          "leaky_vs_honest_test_gap": 0.1484375,
          "optimism_gap": 0.1376953125
        },
        "honest": {
          "audit_correct": 530,
          "audit_total": 1024,
          "feature_index": 80,
          "polarity": 1,
          "selection_correct": 152,
          "selection_total": 256,
          "test_correct": 126,
          "test_total": 256
        },
        "k": 512,
        "leaky": {
          "audit_correct": 515,
          "audit_total": 1024,
          "feature_index": 190,
          "polarity": 1,
          "test_correct": 164,
          "test_total": 256
        },
        "seed": 19
      },
      {
        "candidate_pair_count": 2048,
        "derived": {
          "honest_audit_accuracy": 0.4970703125,
          "honest_selection_accuracy": 0.6171875,
          "honest_test_accuracy": 0.55859375,
          "leaky_audit_accuracy": 0.5029296875,
          "leaky_reported_test_accuracy": 0.640625,
          "leaky_vs_honest_test_gap": 0.08203125,
          "optimism_gap": 0.1376953125
        },
        "honest": {
          "audit_correct": 509,
          "audit_total": 1024,
          "feature_index": 705,
          "polarity": -1,
          "selection_correct": 158,
          "selection_total": 256,
          "test_correct": 143,
          "test_total": 256
        },
        "k": 1024,
        "leaky": {
          "audit_correct": 515,
          "audit_total": 1024,
          "feature_index": 190,
          "polarity": 1,
          "test_correct": 164,
          "test_total": 256
        },
        "seed": 19
      },
      {
        "candidate_pair_count": 2,
        "derived": {
          "honest_audit_accuracy": 0.5185546875,
          "honest_selection_accuracy": 0.50390625,
          "honest_test_accuracy": 0.47265625,
          "leaky_audit_accuracy": 0.4814453125,
          "leaky_reported_test_accuracy": 0.52734375,
          "leaky_vs_honest_test_gap": 0.0546875,
          "optimism_gap": 0.0458984375
        },
        "honest": {
          "audit_correct": 531,
          "audit_total": 1024,
          "feature_index": 0,
          "polarity": 1,
          "selection_correct": 129,
          "selection_total": 256,
          "test_correct": 121,
          "test_total": 256
        },
        "k": 1,
        "leaky": {
          "audit_correct": 493,
          "audit_total": 1024,
          "feature_index": 0,
          "polarity": -1,
          "test_correct": 135,
          "test_total": 256
        },
        "seed": 20
      },
      {
        "candidate_pair_count": 4,
        "derived": {
          "honest_audit_accuracy": 0.5068359375,
          "honest_selection_accuracy": 0.51953125,
          "honest_test_accuracy": 0.52734375,
          "leaky_audit_accuracy": 0.4814453125,
          "leaky_reported_test_accuracy": 0.52734375,
          "leaky_vs_honest_test_gap": 0.0,
          "optimism_gap": 0.0458984375
        },
        "honest": {
          "audit_correct": 519,
          "audit_total": 1024,
          "feature_index": 1,
          "polarity": 1,
          "selection_correct": 133,
          "selection_total": 256,
          "test_correct": 135,
          "test_total": 256
        },
        "k": 2,
        "leaky": {
          "audit_correct": 493,
          "audit_total": 1024,
          "feature_index": 0,
          "polarity": -1,
          "test_correct": 135,
          "test_total": 256
        },
        "seed": 20
      },
      {
        "candidate_pair_count": 8,
        "derived": {
          "honest_audit_accuracy": 0.5068359375,
          "honest_selection_accuracy": 0.51953125,
          "honest_test_accuracy": 0.52734375,
          "leaky_audit_accuracy": 0.4814453125,
          "leaky_reported_test_accuracy": 0.52734375,
          "leaky_vs_honest_test_gap": 0.0,
          "optimism_gap": 0.0458984375
        },
        "honest": {
          "audit_correct": 519,
          "audit_total": 1024,
          "feature_index": 1,
          "polarity": 1,
          "selection_correct": 133,
          "selection_total": 256,
          "test_correct": 135,
          "test_total": 256
        },
        "k": 4,
        "leaky": {
          "audit_correct": 493,
          "audit_total": 1024,
          "feature_index": 0,
          "polarity": -1,
          "test_correct": 135,
          "test_total": 256
        },
        "seed": 20
      },
      {
        "candidate_pair_count": 16,
        "derived": {
          "honest_audit_accuracy": 0.5107421875,
          "honest_selection_accuracy": 0.53125,
          "honest_test_accuracy": 0.44921875,
          "leaky_audit_accuracy": 0.4892578125,
          "leaky_reported_test_accuracy": 0.55078125,
          "leaky_vs_honest_test_gap": 0.1015625,
          "optimism_gap": 0.0615234375
        },
        "honest": {
          "audit_correct": 523,
          "audit_total": 1024,
          "feature_index": 6,
          "polarity": 1,
          "selection_correct": 136,
          "selection_total": 256,
          "test_correct": 115,
          "test_total": 256
        },
        "k": 8,
        "leaky": {
          "audit_correct": 501,
          "audit_total": 1024,
          "feature_index": 6,
          "polarity": -1,
          "test_correct": 141,
          "test_total": 256
        },
        "seed": 20
      },
      {
        "candidate_pair_count": 32,
        "derived": {
          "honest_audit_accuracy": 0.505859375,
          "honest_selection_accuracy": 0.56640625,
          "honest_test_accuracy": 0.4921875,
          "leaky_audit_accuracy": 0.5048828125,
          "leaky_reported_test_accuracy": 0.55859375,
          "leaky_vs_honest_test_gap": 0.06640625,
          "optimism_gap": 0.0537109375
        },
        "honest": {
          "audit_correct": 518,
          "audit_total": 1024,
          "feature_index": 14,
          "polarity": 1,
          "selection_correct": 145,
          "selection_total": 256,
          "test_correct": 126,
          "test_total": 256
        },
        "k": 16,
        "leaky": {
          "audit_correct": 517,
          "audit_total": 1024,
          "feature_index": 15,
          "polarity": -1,
          "test_correct": 143,
          "test_total": 256
        },
        "seed": 20
      },
      {
        "candidate_pair_count": 64,
        "derived": {
          "honest_audit_accuracy": 0.505859375,
          "honest_selection_accuracy": 0.56640625,
          "honest_test_accuracy": 0.4921875,
          "leaky_audit_accuracy": 0.5048828125,
          "leaky_reported_test_accuracy": 0.55859375,
          "leaky_vs_honest_test_gap": 0.06640625,
          "optimism_gap": 0.0537109375
        },
        "honest": {
          "audit_correct": 518,
          "audit_total": 1024,
          "feature_index": 14,
          "polarity": 1,
          "selection_correct": 145,
          "selection_total": 256,
          "test_correct": 126,
          "test_total": 256
        },
        "k": 32,
        "leaky": {
          "audit_correct": 517,
          "audit_total": 1024,
          "feature_index": 15,
          "polarity": -1,
          "test_correct": 143,
          "test_total": 256
        },
        "seed": 20
      },
      {
        "candidate_pair_count": 128,
        "derived": {
          "honest_audit_accuracy": 0.505859375,
          "honest_selection_accuracy": 0.56640625,
          "honest_test_accuracy": 0.4921875,
          "leaky_audit_accuracy": 0.5205078125,
          "leaky_reported_test_accuracy": 0.56640625,
          "leaky_vs_honest_test_gap": 0.07421875,
          "optimism_gap": 0.0458984375
        },
        "honest": {
          "audit_correct": 518,
          "audit_total": 1024,
          "feature_index": 14,
          "polarity": 1,
          "selection_correct": 145,
          "selection_total": 256,
          "test_correct": 126,
          "test_total": 256
        },
        "k": 64,
        "leaky": {
          "audit_correct": 533,
          "audit_total": 1024,
          "feature_index": 34,
          "polarity": -1,
          "test_correct": 145,
          "test_total": 256
        },
        "seed": 20
      },
      {
        "candidate_pair_count": 256,
        "derived": {
          "honest_audit_accuracy": 0.505859375,
          "honest_selection_accuracy": 0.56640625,
          "honest_test_accuracy": 0.4921875,
          "leaky_audit_accuracy": 0.5009765625,
          "leaky_reported_test_accuracy": 0.58203125,
          "leaky_vs_honest_test_gap": 0.08984375,
          "optimism_gap": 0.0810546875
        },
        "honest": {
          "audit_correct": 518,
          "audit_total": 1024,
          "feature_index": 14,
          "polarity": 1,
          "selection_correct": 145,
          "selection_total": 256,
          "test_correct": 126,
          "test_total": 256
        },
        "k": 128,
        "leaky": {
          "audit_correct": 513,
          "audit_total": 1024,
          "feature_index": 65,
          "polarity": 1,
          "test_correct": 149,
          "test_total": 256
        },
        "seed": 20
      },
      {
        "candidate_pair_count": 512,
        "derived": {
          "honest_audit_accuracy": 0.5029296875,
          "honest_selection_accuracy": 0.5859375,
          "honest_test_accuracy": 0.48046875,
          "leaky_audit_accuracy": 0.5048828125,
          "leaky_reported_test_accuracy": 0.5859375,
          "leaky_vs_honest_test_gap": 0.10546875,
          "optimism_gap": 0.0810546875
        },
        "honest": {
          "audit_correct": 515,
          "audit_total": 1024,
          "feature_index": 152,
          "polarity": 1,
          "selection_correct": 150,
          "selection_total": 256,
          "test_correct": 123,
          "test_total": 256
        },
        "k": 256,
        "leaky": {
          "audit_correct": 517,
          "audit_total": 1024,
          "feature_index": 176,
          "polarity": -1,
          "test_correct": 150,
          "test_total": 256
        },
        "seed": 20
      },
      {
        "candidate_pair_count": 1024,
        "derived": {
          "honest_audit_accuracy": 0.5029296875,
          "honest_selection_accuracy": 0.5859375,
          "honest_test_accuracy": 0.48046875,
          "leaky_audit_accuracy": 0.501953125,
          "leaky_reported_test_accuracy": 0.6015625,
          "leaky_vs_honest_test_gap": 0.12109375,
          "optimism_gap": 0.099609375
        },
        "honest": {
          "audit_correct": 515,
          "audit_total": 1024,
          "feature_index": 152,
          "polarity": 1,
          "selection_correct": 150,
          "selection_total": 256,
          "test_correct": 123,
          "test_total": 256
        },
        "k": 512,
        "leaky": {
          "audit_correct": 514,
          "audit_total": 1024,
          "feature_index": 388,
          "polarity": -1,
          "test_correct": 154,
          "test_total": 256
        },
        "seed": 20
      },
      {
        "candidate_pair_count": 2048,
        "derived": {
          "honest_audit_accuracy": 0.4931640625,
          "honest_selection_accuracy": 0.59375,
          "honest_test_accuracy": 0.44921875,
          "leaky_audit_accuracy": 0.501953125,
          "leaky_reported_test_accuracy": 0.6015625,
          "leaky_vs_honest_test_gap": 0.15234375,
          "optimism_gap": 0.099609375
        },
        "honest": {
          "audit_correct": 505,
          "audit_total": 1024,
          "feature_index": 992,
          "polarity": -1,
          "selection_correct": 152,
          "selection_total": 256,
          "test_correct": 115,
          "test_total": 256
        },
        "k": 1024,
        "leaky": {
          "audit_correct": 514,
          "audit_total": 1024,
          "feature_index": 388,
          "polarity": -1,
          "test_correct": 154,
          "test_total": 256
        },
        "seed": 20
      },
      {
        "candidate_pair_count": 2,
        "derived": {
          "honest_audit_accuracy": 0.529296875,
          "honest_selection_accuracy": 0.50390625,
          "honest_test_accuracy": 0.515625,
          "leaky_audit_accuracy": 0.529296875,
          "leaky_reported_test_accuracy": 0.515625,
          "leaky_vs_honest_test_gap": 0.0,
          "optimism_gap": -0.013671875
        },
        "honest": {
          "audit_correct": 542,
          "audit_total": 1024,
          "feature_index": 0,
          "polarity": 1,
          "selection_correct": 129,
          "selection_total": 256,
          "test_correct": 132,
          "test_total": 256
        },
        "k": 1,
        "leaky": {
          "audit_correct": 542,
          "audit_total": 1024,
          "feature_index": 0,
          "polarity": 1,
          "test_correct": 132,
          "test_total": 256
        },
        "seed": 21
      },
      {
        "candidate_pair_count": 4,
        "derived": {
          "honest_audit_accuracy": 0.486328125,
          "honest_selection_accuracy": 0.5078125,
          "honest_test_accuracy": 0.57421875,
          "leaky_audit_accuracy": 0.486328125,
          "leaky_reported_test_accuracy": 0.57421875,
          "leaky_vs_honest_test_gap": 0.0,
          "optimism_gap": 0.087890625
        },
        "honest": {
          "audit_correct": 498,
          "audit_total": 1024,
          "feature_index": 1,
          "polarity": 1,
          "selection_correct": 130,
          "selection_total": 256,
          "test_correct": 147,
          "test_total": 256
        },
        "k": 2,
        "leaky": {
          "audit_correct": 498,
          "audit_total": 1024,
          "feature_index": 1,
          "polarity": 1,
          "test_correct": 147,
          "test_total": 256
        },
        "seed": 21
      },
      {
        "candidate_pair_count": 8,
        "derived": {
          "honest_audit_accuracy": 0.5048828125,
          "honest_selection_accuracy": 0.5546875,
          "honest_test_accuracy": 0.54296875,
          "leaky_audit_accuracy": 0.486328125,
          "leaky_reported_test_accuracy": 0.57421875,
          "leaky_vs_honest_test_gap": 0.03125,
          "optimism_gap": 0.087890625
        },
        "honest": {
          "audit_correct": 517,
          "audit_total": 1024,
          "feature_index": 2,
          "polarity": -1,
          "selection_correct": 142,
          "selection_total": 256,
          "test_correct": 139,
          "test_total": 256
        },
        "k": 4,
        "leaky": {
          "audit_correct": 498,
          "audit_total": 1024,
          "feature_index": 1,
          "polarity": 1,
          "test_correct": 147,
          "test_total": 256
        },
        "seed": 21
      },
      {
        "candidate_pair_count": 16,
        "derived": {
          "honest_audit_accuracy": 0.4990234375,
          "honest_selection_accuracy": 0.5703125,
          "honest_test_accuracy": 0.5390625,
          "leaky_audit_accuracy": 0.486328125,
          "leaky_reported_test_accuracy": 0.57421875,
          "leaky_vs_honest_test_gap": 0.03515625,
          "optimism_gap": 0.087890625
        },
        "honest": {
          "audit_correct": 511,
          "audit_total": 1024,
          "feature_index": 5,
          "polarity": 1,
          "selection_correct": 146,
          "selection_total": 256,
          "test_correct": 138,
          "test_total": 256
        },
        "k": 8,
        "leaky": {
          "audit_correct": 498,
          "audit_total": 1024,
          "feature_index": 1,
          "polarity": 1,
          "test_correct": 147,
          "test_total": 256
        },
        "seed": 21
      },
      {
        "candidate_pair_count": 32,
        "derived": {
          "honest_audit_accuracy": 0.4990234375,
          "honest_selection_accuracy": 0.5703125,
          "honest_test_accuracy": 0.5390625,
          "leaky_audit_accuracy": 0.486328125,
          "leaky_reported_test_accuracy": 0.57421875,
          "leaky_vs_honest_test_gap": 0.03515625,
          "optimism_gap": 0.087890625
        },
        "honest": {
          "audit_correct": 511,
          "audit_total": 1024,
          "feature_index": 5,
          "polarity": 1,
          "selection_correct": 146,
          "selection_total": 256,
          "test_correct": 138,
          "test_total": 256
        },
        "k": 16,
        "leaky": {
          "audit_correct": 498,
          "audit_total": 1024,
          "feature_index": 1,
          "polarity": 1,
          "test_correct": 147,
          "test_total": 256
        },
        "seed": 21
      },
      {
        "candidate_pair_count": 64,
        "derived": {
          "honest_audit_accuracy": 0.4951171875,
          "honest_selection_accuracy": 0.58984375,
          "honest_test_accuracy": 0.53125,
          "leaky_audit_accuracy": 0.486328125,
          "leaky_reported_test_accuracy": 0.57421875,
          "leaky_vs_honest_test_gap": 0.04296875,
          "optimism_gap": 0.087890625
        },
        "honest": {
          "audit_correct": 507,
          "audit_total": 1024,
          "feature_index": 29,
          "polarity": -1,
          "selection_correct": 151,
          "selection_total": 256,
          "test_correct": 136,
          "test_total": 256
        },
        "k": 32,
        "leaky": {
          "audit_correct": 498,
          "audit_total": 1024,
          "feature_index": 1,
          "polarity": 1,
          "test_correct": 147,
          "test_total": 256
        },
        "seed": 21
      },
      {
        "candidate_pair_count": 128,
        "derived": {
          "honest_audit_accuracy": 0.4951171875,
          "honest_selection_accuracy": 0.58984375,
          "honest_test_accuracy": 0.53125,
          "leaky_audit_accuracy": 0.486328125,
          "leaky_reported_test_accuracy": 0.57421875,
          "leaky_vs_honest_test_gap": 0.04296875,
          "optimism_gap": 0.087890625
        },
        "honest": {
          "audit_correct": 507,
          "audit_total": 1024,
          "feature_index": 29,
          "polarity": -1,
          "selection_correct": 151,
          "selection_total": 256,
          "test_correct": 136,
          "test_total": 256
        },
        "k": 64,
        "leaky": {
          "audit_correct": 498,
          "audit_total": 1024,
          "feature_index": 1,
          "polarity": 1,
          "test_correct": 147,
          "test_total": 256
        },
        "seed": 21
      },
      {
        "candidate_pair_count": 256,
        "derived": {
          "honest_audit_accuracy": 0.4951171875,
          "honest_selection_accuracy": 0.58984375,
          "honest_test_accuracy": 0.53125,
          "leaky_audit_accuracy": 0.486328125,
          "leaky_reported_test_accuracy": 0.578125,
          "leaky_vs_honest_test_gap": 0.046875,
          "optimism_gap": 0.091796875
        },
        "honest": {
          "audit_correct": 507,
          "audit_total": 1024,
          "feature_index": 29,
          "polarity": -1,
          "selection_correct": 151,
          "selection_total": 256,
          "test_correct": 136,
          "test_total": 256
        },
        "k": 128,
        "leaky": {
          "audit_correct": 498,
          "audit_total": 1024,
          "feature_index": 98,
          "polarity": 1,
          "test_correct": 148,
          "test_total": 256
        },
        "seed": 21
      },
      {
        "candidate_pair_count": 512,
        "derived": {
          "honest_audit_accuracy": 0.4951171875,
          "honest_selection_accuracy": 0.58984375,
          "honest_test_accuracy": 0.53125,
          "leaky_audit_accuracy": 0.4931640625,
          "leaky_reported_test_accuracy": 0.6015625,
          "leaky_vs_honest_test_gap": 0.0703125,
          "optimism_gap": 0.1083984375
        },
        "honest": {
          "audit_correct": 507,
          "audit_total": 1024,
          "feature_index": 29,
          "polarity": -1,
          "selection_correct": 151,
          "selection_total": 256,
          "test_correct": 136,
          "test_total": 256
        },
        "k": 256,
        "leaky": {
          "audit_correct": 505,
          "audit_total": 1024,
          "feature_index": 181,
          "polarity": -1,
          "test_correct": 154,
          "test_total": 256
        },
        "seed": 21
      },
      {
        "candidate_pair_count": 1024,
        "derived": {
          "honest_audit_accuracy": 0.4951171875,
          "honest_selection_accuracy": 0.58984375,
          "honest_test_accuracy": 0.53125,
          "leaky_audit_accuracy": 0.4931640625,
          "leaky_reported_test_accuracy": 0.6015625,
          "leaky_vs_honest_test_gap": 0.0703125,
          "optimism_gap": 0.1083984375
        },
        "honest": {
          "audit_correct": 507,
          "audit_total": 1024,
          "feature_index": 29,
          "polarity": -1,
          "selection_correct": 151,
          "selection_total": 256,
          "test_correct": 136,
          "test_total": 256
        },
        "k": 512,
        "leaky": {
          "audit_correct": 505,
          "audit_total": 1024,
          "feature_index": 181,
          "polarity": -1,
          "test_correct": 154,
          "test_total": 256
        },
        "seed": 21
      },
      {
        "candidate_pair_count": 2048,
        "derived": {
          "honest_audit_accuracy": 0.4951171875,
          "honest_selection_accuracy": 0.58984375,
          "honest_test_accuracy": 0.53125,
          "leaky_audit_accuracy": 0.4931640625,
          "leaky_reported_test_accuracy": 0.6015625,
          "leaky_vs_honest_test_gap": 0.0703125,
          "optimism_gap": 0.1083984375
        },
        "honest": {
          "audit_correct": 507,
          "audit_total": 1024,
          "feature_index": 29,
          "polarity": -1,
          "selection_correct": 151,
          "selection_total": 256,
          "test_correct": 136,
          "test_total": 256
        },
        "k": 1024,
        "leaky": {
          "audit_correct": 505,
          "audit_total": 1024,
          "feature_index": 181,
          "polarity": -1,
          "test_correct": 154,
          "test_total": 256
        },
        "seed": 21
      },
      {
        "candidate_pair_count": 2,
        "derived": {
          "honest_audit_accuracy": 0.5078125,
          "honest_selection_accuracy": 0.5234375,
          "honest_test_accuracy": 0.47265625,
          "leaky_audit_accuracy": 0.4921875,
          "leaky_reported_test_accuracy": 0.52734375,
          "leaky_vs_honest_test_gap": 0.0546875,
          "optimism_gap": 0.03515625
        },
        "honest": {
          "audit_correct": 520,
          "audit_total": 1024,
          "feature_index": 0,
          "polarity": 1,
          "selection_correct": 134,
          "selection_total": 256,
          "test_correct": 121,
          "test_total": 256
        },
        "k": 1,
        "leaky": {
          "audit_correct": 504,
          "audit_total": 1024,
          "feature_index": 0,
          "polarity": -1,
          "test_correct": 135,
          "test_total": 256
        },
        "seed": 22
      },
      {
        "candidate_pair_count": 4,
        "derived": {
          "honest_audit_accuracy": 0.5078125,
          "honest_selection_accuracy": 0.5234375,
          "honest_test_accuracy": 0.47265625,
          "leaky_audit_accuracy": 0.4921875,
          "leaky_reported_test_accuracy": 0.52734375,
          "leaky_vs_honest_test_gap": 0.0546875,
          "optimism_gap": 0.03515625
        },
        "honest": {
          "audit_correct": 520,
          "audit_total": 1024,
          "feature_index": 0,
          "polarity": 1,
          "selection_correct": 134,
          "selection_total": 256,
          "test_correct": 121,
          "test_total": 256
        },
        "k": 2,
        "leaky": {
          "audit_correct": 504,
          "audit_total": 1024,
          "feature_index": 0,
          "polarity": -1,
          "test_correct": 135,
          "test_total": 256
        },
        "seed": 22
      },
      {
        "candidate_pair_count": 8,
        "derived": {
          "honest_audit_accuracy": 0.4853515625,
          "honest_selection_accuracy": 0.5390625,
          "honest_test_accuracy": 0.51953125,
          "leaky_audit_accuracy": 0.4921875,
          "leaky_reported_test_accuracy": 0.52734375,
          "leaky_vs_honest_test_gap": 0.0078125,
          "optimism_gap": 0.03515625
        },
        "honest": {
          "audit_correct": 497,
          "audit_total": 1024,
          "feature_index": 2,
          "polarity": -1,
          "selection_correct": 138,
          "selection_total": 256,
          "test_correct": 133,
          "test_total": 256
        },
        "k": 4,
        "leaky": {
          "audit_correct": 504,
          "audit_total": 1024,
          "feature_index": 0,
          "polarity": -1,
          "test_correct": 135,
          "test_total": 256
        },
        "seed": 22
      },
      {
        "candidate_pair_count": 16,
        "derived": {
          "honest_audit_accuracy": 0.5029296875,
          "honest_selection_accuracy": 0.5625,
          "honest_test_accuracy": 0.51171875,
          "leaky_audit_accuracy": 0.4921875,
          "leaky_reported_test_accuracy": 0.5390625,
          "leaky_vs_honest_test_gap": 0.02734375,
          "optimism_gap": 0.046875
        },
        "honest": {
          "audit_correct": 515,
          "audit_total": 1024,
          "feature_index": 4,
          "polarity": -1,
          "selection_correct": 144,
          "selection_total": 256,
          "test_correct": 131,
          "test_total": 256
        },
        "k": 8,
        "leaky": {
          "audit_correct": 504,
          "audit_total": 1024,
          "feature_index": 5,
          "polarity": 1,
          "test_correct": 138,
          "test_total": 256
        },
        "seed": 22
      },
      {
        "candidate_pair_count": 32,
        "derived": {
          "honest_audit_accuracy": 0.5068359375,
          "honest_selection_accuracy": 0.57421875,
          "honest_test_accuracy": 0.48828125,
          "leaky_audit_accuracy": 0.509765625,
          "leaky_reported_test_accuracy": 0.546875,
          "leaky_vs_honest_test_gap": 0.05859375,
          "optimism_gap": 0.037109375
        },
        "honest": {
          "audit_correct": 519,
          "audit_total": 1024,
          "feature_index": 12,
          "polarity": 1,
          "selection_correct": 147,
          "selection_total": 256,
          "test_correct": 125,
          "test_total": 256
        },
        "k": 16,
        "leaky": {
          "audit_correct": 522,
          "audit_total": 1024,
          "feature_index": 14,
          "polarity": 1,
          "test_correct": 140,
          "test_total": 256
        },
        "seed": 22
      },
      {
        "candidate_pair_count": 64,
        "derived": {
          "honest_audit_accuracy": 0.5068359375,
          "honest_selection_accuracy": 0.57421875,
          "honest_test_accuracy": 0.48828125,
          "leaky_audit_accuracy": 0.515625,
          "leaky_reported_test_accuracy": 0.5625,
          "leaky_vs_honest_test_gap": 0.07421875,
          "optimism_gap": 0.046875
        },
        "honest": {
          "audit_correct": 519,
          "audit_total": 1024,
          "feature_index": 12,
          "polarity": 1,
          "selection_correct": 147,
          "selection_total": 256,
          "test_correct": 125,
          "test_total": 256
        },
        "k": 32,
        "leaky": {
          "audit_correct": 528,
          "audit_total": 1024,
          "feature_index": 28,
          "polarity": -1,
          "test_correct": 144,
          "test_total": 256
        },
        "seed": 22
      },
      {
        "candidate_pair_count": 128,
        "derived": {
          "honest_audit_accuracy": 0.5068359375,
          "honest_selection_accuracy": 0.57421875,
          "honest_test_accuracy": 0.48828125,
          "leaky_audit_accuracy": 0.5009765625,
          "leaky_reported_test_accuracy": 0.5703125,
          "leaky_vs_honest_test_gap": 0.08203125,
          "optimism_gap": 0.0693359375
        },
        "honest": {
          "audit_correct": 519,
          "audit_total": 1024,
          "feature_index": 12,
          "polarity": 1,
          "selection_correct": 147,
          "selection_total": 256,
          "test_correct": 125,
          "test_total": 256
        },
        "k": 64,
        "leaky": {
          "audit_correct": 513,
          "audit_total": 1024,
          "feature_index": 40,
          "polarity": 1,
          "test_correct": 146,
          "test_total": 256
        },
        "seed": 22
      },
      {
        "candidate_pair_count": 256,
        "derived": {
          "honest_audit_accuracy": 0.5068359375,
          "honest_selection_accuracy": 0.57421875,
          "honest_test_accuracy": 0.48828125,
          "leaky_audit_accuracy": 0.5009765625,
          "leaky_reported_test_accuracy": 0.60546875,
          "leaky_vs_honest_test_gap": 0.1171875,
          "optimism_gap": 0.1044921875
        },
        "honest": {
          "audit_correct": 519,
          "audit_total": 1024,
          "feature_index": 12,
          "polarity": 1,
          "selection_correct": 147,
          "selection_total": 256,
          "test_correct": 125,
          "test_total": 256
        },
        "k": 128,
        "leaky": {
          "audit_correct": 513,
          "audit_total": 1024,
          "feature_index": 92,
          "polarity": -1,
          "test_correct": 155,
          "test_total": 256
        },
        "seed": 22
      },
      {
        "candidate_pair_count": 512,
        "derived": {
          "honest_audit_accuracy": 0.5068359375,
          "honest_selection_accuracy": 0.58203125,
          "honest_test_accuracy": 0.49609375,
          "leaky_audit_accuracy": 0.5009765625,
          "leaky_reported_test_accuracy": 0.60546875,
          "leaky_vs_honest_test_gap": 0.109375,
          "optimism_gap": 0.1044921875
        },
        "honest": {
          "audit_correct": 519,
          "audit_total": 1024,
          "feature_index": 223,
          "polarity": 1,
          "selection_correct": 149,
          "selection_total": 256,
          "test_correct": 127,
          "test_total": 256
        },
        "k": 256,
        "leaky": {
          "audit_correct": 513,
          "audit_total": 1024,
          "feature_index": 92,
          "polarity": -1,
          "test_correct": 155,
          "test_total": 256
        },
        "seed": 22
      },
      {
        "candidate_pair_count": 1024,
        "derived": {
          "honest_audit_accuracy": 0.5068359375,
          "honest_selection_accuracy": 0.58203125,
          "honest_test_accuracy": 0.49609375,
          "leaky_audit_accuracy": 0.5009765625,
          "leaky_reported_test_accuracy": 0.60546875,
          "leaky_vs_honest_test_gap": 0.109375,
          "optimism_gap": 0.1044921875
        },
        "honest": {
          "audit_correct": 519,
          "audit_total": 1024,
          "feature_index": 223,
          "polarity": 1,
          "selection_correct": 149,
          "selection_total": 256,
          "test_correct": 127,
          "test_total": 256
        },
        "k": 512,
        "leaky": {
          "audit_correct": 513,
          "audit_total": 1024,
          "feature_index": 92,
          "polarity": -1,
          "test_correct": 155,
          "test_total": 256
        },
        "seed": 22
      },
      {
        "candidate_pair_count": 2048,
        "derived": {
          "honest_audit_accuracy": 0.4814453125,
          "honest_selection_accuracy": 0.609375,
          "honest_test_accuracy": 0.4375,
          "leaky_audit_accuracy": 0.5009765625,
          "leaky_reported_test_accuracy": 0.60546875,
          "leaky_vs_honest_test_gap": 0.16796875,
          "optimism_gap": 0.1044921875
        },
        "honest": {
          "audit_correct": 493,
          "audit_total": 1024,
          "feature_index": 877,
          "polarity": -1,
          "selection_correct": 156,
          "selection_total": 256,
          "test_correct": 112,
          "test_total": 256
        },
        "k": 1024,
        "leaky": {
          "audit_correct": 513,
          "audit_total": 1024,
          "feature_index": 92,
          "polarity": -1,
          "test_correct": 155,
          "test_total": 256
        },
        "seed": 22
      },
      {
        "candidate_pair_count": 2,
        "derived": {
          "honest_audit_accuracy": 0.4736328125,
          "honest_selection_accuracy": 0.5234375,
          "honest_test_accuracy": 0.55859375,
          "leaky_audit_accuracy": 0.4736328125,
          "leaky_reported_test_accuracy": 0.55859375,
          "leaky_vs_honest_test_gap": 0.0,
          "optimism_gap": 0.0849609375
        },
        "honest": {
          "audit_correct": 485,
          "audit_total": 1024,
          "feature_index": 0,
          "polarity": 1,
          "selection_correct": 134,
          "selection_total": 256,
          "test_correct": 143,
          "test_total": 256
        },
        "k": 1,
        "leaky": {
          "audit_correct": 485,
          "audit_total": 1024,
          "feature_index": 0,
          "polarity": 1,
          "test_correct": 143,
          "test_total": 256
        },
        "seed": 23
      },
      {
        "candidate_pair_count": 4,
        "derived": {
          "honest_audit_accuracy": 0.4736328125,
          "honest_selection_accuracy": 0.5234375,
          "honest_test_accuracy": 0.55859375,
          "leaky_audit_accuracy": 0.4736328125,
          "leaky_reported_test_accuracy": 0.55859375,
          "leaky_vs_honest_test_gap": 0.0,
          "optimism_gap": 0.0849609375
        },
        "honest": {
          "audit_correct": 485,
          "audit_total": 1024,
          "feature_index": 0,
          "polarity": 1,
          "selection_correct": 134,
          "selection_total": 256,
          "test_correct": 143,
          "test_total": 256
        },
        "k": 2,
        "leaky": {
          "audit_correct": 485,
          "audit_total": 1024,
          "feature_index": 0,
          "polarity": 1,
          "test_correct": 143,
          "test_total": 256
        },
        "seed": 23
      },
      {
        "candidate_pair_count": 8,
        "derived": {
          "honest_audit_accuracy": 0.517578125,
          "honest_selection_accuracy": 0.53515625,
          "honest_test_accuracy": 0.53125,
          "leaky_audit_accuracy": 0.4736328125,
          "leaky_reported_test_accuracy": 0.55859375,
          "leaky_vs_honest_test_gap": 0.02734375,
          "optimism_gap": 0.0849609375
        },
        "honest": {
          "audit_correct": 530,
          "audit_total": 1024,
          "feature_index": 2,
          "polarity": -1,
          "selection_correct": 137,
          "selection_total": 256,
          "test_correct": 136,
          "test_total": 256
        },
        "k": 4,
        "leaky": {
          "audit_correct": 485,
          "audit_total": 1024,
          "feature_index": 0,
          "polarity": 1,
          "test_correct": 143,
          "test_total": 256
        },
        "seed": 23
      },
      {
        "candidate_pair_count": 16,
        "derived": {
          "honest_audit_accuracy": 0.4873046875,
          "honest_selection_accuracy": 0.5625,
          "honest_test_accuracy": 0.5234375,
          "leaky_audit_accuracy": 0.4736328125,
          "leaky_reported_test_accuracy": 0.55859375,
          "leaky_vs_honest_test_gap": 0.03515625,
          "optimism_gap": 0.0849609375
        },
        "honest": {
          "audit_correct": 499,
          "audit_total": 1024,
          "feature_index": 7,
          "polarity": 1,
          "selection_correct": 144,
          "selection_total": 256,
          "test_correct": 134,
          "test_total": 256
        },
        "k": 8,
        "leaky": {
          "audit_correct": 485,
          "audit_total": 1024,
          "feature_index": 0,
          "polarity": 1,
          "test_correct": 143,
          "test_total": 256
        },
        "seed": 23
      },
      {
        "candidate_pair_count": 32,
        "derived": {
          "honest_audit_accuracy": 0.4873046875,
          "honest_selection_accuracy": 0.5625,
          "honest_test_accuracy": 0.5234375,
          "leaky_audit_accuracy": 0.4736328125,
          "leaky_reported_test_accuracy": 0.55859375,
          "leaky_vs_honest_test_gap": 0.03515625,
          "optimism_gap": 0.0849609375
        },
        "honest": {
          "audit_correct": 499,
          "audit_total": 1024,
          "feature_index": 7,
          "polarity": 1,
          "selection_correct": 144,
          "selection_total": 256,
          "test_correct": 134,
          "test_total": 256
        },
        "k": 16,
        "leaky": {
          "audit_correct": 485,
          "audit_total": 1024,
          "feature_index": 0,
          "polarity": 1,
          "test_correct": 143,
          "test_total": 256
        },
        "seed": 23
      },
      {
        "candidate_pair_count": 64,
        "derived": {
          "honest_audit_accuracy": 0.4873046875,
          "honest_selection_accuracy": 0.5625,
          "honest_test_accuracy": 0.5234375,
          "leaky_audit_accuracy": 0.4736328125,
          "leaky_reported_test_accuracy": 0.55859375,
          "leaky_vs_honest_test_gap": 0.03515625,
          "optimism_gap": 0.0849609375
        },
        "honest": {
          "audit_correct": 499,
          "audit_total": 1024,
          "feature_index": 7,
          "polarity": 1,
          "selection_correct": 144,
          "selection_total": 256,
          "test_correct": 134,
          "test_total": 256
        },
        "k": 32,
        "leaky": {
          "audit_correct": 485,
          "audit_total": 1024,
          "feature_index": 0,
          "polarity": 1,
          "test_correct": 143,
          "test_total": 256
        },
        "seed": 23
      },
      {
        "candidate_pair_count": 128,
        "derived": {
          "honest_audit_accuracy": 0.4873046875,
          "honest_selection_accuracy": 0.5625,
          "honest_test_accuracy": 0.5234375,
          "leaky_audit_accuracy": 0.4892578125,
          "leaky_reported_test_accuracy": 0.578125,
          "leaky_vs_honest_test_gap": 0.0546875,
          "optimism_gap": 0.0888671875
        },
        "honest": {
          "audit_correct": 499,
          "audit_total": 1024,
          "feature_index": 7,
          "polarity": 1,
          "selection_correct": 144,
          "selection_total": 256,
          "test_correct": 134,
          "test_total": 256
        },
        "k": 64,
        "leaky": {
          "audit_correct": 501,
          "audit_total": 1024,
          "feature_index": 56,
          "polarity": -1,
          "test_correct": 148,
          "test_total": 256
        },
        "seed": 23
      },
      {
        "candidate_pair_count": 256,
        "derived": {
          "honest_audit_accuracy": 0.486328125,
          "honest_selection_accuracy": 0.578125,
          "honest_test_accuracy": 0.4375,
          "leaky_audit_accuracy": 0.4892578125,
          "leaky_reported_test_accuracy": 0.578125,
          "leaky_vs_honest_test_gap": 0.140625,
          "optimism_gap": 0.0888671875
        },
        "honest": {
          "audit_correct": 498,
          "audit_total": 1024,
          "feature_index": 104,
          "polarity": 1,
          "selection_correct": 148,
          "selection_total": 256,
          "test_correct": 112,
          "test_total": 256
        },
        "k": 128,
        "leaky": {
          "audit_correct": 501,
          "audit_total": 1024,
          "feature_index": 56,
          "polarity": -1,
          "test_correct": 148,
          "test_total": 256
        },
        "seed": 23
      },
      {
        "candidate_pair_count": 512,
        "derived": {
          "honest_audit_accuracy": 0.486328125,
          "honest_selection_accuracy": 0.578125,
          "honest_test_accuracy": 0.4375,
          "leaky_audit_accuracy": 0.501953125,
          "leaky_reported_test_accuracy": 0.58984375,
          "leaky_vs_honest_test_gap": 0.15234375,
          "optimism_gap": 0.087890625
        },
        "honest": {
          "audit_correct": 498,
          "audit_total": 1024,
          "feature_index": 104,
          "polarity": 1,
          "selection_correct": 148,
          "selection_total": 256,
          "test_correct": 112,
          "test_total": 256
        },
        "k": 256,
        "leaky": {
          "audit_correct": 514,
          "audit_total": 1024,
          "feature_index": 182,
          "polarity": 1,
          "test_correct": 151,
          "test_total": 256
        },
        "seed": 23
      },
      {
        "candidate_pair_count": 1024,
        "derived": {
          "honest_audit_accuracy": 0.5146484375,
          "honest_selection_accuracy": 0.6015625,
          "honest_test_accuracy": 0.4921875,
          "leaky_audit_accuracy": 0.50390625,
          "leaky_reported_test_accuracy": 0.59765625,
          "leaky_vs_honest_test_gap": 0.10546875,
          "optimism_gap": 0.09375
        },
        "honest": {
          "audit_correct": 527,
          "audit_total": 1024,
          "feature_index": 506,
          "polarity": -1,
          "selection_correct": 154,
          "selection_total": 256,
          "test_correct": 126,
          "test_total": 256
        },
        "k": 512,
        "leaky": {
          "audit_correct": 516,
          "audit_total": 1024,
          "feature_index": 342,
          "polarity": -1,
          "test_correct": 153,
          "test_total": 256
        },
        "seed": 23
      },
      {
        "candidate_pair_count": 2048,
        "derived": {
          "honest_audit_accuracy": 0.5146484375,
          "honest_selection_accuracy": 0.6015625,
          "honest_test_accuracy": 0.4921875,
          "leaky_audit_accuracy": 0.5,
          "leaky_reported_test_accuracy": 0.609375,
          "leaky_vs_honest_test_gap": 0.1171875,
          "optimism_gap": 0.109375
        },
        "honest": {
          "audit_correct": 527,
          "audit_total": 1024,
          "feature_index": 506,
          "polarity": -1,
          "selection_correct": 154,
          "selection_total": 256,
          "test_correct": 126,
          "test_total": 256
        },
        "k": 1024,
        "leaky": {
          "audit_correct": 512,
          "audit_total": 1024,
          "feature_index": 810,
          "polarity": 1,
          "test_correct": 156,
          "test_total": 256
        },
        "seed": 23
      },
      {
        "candidate_pair_count": 2,
        "derived": {
          "honest_audit_accuracy": 0.498046875,
          "honest_selection_accuracy": 0.50390625,
          "honest_test_accuracy": 0.50390625,
          "leaky_audit_accuracy": 0.498046875,
          "leaky_reported_test_accuracy": 0.50390625,
          "leaky_vs_honest_test_gap": 0.0,
          "optimism_gap": 0.005859375
        },
        "honest": {
          "audit_correct": 510,
          "audit_total": 1024,
          "feature_index": 0,
          "polarity": 1,
          "selection_correct": 129,
          "selection_total": 256,
          "test_correct": 129,
          "test_total": 256
        },
        "k": 1,
        "leaky": {
          "audit_correct": 510,
          "audit_total": 1024,
          "feature_index": 0,
          "polarity": 1,
          "test_correct": 129,
          "test_total": 256
        },
        "seed": 24
      },
      {
        "candidate_pair_count": 4,
        "derived": {
          "honest_audit_accuracy": 0.486328125,
          "honest_selection_accuracy": 0.5390625,
          "honest_test_accuracy": 0.50390625,
          "leaky_audit_accuracy": 0.498046875,
          "leaky_reported_test_accuracy": 0.50390625,
          "leaky_vs_honest_test_gap": 0.0,
          "optimism_gap": 0.005859375
        },
        "honest": {
          "audit_correct": 498,
          "audit_total": 1024,
          "feature_index": 1,
          "polarity": -1,
          "selection_correct": 138,
          "selection_total": 256,
          "test_correct": 129,
          "test_total": 256
        },
        "k": 2,
        "leaky": {
          "audit_correct": 510,
          "audit_total": 1024,
          "feature_index": 0,
          "polarity": 1,
          "test_correct": 129,
          "test_total": 256
        },
        "seed": 24
      },
      {
        "candidate_pair_count": 8,
        "derived": {
          "honest_audit_accuracy": 0.5166015625,
          "honest_selection_accuracy": 0.5625,
          "honest_test_accuracy": 0.5703125,
          "leaky_audit_accuracy": 0.5166015625,
          "leaky_reported_test_accuracy": 0.5703125,
          "leaky_vs_honest_test_gap": 0.0,
          "optimism_gap": 0.0537109375
        },
        "honest": {
          "audit_correct": 529,
          "audit_total": 1024,
          "feature_index": 2,
          "polarity": -1,
          "selection_correct": 144,
          "selection_total": 256,
          "test_correct": 146,
          "test_total": 256
        },
        "k": 4,
        "leaky": {
          "audit_correct": 529,
          "audit_total": 1024,
          "feature_index": 2,
          "polarity": -1,
          "test_correct": 146,
          "test_total": 256
        },
        "seed": 24
      },
      {
        "candidate_pair_count": 16,
        "derived": {
          "honest_audit_accuracy": 0.48828125,
          "honest_selection_accuracy": 0.57421875,
          "honest_test_accuracy": 0.47265625,
          "leaky_audit_accuracy": 0.5166015625,
          "leaky_reported_test_accuracy": 0.5703125,
          "leaky_vs_honest_test_gap": 0.09765625,
          "optimism_gap": 0.0537109375
        },
        "honest": {
          "audit_correct": 500,
          "audit_total": 1024,
          "feature_index": 6,
          "polarity": 1,
          "selection_correct": 147,
          "selection_total": 256,
          "test_correct": 121,
          "test_total": 256
        },
        "k": 8,
        "leaky": {
          "audit_correct": 529,
          "audit_total": 1024,
          "feature_index": 2,
          "polarity": -1,
          "test_correct": 146,
          "test_total": 256
        },
        "seed": 24
      },
      {
        "candidate_pair_count": 32,
        "derived": {
          "honest_audit_accuracy": 0.48828125,
          "honest_selection_accuracy": 0.57421875,
          "honest_test_accuracy": 0.47265625,
          "leaky_audit_accuracy": 0.5166015625,
          "leaky_reported_test_accuracy": 0.5703125,
          "leaky_vs_honest_test_gap": 0.09765625,
          "optimism_gap": 0.0537109375
        },
        "honest": {
          "audit_correct": 500,
          "audit_total": 1024,
          "feature_index": 6,
          "polarity": 1,
          "selection_correct": 147,
          "selection_total": 256,
          "test_correct": 121,
          "test_total": 256
        },
        "k": 16,
        "leaky": {
          "audit_correct": 529,
          "audit_total": 1024,
          "feature_index": 2,
          "polarity": -1,
          "test_correct": 146,
          "test_total": 256
        },
        "seed": 24
      },
      {
        "candidate_pair_count": 64,
        "derived": {
          "honest_audit_accuracy": 0.48828125,
          "honest_selection_accuracy": 0.57421875,
          "honest_test_accuracy": 0.47265625,
          "leaky_audit_accuracy": 0.5166015625,
          "leaky_reported_test_accuracy": 0.5703125,
          "leaky_vs_honest_test_gap": 0.09765625,
          "optimism_gap": 0.0537109375
        },
        "honest": {
          "audit_correct": 500,
          "audit_total": 1024,
          "feature_index": 6,
          "polarity": 1,
          "selection_correct": 147,
          "selection_total": 256,
          "test_correct": 121,
          "test_total": 256
        },
        "k": 32,
        "leaky": {
          "audit_correct": 529,
          "audit_total": 1024,
          "feature_index": 2,
          "polarity": -1,
          "test_correct": 146,
          "test_total": 256
        },
        "seed": 24
      },
      {
        "candidate_pair_count": 128,
        "derived": {
          "honest_audit_accuracy": 0.48046875,
          "honest_selection_accuracy": 0.5859375,
          "honest_test_accuracy": 0.5390625,
          "leaky_audit_accuracy": 0.5166015625,
          "leaky_reported_test_accuracy": 0.5703125,
          "leaky_vs_honest_test_gap": 0.03125,
          "optimism_gap": 0.0537109375
        },
        "honest": {
          "audit_correct": 492,
          "audit_total": 1024,
          "feature_index": 33,
          "polarity": -1,
          "selection_correct": 150,
          "selection_total": 256,
          "test_correct": 138,
          "test_total": 256
        },
        "k": 64,
        "leaky": {
          "audit_correct": 529,
          "audit_total": 1024,
          "feature_index": 2,
          "polarity": -1,
          "test_correct": 146,
          "test_total": 256
        },
        "seed": 24
      },
      {
        "candidate_pair_count": 256,
        "derived": {
          "honest_audit_accuracy": 0.486328125,
          "honest_selection_accuracy": 0.59375,
          "honest_test_accuracy": 0.5234375,
          "leaky_audit_accuracy": 0.5166015625,
          "leaky_reported_test_accuracy": 0.5703125,
          "leaky_vs_honest_test_gap": 0.046875,
          "optimism_gap": 0.0537109375
        },
        "honest": {
          "audit_correct": 498,
          "audit_total": 1024,
          "feature_index": 97,
          "polarity": -1,
          "selection_correct": 152,
          "selection_total": 256,
          "test_correct": 134,
          "test_total": 256
        },
        "k": 128,
        "leaky": {
          "audit_correct": 529,
          "audit_total": 1024,
          "feature_index": 2,
          "polarity": -1,
          "test_correct": 146,
          "test_total": 256
        },
        "seed": 24
      },
      {
        "candidate_pair_count": 512,
        "derived": {
          "honest_audit_accuracy": 0.486328125,
          "honest_selection_accuracy": 0.59375,
          "honest_test_accuracy": 0.5234375,
          "leaky_audit_accuracy": 0.5087890625,
          "leaky_reported_test_accuracy": 0.58203125,
          "leaky_vs_honest_test_gap": 0.05859375,
          "optimism_gap": 0.0732421875
        },
        "honest": {
          "audit_correct": 498,
          "audit_total": 1024,
          "feature_index": 97,
          "polarity": -1,
          "selection_correct": 152,
          "selection_total": 256,
          "test_correct": 134,
          "test_total": 256
        },
        "k": 256,
        "leaky": {
          "audit_correct": 521,
          "audit_total": 1024,
          "feature_index": 135,
          "polarity": -1,
          "test_correct": 149,
          "test_total": 256
        },
        "seed": 24
      },
      {
        "candidate_pair_count": 1024,
        "derived": {
          "honest_audit_accuracy": 0.5205078125,
          "honest_selection_accuracy": 0.59765625,
          "honest_test_accuracy": 0.50390625,
          "leaky_audit_accuracy": 0.4853515625,
          "leaky_reported_test_accuracy": 0.59765625,
          "leaky_vs_honest_test_gap": 0.09375,
          "optimism_gap": 0.1123046875
        },
        "honest": {
          "audit_correct": 533,
          "audit_total": 1024,
          "feature_index": 454,
          "polarity": -1,
          "selection_correct": 153,
          "selection_total": 256,
          "test_correct": 129,
          "test_total": 256
        },
        "k": 512,
        "leaky": {
          "audit_correct": 497,
          "audit_total": 1024,
          "feature_index": 355,
          "polarity": -1,
          "test_correct": 153,
          "test_total": 256
        },
        "seed": 24
      },
      {
        "candidate_pair_count": 2048,
        "derived": {
          "honest_audit_accuracy": 0.5263671875,
          "honest_selection_accuracy": 0.60546875,
          "honest_test_accuracy": 0.4921875,
          "leaky_audit_accuracy": 0.5146484375,
          "leaky_reported_test_accuracy": 0.6171875,
          "leaky_vs_honest_test_gap": 0.125,
          "optimism_gap": 0.1025390625
        },
        "honest": {
          "audit_correct": 539,
          "audit_total": 1024,
          "feature_index": 947,
          "polarity": 1,
          "selection_correct": 155,
          "selection_total": 256,
          "test_correct": 126,
          "test_total": 256
        },
        "k": 1024,
        "leaky": {
          "audit_correct": 527,
          "audit_total": 1024,
          "feature_index": 653,
          "polarity": -1,
          "test_correct": 158,
          "test_total": 256
        },
        "seed": 24
      },
      {
        "candidate_pair_count": 2,
        "derived": {
          "honest_audit_accuracy": 0.49609375,
          "honest_selection_accuracy": 0.51953125,
          "honest_test_accuracy": 0.4921875,
          "leaky_audit_accuracy": 0.50390625,
          "leaky_reported_test_accuracy": 0.5078125,
          "leaky_vs_honest_test_gap": 0.015625,
          "optimism_gap": 0.00390625
        },
        "honest": {
          "audit_correct": 508,
          "audit_total": 1024,
          "feature_index": 0,
          "polarity": -1,
          "selection_correct": 133,
          "selection_total": 256,
          "test_correct": 126,
          "test_total": 256
        },
        "k": 1,
        "leaky": {
          "audit_correct": 516,
          "audit_total": 1024,
          "feature_index": 0,
          "polarity": 1,
          "test_correct": 130,
          "test_total": 256
        },
        "seed": 25
      },
      {
        "candidate_pair_count": 4,
        "derived": {
          "honest_audit_accuracy": 0.49609375,
          "honest_selection_accuracy": 0.51953125,
          "honest_test_accuracy": 0.4921875,
          "leaky_audit_accuracy": 0.4951171875,
          "leaky_reported_test_accuracy": 0.5234375,
          "leaky_vs_honest_test_gap": 0.03125,
          "optimism_gap": 0.0283203125
        },
        "honest": {
          "audit_correct": 508,
          "audit_total": 1024,
          "feature_index": 0,
          "polarity": -1,
          "selection_correct": 133,
          "selection_total": 256,
          "test_correct": 126,
          "test_total": 256
        },
        "k": 2,
        "leaky": {
          "audit_correct": 507,
          "audit_total": 1024,
          "feature_index": 1,
          "polarity": -1,
          "test_correct": 134,
          "test_total": 256
        },
        "seed": 25
      },
      {
        "candidate_pair_count": 8,
        "derived": {
          "honest_audit_accuracy": 0.486328125,
          "honest_selection_accuracy": 0.55078125,
          "honest_test_accuracy": 0.4921875,
          "leaky_audit_accuracy": 0.4970703125,
          "leaky_reported_test_accuracy": 0.55078125,
          "leaky_vs_honest_test_gap": 0.05859375,
          "optimism_gap": 0.0537109375
        },
        "honest": {
          "audit_correct": 498,
          "audit_total": 1024,
          "feature_index": 2,
          "polarity": -1,
          "selection_correct": 141,
          "selection_total": 256,
          "test_correct": 126,
          "test_total": 256
        },
        "k": 4,
        "leaky": {
          "audit_correct": 509,
          "audit_total": 1024,
          "feature_index": 3,
          "polarity": 1,
          "test_correct": 141,
          "test_total": 256
        },
        "seed": 25
      },
      {
        "candidate_pair_count": 16,
        "derived": {
          "honest_audit_accuracy": 0.4912109375,
          "honest_selection_accuracy": 0.5546875,
          "honest_test_accuracy": 0.4921875,
          "leaky_audit_accuracy": 0.4970703125,
          "leaky_reported_test_accuracy": 0.55078125,
          "leaky_vs_honest_test_gap": 0.05859375,
          "optimism_gap": 0.0537109375
        },
        "honest": {
          "audit_correct": 503,
          "audit_total": 1024,
          "feature_index": 5,
          "polarity": 1,
          "selection_correct": 142,
          "selection_total": 256,
          "test_correct": 126,
          "test_total": 256
        },
        "k": 8,
        "leaky": {
          "audit_correct": 509,
          "audit_total": 1024,
          "feature_index": 3,
          "polarity": 1,
          "test_correct": 141,
          "test_total": 256
        },
        "seed": 25
      },
      {
        "candidate_pair_count": 32,
        "derived": {
          "honest_audit_accuracy": 0.4912109375,
          "honest_selection_accuracy": 0.5546875,
          "honest_test_accuracy": 0.4921875,
          "leaky_audit_accuracy": 0.4970703125,
          "leaky_reported_test_accuracy": 0.55078125,
          "leaky_vs_honest_test_gap": 0.05859375,
          "optimism_gap": 0.0537109375
        },
        "honest": {
          "audit_correct": 503,
          "audit_total": 1024,
          "feature_index": 5,
          "polarity": 1,
          "selection_correct": 142,
          "selection_total": 256,
          "test_correct": 126,
          "test_total": 256
        },
        "k": 16,
        "leaky": {
          "audit_correct": 509,
          "audit_total": 1024,
          "feature_index": 3,
          "polarity": 1,
          "test_correct": 141,
          "test_total": 256
        },
        "seed": 25
      },
      {
        "candidate_pair_count": 64,
        "derived": {
          "honest_audit_accuracy": 0.4833984375,
          "honest_selection_accuracy": 0.57421875,
          "honest_test_accuracy": 0.48828125,
          "leaky_audit_accuracy": 0.4970703125,
          "leaky_reported_test_accuracy": 0.55078125,
          "leaky_vs_honest_test_gap": 0.0625,
          "optimism_gap": 0.0537109375
        },
        "honest": {
          "audit_correct": 495,
          "audit_total": 1024,
          "feature_index": 25,
          "polarity": 1,
          "selection_correct": 147,
          "selection_total": 256,
          "test_correct": 125,
          "test_total": 256
        },
        "k": 32,
        "leaky": {
          "audit_correct": 509,
          "audit_total": 1024,
          "feature_index": 3,
          "polarity": 1,
          "test_correct": 141,
          "test_total": 256
        },
        "seed": 25
      },
      {
        "candidate_pair_count": 128,
        "derived": {
          "honest_audit_accuracy": 0.484375,
          "honest_selection_accuracy": 0.5859375,
          "honest_test_accuracy": 0.44921875,
          "leaky_audit_accuracy": 0.50390625,
          "leaky_reported_test_accuracy": 0.56640625,
          "leaky_vs_honest_test_gap": 0.1171875,
          "optimism_gap": 0.0625
        },
        "honest": {
          "audit_correct": 496,
          "audit_total": 1024,
          "feature_index": 55,
          "polarity": 1,
          "selection_correct": 150,
          "selection_total": 256,
          "test_correct": 115,
          "test_total": 256
        },
        "k": 64,
        "leaky": {
          "audit_correct": 516,
          "audit_total": 1024,
          "feature_index": 54,
          "polarity": 1,
          "test_correct": 145,
          "test_total": 256
        },
        "seed": 25
      },
      {
        "candidate_pair_count": 256,
        "derived": {
          "honest_audit_accuracy": 0.484375,
          "honest_selection_accuracy": 0.5859375,
          "honest_test_accuracy": 0.44921875,
          "leaky_audit_accuracy": 0.494140625,
          "leaky_reported_test_accuracy": 0.60546875,
          "leaky_vs_honest_test_gap": 0.15625,
          "optimism_gap": 0.111328125
        },
        "honest": {
          "audit_correct": 496,
          "audit_total": 1024,
          "feature_index": 55,
          "polarity": 1,
          "selection_correct": 150,
          "selection_total": 256,
          "test_correct": 115,
          "test_total": 256
        },
        "k": 128,
        "leaky": {
          "audit_correct": 506,
          "audit_total": 1024,
          "feature_index": 78,
          "polarity": 1,
          "test_correct": 155,
          "test_total": 256
        },
        "seed": 25
      },
      {
        "candidate_pair_count": 512,
        "derived": {
          "honest_audit_accuracy": 0.484375,
          "honest_selection_accuracy": 0.59765625,
          "honest_test_accuracy": 0.53515625,
          "leaky_audit_accuracy": 0.494140625,
          "leaky_reported_test_accuracy": 0.60546875,
          "leaky_vs_honest_test_gap": 0.0703125,
          "optimism_gap": 0.111328125
        },
        "honest": {
          "audit_correct": 496,
          "audit_total": 1024,
          "feature_index": 250,
          "polarity": 1,
          "selection_correct": 153,
          "selection_total": 256,
          "test_correct": 137,
          "test_total": 256
        },
        "k": 256,
        "leaky": {
          "audit_correct": 506,
          "audit_total": 1024,
          "feature_index": 78,
          "polarity": 1,
          "test_correct": 155,
          "test_total": 256
        },
        "seed": 25
      },
      {
        "candidate_pair_count": 1024,
        "derived": {
          "honest_audit_accuracy": 0.484375,
          "honest_selection_accuracy": 0.59765625,
          "honest_test_accuracy": 0.53515625,
          "leaky_audit_accuracy": 0.494140625,
          "leaky_reported_test_accuracy": 0.60546875,
          "leaky_vs_honest_test_gap": 0.0703125,
          "optimism_gap": 0.111328125
        },
        "honest": {
          "audit_correct": 496,
          "audit_total": 1024,
          "feature_index": 250,
          "polarity": 1,
          "selection_correct": 153,
          "selection_total": 256,
          "test_correct": 137,
          "test_total": 256
        },
        "k": 512,
        "leaky": {
          "audit_correct": 506,
          "audit_total": 1024,
          "feature_index": 78,
          "polarity": 1,
          "test_correct": 155,
          "test_total": 256
        },
        "seed": 25
      },
      {
        "candidate_pair_count": 2048,
        "derived": {
          "honest_audit_accuracy": 0.49609375,
          "honest_selection_accuracy": 0.61328125,
          "honest_test_accuracy": 0.5234375,
          "leaky_audit_accuracy": 0.494140625,
          "leaky_reported_test_accuracy": 0.60546875,
          "leaky_vs_honest_test_gap": 0.08203125,
          "optimism_gap": 0.111328125
        },
        "honest": {
          "audit_correct": 508,
          "audit_total": 1024,
          "feature_index": 659,
          "polarity": 1,
          "selection_correct": 157,
          "selection_total": 256,
          "test_correct": 134,
          "test_total": 256
        },
        "k": 1024,
        "leaky": {
          "audit_correct": 506,
          "audit_total": 1024,
          "feature_index": 78,
          "polarity": 1,
          "test_correct": 155,
          "test_total": 256
        },
        "seed": 25
      },
      {
        "candidate_pair_count": 2,
        "derived": {
          "honest_audit_accuracy": 0.4814453125,
          "honest_selection_accuracy": 0.5,
          "honest_test_accuracy": 0.51171875,
          "leaky_audit_accuracy": 0.4814453125,
          "leaky_reported_test_accuracy": 0.51171875,
          "leaky_vs_honest_test_gap": 0.0,
          "optimism_gap": 0.0302734375
        },
        "honest": {
          "audit_correct": 493,
          "audit_total": 1024,
          "feature_index": 0,
          "polarity": 1,
          "selection_correct": 128,
          "selection_total": 256,
          "test_correct": 131,
          "test_total": 256
        },
        "k": 1,
        "leaky": {
          "audit_correct": 493,
          "audit_total": 1024,
          "feature_index": 0,
          "polarity": 1,
          "test_correct": 131,
          "test_total": 256
        },
        "seed": 26
      },
      {
        "candidate_pair_count": 4,
        "derived": {
          "honest_audit_accuracy": 0.5166015625,
          "honest_selection_accuracy": 0.5078125,
          "honest_test_accuracy": 0.46484375,
          "leaky_audit_accuracy": 0.4833984375,
          "leaky_reported_test_accuracy": 0.53515625,
          "leaky_vs_honest_test_gap": 0.0703125,
          "optimism_gap": 0.0517578125
        },
        "honest": {
          "audit_correct": 529,
          "audit_total": 1024,
          "feature_index": 1,
          "polarity": 1,
          "selection_correct": 130,
          "selection_total": 256,
          "test_correct": 119,
          "test_total": 256
        },
        "k": 2,
        "leaky": {
          "audit_correct": 495,
          "audit_total": 1024,
          "feature_index": 1,
          "polarity": -1,
          "test_correct": 137,
          "test_total": 256
        },
        "seed": 26
      },
      {
        "candidate_pair_count": 8,
        "derived": {
          "honest_audit_accuracy": 0.5087890625,
          "honest_selection_accuracy": 0.53125,
          "honest_test_accuracy": 0.5390625,
          "leaky_audit_accuracy": 0.4912109375,
          "leaky_reported_test_accuracy": 0.54296875,
          "leaky_vs_honest_test_gap": 0.00390625,
          "optimism_gap": 0.0517578125
        },
        "honest": {
          "audit_correct": 521,
          "audit_total": 1024,
          "feature_index": 2,
          "polarity": 1,
          "selection_correct": 136,
          "selection_total": 256,
          "test_correct": 138,
          "test_total": 256
        },
        "k": 4,
        "leaky": {
          "audit_correct": 503,
          "audit_total": 1024,
          "feature_index": 3,
          "polarity": -1,
          "test_correct": 139,
          "test_total": 256
        },
        "seed": 26
      },
      {
        "candidate_pair_count": 16,
        "derived": {
          "honest_audit_accuracy": 0.48046875,
          "honest_selection_accuracy": 0.5625,
          "honest_test_accuracy": 0.53515625,
          "leaky_audit_accuracy": 0.490234375,
          "leaky_reported_test_accuracy": 0.546875,
          "leaky_vs_honest_test_gap": 0.01171875,
          "optimism_gap": 0.056640625
        },
        "honest": {
          "audit_correct": 492,
          "audit_total": 1024,
          "feature_index": 7,
          "polarity": 1,
          "selection_correct": 144,
          "selection_total": 256,
          "test_correct": 137,
          "test_total": 256
        },
        "k": 8,
        "leaky": {
          "audit_correct": 502,
          "audit_total": 1024,
          "feature_index": 4,
          "polarity": -1,
          "test_correct": 140,
          "test_total": 256
        },
        "seed": 26
      },
      {
        "candidate_pair_count": 32,
        "derived": {
          "honest_audit_accuracy": 0.48046875,
          "honest_selection_accuracy": 0.5625,
          "honest_test_accuracy": 0.53515625,
          "leaky_audit_accuracy": 0.5205078125,
          "leaky_reported_test_accuracy": 0.5625,
          "leaky_vs_honest_test_gap": 0.02734375,
          "optimism_gap": 0.0419921875
        },
        "honest": {
          "audit_correct": 492,
          "audit_total": 1024,
          "feature_index": 7,
          "polarity": 1,
          "selection_correct": 144,
          "selection_total": 256,
          "test_correct": 137,
          "test_total": 256
        },
        "k": 16,
        "leaky": {
          "audit_correct": 533,
          "audit_total": 1024,
          "feature_index": 10,
          "polarity": -1,
          "test_correct": 144,
          "test_total": 256
        },
        "seed": 26
      },
      {
        "candidate_pair_count": 64,
        "derived": {
          "honest_audit_accuracy": 0.5078125,
          "honest_selection_accuracy": 0.5703125,
          "honest_test_accuracy": 0.50390625,
          "leaky_audit_accuracy": 0.53515625,
          "leaky_reported_test_accuracy": 0.59765625,
          "leaky_vs_honest_test_gap": 0.09375,
          "optimism_gap": 0.0625
        },
        "honest": {
          "audit_correct": 520,
          "audit_total": 1024,
          "feature_index": 24,
          "polarity": -1,
          "selection_correct": 146,
          "selection_total": 256,
          "test_correct": 129,
          "test_total": 256
        },
        "k": 32,
        "leaky": {
          "audit_correct": 548,
          "audit_total": 1024,
          "feature_index": 19,
          "polarity": -1,
          "test_correct": 153,
          "test_total": 256
        },
        "seed": 26
      },
      {
        "candidate_pair_count": 128,
        "derived": {
          "honest_audit_accuracy": 0.4814453125,
          "honest_selection_accuracy": 0.58984375,
          "honest_test_accuracy": 0.46875,
          "leaky_audit_accuracy": 0.53515625,
          "leaky_reported_test_accuracy": 0.59765625,
          "leaky_vs_honest_test_gap": 0.12890625,
          "optimism_gap": 0.0625
        },
        "honest": {
          "audit_correct": 493,
          "audit_total": 1024,
          "feature_index": 45,
          "polarity": 1,
          "selection_correct": 151,
          "selection_total": 256,
          "test_correct": 120,
          "test_total": 256
        },
        "k": 64,
        "leaky": {
          "audit_correct": 548,
          "audit_total": 1024,
          "feature_index": 19,
          "polarity": -1,
          "test_correct": 153,
          "test_total": 256
        },
        "seed": 26
      },
      {
        "candidate_pair_count": 256,
        "derived": {
          "honest_audit_accuracy": 0.4814453125,
          "honest_selection_accuracy": 0.58984375,
          "honest_test_accuracy": 0.46875,
          "leaky_audit_accuracy": 0.53515625,
          "leaky_reported_test_accuracy": 0.59765625,
          "leaky_vs_honest_test_gap": 0.12890625,
          "optimism_gap": 0.0625
        },
        "honest": {
          "audit_correct": 493,
          "audit_total": 1024,
          "feature_index": 45,
          "polarity": 1,
          "selection_correct": 151,
          "selection_total": 256,
          "test_correct": 120,
          "test_total": 256
        },
        "k": 128,
        "leaky": {
          "audit_correct": 548,
          "audit_total": 1024,
          "feature_index": 19,
          "polarity": -1,
          "test_correct": 153,
          "test_total": 256
        },
        "seed": 26
      },
      {
        "candidate_pair_count": 512,
        "derived": {
          "honest_audit_accuracy": 0.4814453125,
          "honest_selection_accuracy": 0.58984375,
          "honest_test_accuracy": 0.46875,
          "leaky_audit_accuracy": 0.53515625,
          "leaky_reported_test_accuracy": 0.59765625,
          "leaky_vs_honest_test_gap": 0.12890625,
          "optimism_gap": 0.0625
        },
        "honest": {
          "audit_correct": 493,
          "audit_total": 1024,
          "feature_index": 45,
          "polarity": 1,
          "selection_correct": 151,
          "selection_total": 256,
          "test_correct": 120,
          "test_total": 256
        },
        "k": 256,
        "leaky": {
          "audit_correct": 548,
          "audit_total": 1024,
          "feature_index": 19,
          "polarity": -1,
          "test_correct": 153,
          "test_total": 256
        },
        "seed": 26
      },
      {
        "candidate_pair_count": 1024,
        "derived": {
          "honest_audit_accuracy": 0.5,
          "honest_selection_accuracy": 0.59375,
          "honest_test_accuracy": 0.49609375,
          "leaky_audit_accuracy": 0.53515625,
          "leaky_reported_test_accuracy": 0.59765625,
          "leaky_vs_honest_test_gap": 0.1015625,
          "optimism_gap": 0.0625
        },
        "honest": {
          "audit_correct": 512,
          "audit_total": 1024,
          "feature_index": 339,
          "polarity": 1,
          "selection_correct": 152,
          "selection_total": 256,
          "test_correct": 127,
          "test_total": 256
        },
        "k": 512,
        "leaky": {
          "audit_correct": 548,
          "audit_total": 1024,
          "feature_index": 19,
          "polarity": -1,
          "test_correct": 153,
          "test_total": 256
        },
        "seed": 26
      },
      {
        "candidate_pair_count": 2048,
        "derived": {
          "honest_audit_accuracy": 0.5,
          "honest_selection_accuracy": 0.59375,
          "honest_test_accuracy": 0.49609375,
          "leaky_audit_accuracy": 0.4970703125,
          "leaky_reported_test_accuracy": 0.6171875,
          "leaky_vs_honest_test_gap": 0.12109375,
          "optimism_gap": 0.1201171875
        },
        "honest": {
          "audit_correct": 512,
          "audit_total": 1024,
          "feature_index": 339,
          "polarity": 1,
          "selection_correct": 152,
          "selection_total": 256,
          "test_correct": 127,
          "test_total": 256
        },
        "k": 1024,
        "leaky": {
          "audit_correct": 509,
          "audit_total": 1024,
          "feature_index": 1020,
          "polarity": 1,
          "test_correct": 158,
          "test_total": 256
        },
        "seed": 26
      },
      {
        "candidate_pair_count": 2,
        "derived": {
          "honest_audit_accuracy": 0.501953125,
          "honest_selection_accuracy": 0.5234375,
          "honest_test_accuracy": 0.55859375,
          "leaky_audit_accuracy": 0.501953125,
          "leaky_reported_test_accuracy": 0.55859375,
          "leaky_vs_honest_test_gap": 0.0,
          "optimism_gap": 0.056640625
        },
        "honest": {
          "audit_correct": 514,
          "audit_total": 1024,
          "feature_index": 0,
          "polarity": -1,
          "selection_correct": 134,
          "selection_total": 256,
          "test_correct": 143,
          "test_total": 256
        },
        "k": 1,
        "leaky": {
          "audit_correct": 514,
          "audit_total": 1024,
          "feature_index": 0,
          "polarity": -1,
          "test_correct": 143,
          "test_total": 256
        },
        "seed": 27
      },
      {
        "candidate_pair_count": 4,
        "derived": {
          "honest_audit_accuracy": 0.501953125,
          "honest_selection_accuracy": 0.5234375,
          "honest_test_accuracy": 0.55859375,
          "leaky_audit_accuracy": 0.501953125,
          "leaky_reported_test_accuracy": 0.55859375,
          "leaky_vs_honest_test_gap": 0.0,
          "optimism_gap": 0.056640625
        },
        "honest": {
          "audit_correct": 514,
          "audit_total": 1024,
          "feature_index": 0,
          "polarity": -1,
          "selection_correct": 134,
          "selection_total": 256,
          "test_correct": 143,
          "test_total": 256
        },
        "k": 2,
        "leaky": {
          "audit_correct": 514,
          "audit_total": 1024,
          "feature_index": 0,
          "polarity": -1,
          "test_correct": 143,
          "test_total": 256
        },
        "seed": 27
      },
      {
        "candidate_pair_count": 8,
        "derived": {
          "honest_audit_accuracy": 0.501953125,
          "honest_selection_accuracy": 0.5234375,
          "honest_test_accuracy": 0.55859375,
          "leaky_audit_accuracy": 0.501953125,
          "leaky_reported_test_accuracy": 0.55859375,
          "leaky_vs_honest_test_gap": 0.0,
          "optimism_gap": 0.056640625
        },
        "honest": {
          "audit_correct": 514,
          "audit_total": 1024,
          "feature_index": 0,
          "polarity": -1,
          "selection_correct": 134,
          "selection_total": 256,
          "test_correct": 143,
          "test_total": 256
        },
        "k": 4,
        "leaky": {
          "audit_correct": 514,
          "audit_total": 1024,
          "feature_index": 0,
          "polarity": -1,
          "test_correct": 143,
          "test_total": 256
        },
        "seed": 27
      },
      {
        "candidate_pair_count": 16,
        "derived": {
          "honest_audit_accuracy": 0.498046875,
          "honest_selection_accuracy": 0.53125,
          "honest_test_accuracy": 0.5703125,
          "leaky_audit_accuracy": 0.498046875,
          "leaky_reported_test_accuracy": 0.5703125,
          "leaky_vs_honest_test_gap": 0.0,
          "optimism_gap": 0.072265625
        },
        "honest": {
          "audit_correct": 510,
          "audit_total": 1024,
          "feature_index": 5,
          "polarity": 1,
          "selection_correct": 136,
          "selection_total": 256,
          "test_correct": 146,
          "test_total": 256
        },
        "k": 8,
        "leaky": {
          "audit_correct": 510,
          "audit_total": 1024,
          "feature_index": 5,
          "polarity": 1,
          "test_correct": 146,
          "test_total": 256
        },
        "seed": 27
      },
      {
        "candidate_pair_count": 32,
        "derived": {
          "honest_audit_accuracy": 0.4990234375,
          "honest_selection_accuracy": 0.5546875,
          "honest_test_accuracy": 0.484375,
          "leaky_audit_accuracy": 0.498046875,
          "leaky_reported_test_accuracy": 0.5703125,
          "leaky_vs_honest_test_gap": 0.0859375,
          "optimism_gap": 0.072265625
        },
        "honest": {
          "audit_correct": 511,
          "audit_total": 1024,
          "feature_index": 13,
          "polarity": -1,
          "selection_correct": 142,
          "selection_total": 256,
          "test_correct": 124,
          "test_total": 256
        },
        "k": 16,
        "leaky": {
          "audit_correct": 510,
          "audit_total": 1024,
          "feature_index": 5,
          "polarity": 1,
          "test_correct": 146,
          "test_total": 256
        },
        "seed": 27
      },
      {
        "candidate_pair_count": 64,
        "derived": {
          "honest_audit_accuracy": 0.4970703125,
          "honest_selection_accuracy": 0.55859375,
          "honest_test_accuracy": 0.5,
          "leaky_audit_accuracy": 0.498046875,
          "leaky_reported_test_accuracy": 0.5703125,
          "leaky_vs_honest_test_gap": 0.0703125,
          "optimism_gap": 0.072265625
        },
        "honest": {
          "audit_correct": 509,
          "audit_total": 1024,
          "feature_index": 17,
          "polarity": 1,
          "selection_correct": 143,
          "selection_total": 256,
          "test_correct": 128,
          "test_total": 256
        },
        "k": 32,
        "leaky": {
          "audit_correct": 510,
          "audit_total": 1024,
          "feature_index": 5,
          "polarity": 1,
          "test_correct": 146,
          "test_total": 256
        },
        "seed": 27
      },
      {
        "candidate_pair_count": 128,
        "derived": {
          "honest_audit_accuracy": 0.4931640625,
          "honest_selection_accuracy": 0.58984375,
          "honest_test_accuracy": 0.51171875,
          "leaky_audit_accuracy": 0.498046875,
          "leaky_reported_test_accuracy": 0.5703125,
          "leaky_vs_honest_test_gap": 0.05859375,
          "optimism_gap": 0.072265625
        },
        "honest": {
          "audit_correct": 505,
          "audit_total": 1024,
          "feature_index": 35,
          "polarity": -1,
          "selection_correct": 151,
          "selection_total": 256,
          "test_correct": 131,
          "test_total": 256
        },
        "k": 64,
        "leaky": {
          "audit_correct": 510,
          "audit_total": 1024,
          "feature_index": 5,
          "polarity": 1,
          "test_correct": 146,
          "test_total": 256
        },
        "seed": 27
      },
      {
        "candidate_pair_count": 256,
        "derived": {
          "honest_audit_accuracy": 0.4931640625,
          "honest_selection_accuracy": 0.58984375,
          "honest_test_accuracy": 0.51171875,
          "leaky_audit_accuracy": 0.494140625,
          "leaky_reported_test_accuracy": 0.58203125,
          "leaky_vs_honest_test_gap": 0.0703125,
          "optimism_gap": 0.087890625
        },
        "honest": {
          "audit_correct": 505,
          "audit_total": 1024,
          "feature_index": 35,
          "polarity": -1,
          "selection_correct": 151,
          "selection_total": 256,
          "test_correct": 131,
          "test_total": 256
        },
        "k": 128,
        "leaky": {
          "audit_correct": 506,
          "audit_total": 1024,
          "feature_index": 75,
          "polarity": -1,
          "test_correct": 149,
          "test_total": 256
        },
        "seed": 27
      },
      {
        "candidate_pair_count": 512,
        "derived": {
          "honest_audit_accuracy": 0.4931640625,
          "honest_selection_accuracy": 0.58984375,
          "honest_test_accuracy": 0.51171875,
          "leaky_audit_accuracy": 0.5107421875,
          "leaky_reported_test_accuracy": 0.58984375,
          "leaky_vs_honest_test_gap": 0.078125,
          "optimism_gap": 0.0791015625
        },
        "honest": {
          "audit_correct": 505,
          "audit_total": 1024,
          "feature_index": 35,
          "polarity": -1,
          "selection_correct": 151,
          "selection_total": 256,
          "test_correct": 131,
          "test_total": 256
        },
        "k": 256,
        "leaky": {
          "audit_correct": 523,
          "audit_total": 1024,
          "feature_index": 174,
          "polarity": -1,
          "test_correct": 151,
          "test_total": 256
        },
        "seed": 27
      },
      {
        "candidate_pair_count": 1024,
        "derived": {
          "honest_audit_accuracy": 0.4931640625,
          "honest_selection_accuracy": 0.58984375,
          "honest_test_accuracy": 0.51171875,
          "leaky_audit_accuracy": 0.4873046875,
          "leaky_reported_test_accuracy": 0.60546875,
          "leaky_vs_honest_test_gap": 0.09375,
          "optimism_gap": 0.1181640625
        },
        "honest": {
          "audit_correct": 505,
          "audit_total": 1024,
          "feature_index": 35,
          "polarity": -1,
          "selection_correct": 151,
          "selection_total": 256,
          "test_correct": 131,
          "test_total": 256
        },
        "k": 512,
        "leaky": {
          "audit_correct": 499,
          "audit_total": 1024,
          "feature_index": 434,
          "polarity": -1,
          "test_correct": 155,
          "test_total": 256
        },
        "seed": 27
      },
      {
        "candidate_pair_count": 2048,
        "derived": {
          "honest_audit_accuracy": 0.5205078125,
          "honest_selection_accuracy": 0.6015625,
          "honest_test_accuracy": 0.5234375,
          "leaky_audit_accuracy": 0.4873046875,
          "leaky_reported_test_accuracy": 0.60546875,
          "leaky_vs_honest_test_gap": 0.08203125,
          "optimism_gap": 0.1181640625
        },
        "honest": {
          "audit_correct": 533,
          "audit_total": 1024,
          "feature_index": 541,
          "polarity": 1,
          "selection_correct": 154,
          "selection_total": 256,
          "test_correct": 134,
          "test_total": 256
        },
        "k": 1024,
        "leaky": {
          "audit_correct": 499,
          "audit_total": 1024,
          "feature_index": 434,
          "polarity": -1,
          "test_correct": 155,
          "test_total": 256
        },
        "seed": 27
      },
      {
        "candidate_pair_count": 2,
        "derived": {
          "honest_audit_accuracy": 0.5283203125,
          "honest_selection_accuracy": 0.50390625,
          "honest_test_accuracy": 0.44921875,
          "leaky_audit_accuracy": 0.4716796875,
          "leaky_reported_test_accuracy": 0.55078125,
          "leaky_vs_honest_test_gap": 0.1015625,
          "optimism_gap": 0.0791015625
        },
        "honest": {
          "audit_correct": 541,
          "audit_total": 1024,
          "feature_index": 0,
          "polarity": 1,
          "selection_correct": 129,
          "selection_total": 256,
          "test_correct": 115,
          "test_total": 256
        },
        "k": 1,
        "leaky": {
          "audit_correct": 483,
          "audit_total": 1024,
          "feature_index": 0,
          "polarity": -1,
          "test_correct": 141,
          "test_total": 256
        },
        "seed": 28
      },
      {
        "candidate_pair_count": 4,
        "derived": {
          "honest_audit_accuracy": 0.4990234375,
          "honest_selection_accuracy": 0.546875,
          "honest_test_accuracy": 0.49609375,
          "leaky_audit_accuracy": 0.4716796875,
          "leaky_reported_test_accuracy": 0.55078125,
          "leaky_vs_honest_test_gap": 0.0546875,
          "optimism_gap": 0.0791015625
        },
        "honest": {
          "audit_correct": 511,
          "audit_total": 1024,
          "feature_index": 1,
          "polarity": -1,
          "selection_correct": 140,
          "selection_total": 256,
          "test_correct": 127,
          "test_total": 256
        },
        "k": 2,
        "leaky": {
          "audit_correct": 483,
          "audit_total": 1024,
          "feature_index": 0,
          "polarity": -1,
          "test_correct": 141,
          "test_total": 256
        },
        "seed": 28
      },
      {
        "candidate_pair_count": 8,
        "derived": {
          "honest_audit_accuracy": 0.4990234375,
          "honest_selection_accuracy": 0.55859375,
          "honest_test_accuracy": 0.5234375,
          "leaky_audit_accuracy": 0.4716796875,
          "leaky_reported_test_accuracy": 0.55078125,
          "leaky_vs_honest_test_gap": 0.02734375,
          "optimism_gap": 0.0791015625
        },
        "honest": {
          "audit_correct": 511,
          "audit_total": 1024,
          "feature_index": 3,
          "polarity": -1,
          "selection_correct": 143,
          "selection_total": 256,
          "test_correct": 134,
          "test_total": 256
        },
        "k": 4,
        "leaky": {
          "audit_correct": 483,
          "audit_total": 1024,
          "feature_index": 0,
          "polarity": -1,
          "test_correct": 141,
          "test_total": 256
        },
        "seed": 28
      },
      {
        "candidate_pair_count": 16,
        "derived": {
          "honest_audit_accuracy": 0.4990234375,
          "honest_selection_accuracy": 0.55859375,
          "honest_test_accuracy": 0.5234375,
          "leaky_audit_accuracy": 0.4716796875,
          "leaky_reported_test_accuracy": 0.55078125,
          "leaky_vs_honest_test_gap": 0.02734375,
          "optimism_gap": 0.0791015625
        },
        "honest": {
          "audit_correct": 511,
          "audit_total": 1024,
          "feature_index": 3,
          "polarity": -1,
          "selection_correct": 143,
          "selection_total": 256,
          "test_correct": 134,
          "test_total": 256
        },
        "k": 8,
        "leaky": {
          "audit_correct": 483,
          "audit_total": 1024,
          "feature_index": 0,
          "polarity": -1,
          "test_correct": 141,
          "test_total": 256
        },
        "seed": 28
      },
      {
        "candidate_pair_count": 32,
        "derived": {
          "honest_audit_accuracy": 0.4990234375,
          "honest_selection_accuracy": 0.55859375,
          "honest_test_accuracy": 0.5234375,
          "leaky_audit_accuracy": 0.4931640625,
          "leaky_reported_test_accuracy": 0.57421875,
          "leaky_vs_honest_test_gap": 0.05078125,
          "optimism_gap": 0.0810546875
        },
        "honest": {
          "audit_correct": 511,
          "audit_total": 1024,
          "feature_index": 3,
          "polarity": -1,
          "selection_correct": 143,
          "selection_total": 256,
          "test_correct": 134,
          "test_total": 256
        },
        "k": 16,
        "leaky": {
          "audit_correct": 505,
          "audit_total": 1024,
          "feature_index": 9,
          "polarity": -1,
          "test_correct": 147,
          "test_total": 256
        },
        "seed": 28
      },
      {
        "candidate_pair_count": 64,
        "derived": {
          "honest_audit_accuracy": 0.4990234375,
          "honest_selection_accuracy": 0.55859375,
          "honest_test_accuracy": 0.5234375,
          "leaky_audit_accuracy": 0.4931640625,
          "leaky_reported_test_accuracy": 0.57421875,
          "leaky_vs_honest_test_gap": 0.05078125,
          "optimism_gap": 0.0810546875
        },
        "honest": {
          "audit_correct": 511,
          "audit_total": 1024,
          "feature_index": 3,
          "polarity": -1,
          "selection_correct": 143,
          "selection_total": 256,
          "test_correct": 134,
          "test_total": 256
        },
        "k": 32,
        "leaky": {
          "audit_correct": 505,
          "audit_total": 1024,
          "feature_index": 9,
          "polarity": -1,
          "test_correct": 147,
          "test_total": 256
        },
        "seed": 28
      },
      {
        "candidate_pair_count": 128,
        "derived": {
          "honest_audit_accuracy": 0.515625,
          "honest_selection_accuracy": 0.57421875,
          "honest_test_accuracy": 0.49609375,
          "leaky_audit_accuracy": 0.501953125,
          "leaky_reported_test_accuracy": 0.578125,
          "leaky_vs_honest_test_gap": 0.08203125,
          "optimism_gap": 0.076171875
        },
        "honest": {
          "audit_correct": 528,
          "audit_total": 1024,
          "feature_index": 42,
          "polarity": -1,
          "selection_correct": 147,
          "selection_total": 256,
          "test_correct": 127,
          "test_total": 256
        },
        "k": 64,
        "leaky": {
          "audit_correct": 514,
          "audit_total": 1024,
          "feature_index": 62,
          "polarity": -1,
          "test_correct": 148,
          "test_total": 256
        },
        "seed": 28
      },
      {
        "candidate_pair_count": 256,
        "derived": {
          "honest_audit_accuracy": 0.5341796875,
          "honest_selection_accuracy": 0.578125,
          "honest_test_accuracy": 0.54296875,
          "leaky_audit_accuracy": 0.501953125,
          "leaky_reported_test_accuracy": 0.578125,
          "leaky_vs_honest_test_gap": 0.03515625,
          "optimism_gap": 0.076171875
        },
        "honest": {
          "audit_correct": 547,
          "audit_total": 1024,
          "feature_index": 86,
          "polarity": 1,
          "selection_correct": 148,
          "selection_total": 256,
          "test_correct": 139,
          "test_total": 256
        },
        "k": 128,
        "leaky": {
          "audit_correct": 514,
          "audit_total": 1024,
          "feature_index": 62,
          "polarity": -1,
          "test_correct": 148,
          "test_total": 256
        },
        "seed": 28
      },
      {
        "candidate_pair_count": 512,
        "derived": {
          "honest_audit_accuracy": 0.509765625,
          "honest_selection_accuracy": 0.59375,
          "honest_test_accuracy": 0.51171875,
          "leaky_audit_accuracy": 0.4892578125,
          "leaky_reported_test_accuracy": 0.59375,
          "leaky_vs_honest_test_gap": 0.08203125,
          "optimism_gap": 0.1044921875
        },
        "honest": {
          "audit_correct": 522,
          "audit_total": 1024,
          "feature_index": 254,
          "polarity": -1,
          "selection_correct": 152,
          "selection_total": 256,
          "test_correct": 131,
          "test_total": 256
        },
        "k": 256,
        "leaky": {
          "audit_correct": 501,
          "audit_total": 1024,
          "feature_index": 206,
          "polarity": -1,
          "test_correct": 152,
          "test_total": 256
        },
        "seed": 28
      },
      {
        "candidate_pair_count": 1024,
        "derived": {
          "honest_audit_accuracy": 0.5087890625,
          "honest_selection_accuracy": 0.59765625,
          "honest_test_accuracy": 0.5078125,
          "leaky_audit_accuracy": 0.5068359375,
          "leaky_reported_test_accuracy": 0.6171875,
          "leaky_vs_honest_test_gap": 0.109375,
          "optimism_gap": 0.1103515625
        },
        "honest": {
          "audit_correct": 521,
          "audit_total": 1024,
          "feature_index": 291,
          "polarity": -1,
          "selection_correct": 153,
          "selection_total": 256,
          "test_correct": 130,
          "test_total": 256
        },
        "k": 512,
        "leaky": {
          "audit_correct": 519,
          "audit_total": 1024,
          "feature_index": 369,
          "polarity": -1,
          "test_correct": 158,
          "test_total": 256
        },
        "seed": 28
      },
      {
        "candidate_pair_count": 2048,
        "derived": {
          "honest_audit_accuracy": 0.4951171875,
          "honest_selection_accuracy": 0.60546875,
          "honest_test_accuracy": 0.50390625,
          "leaky_audit_accuracy": 0.5068359375,
          "leaky_reported_test_accuracy": 0.6171875,
          "leaky_vs_honest_test_gap": 0.11328125,
          "optimism_gap": 0.1103515625
        },
        "honest": {
          "audit_correct": 507,
          "audit_total": 1024,
          "feature_index": 744,
          "polarity": 1,
          "selection_correct": 155,
          "selection_total": 256,
          "test_correct": 129,
          "test_total": 256
        },
        "k": 1024,
        "leaky": {
          "audit_correct": 519,
          "audit_total": 1024,
          "feature_index": 369,
          "polarity": -1,
          "test_correct": 158,
          "test_total": 256
        },
        "seed": 28
      },
      {
        "candidate_pair_count": 2,
        "derived": {
          "honest_audit_accuracy": 0.5029296875,
          "honest_selection_accuracy": 0.5,
          "honest_test_accuracy": 0.53515625,
          "leaky_audit_accuracy": 0.5029296875,
          "leaky_reported_test_accuracy": 0.53515625,
          "leaky_vs_honest_test_gap": 0.0,
          "optimism_gap": 0.0322265625
        },
        "honest": {
          "audit_correct": 515,
          "audit_total": 1024,
          "feature_index": 0,
          "polarity": 1,
          "selection_correct": 128,
          "selection_total": 256,
          "test_correct": 137,
          "test_total": 256
        },
        "k": 1,
        "leaky": {
          "audit_correct": 515,
          "audit_total": 1024,
          "feature_index": 0,
          "polarity": 1,
          "test_correct": 137,
          "test_total": 256
        },
        "seed": 29
      },
      {
        "candidate_pair_count": 4,
        "derived": {
          "honest_audit_accuracy": 0.49609375,
          "honest_selection_accuracy": 0.53125,
          "honest_test_accuracy": 0.56640625,
          "leaky_audit_accuracy": 0.49609375,
          "leaky_reported_test_accuracy": 0.56640625,
          "leaky_vs_honest_test_gap": 0.0,
          "optimism_gap": 0.0703125
        },
        "honest": {
          "audit_correct": 508,
          "audit_total": 1024,
          "feature_index": 1,
          "polarity": -1,
          "selection_correct": 136,
          "selection_total": 256,
          "test_correct": 145,
          "test_total": 256
        },
        "k": 2,
        "leaky": {
          "audit_correct": 508,
          "audit_total": 1024,
          "feature_index": 1,
          "polarity": -1,
          "test_correct": 145,
          "test_total": 256
        },
        "seed": 29
      },
      {
        "candidate_pair_count": 8,
        "derived": {
          "honest_audit_accuracy": 0.5029296875,
          "honest_selection_accuracy": 0.5390625,
          "honest_test_accuracy": 0.50390625,
          "leaky_audit_accuracy": 0.49609375,
          "leaky_reported_test_accuracy": 0.56640625,
          "leaky_vs_honest_test_gap": 0.0625,
          "optimism_gap": 0.0703125
        },
        "honest": {
          "audit_correct": 515,
          "audit_total": 1024,
          "feature_index": 3,
          "polarity": 1,
          "selection_correct": 138,
          "selection_total": 256,
          "test_correct": 129,
          "test_total": 256
        },
        "k": 4,
        "leaky": {
          "audit_correct": 508,
          "audit_total": 1024,
          "feature_index": 1,
          "polarity": -1,
          "test_correct": 145,
          "test_total": 256
        },
        "seed": 29
      },
      {
        "candidate_pair_count": 16,
        "derived": {
          "honest_audit_accuracy": 0.5029296875,
          "honest_selection_accuracy": 0.5390625,
          "honest_test_accuracy": 0.50390625,
          "leaky_audit_accuracy": 0.49609375,
          "leaky_reported_test_accuracy": 0.56640625,
          "leaky_vs_honest_test_gap": 0.0625,
          "optimism_gap": 0.0703125
        },
        "honest": {
          "audit_correct": 515,
          "audit_total": 1024,
          "feature_index": 3,
          "polarity": 1,
          "selection_correct": 138,
          "selection_total": 256,
          "test_correct": 129,
          "test_total": 256
        },
        "k": 8,
        "leaky": {
          "audit_correct": 508,
          "audit_total": 1024,
          "feature_index": 1,
          "polarity": -1,
          "test_correct": 145,
          "test_total": 256
        },
        "seed": 29
      },
      {
        "candidate_pair_count": 32,
        "derived": {
          "honest_audit_accuracy": 0.4921875,
          "honest_selection_accuracy": 0.5703125,
          "honest_test_accuracy": 0.47265625,
          "leaky_audit_accuracy": 0.49609375,
          "leaky_reported_test_accuracy": 0.56640625,
          "leaky_vs_honest_test_gap": 0.09375,
          "optimism_gap": 0.0703125
        },
        "honest": {
          "audit_correct": 504,
          "audit_total": 1024,
          "feature_index": 12,
          "polarity": 1,
          "selection_correct": 146,
          "selection_total": 256,
          "test_correct": 121,
          "test_total": 256
        },
        "k": 16,
        "leaky": {
          "audit_correct": 508,
          "audit_total": 1024,
          "feature_index": 1,
          "polarity": -1,
          "test_correct": 145,
          "test_total": 256
        },
        "seed": 29
      },
      {
        "candidate_pair_count": 64,
        "derived": {
          "honest_audit_accuracy": 0.4921875,
          "honest_selection_accuracy": 0.5703125,
          "honest_test_accuracy": 0.47265625,
          "leaky_audit_accuracy": 0.49609375,
          "leaky_reported_test_accuracy": 0.56640625,
          "leaky_vs_honest_test_gap": 0.09375,
          "optimism_gap": 0.0703125
        },
        "honest": {
          "audit_correct": 504,
          "audit_total": 1024,
          "feature_index": 12,
          "polarity": 1,
          "selection_correct": 146,
          "selection_total": 256,
          "test_correct": 121,
          "test_total": 256
        },
        "k": 32,
        "leaky": {
          "audit_correct": 508,
          "audit_total": 1024,
          "feature_index": 1,
          "polarity": -1,
          "test_correct": 145,
          "test_total": 256
        },
        "seed": 29
      },
      {
        "candidate_pair_count": 128,
        "derived": {
          "honest_audit_accuracy": 0.4794921875,
          "honest_selection_accuracy": 0.59765625,
          "honest_test_accuracy": 0.48046875,
          "leaky_audit_accuracy": 0.498046875,
          "leaky_reported_test_accuracy": 0.57421875,
          "leaky_vs_honest_test_gap": 0.09375,
          "optimism_gap": 0.076171875
        },
        "honest": {
          "audit_correct": 491,
          "audit_total": 1024,
          "feature_index": 33,
          "polarity": 1,
          "selection_correct": 153,
          "selection_total": 256,
          "test_correct": 123,
          "test_total": 256
        },
        "k": 64,
        "leaky": {
          "audit_correct": 510,
          "audit_total": 1024,
          "feature_index": 45,
          "polarity": 1,
          "test_correct": 147,
          "test_total": 256
        },
        "seed": 29
      },
      {
        "candidate_pair_count": 256,
        "derived": {
          "honest_audit_accuracy": 0.4794921875,
          "honest_selection_accuracy": 0.59765625,
          "honest_test_accuracy": 0.48046875,
          "leaky_audit_accuracy": 0.513671875,
          "leaky_reported_test_accuracy": 0.578125,
          "leaky_vs_honest_test_gap": 0.09765625,
          "optimism_gap": 0.064453125
        },
        "honest": {
          "audit_correct": 491,
          "audit_total": 1024,
          "feature_index": 33,
          "polarity": 1,
          "selection_correct": 153,
          "selection_total": 256,
          "test_correct": 123,
          "test_total": 256
        },
        "k": 128,
        "leaky": {
          "audit_correct": 526,
          "audit_total": 1024,
          "feature_index": 108,
          "polarity": -1,
          "test_correct": 148,
          "test_total": 256
        },
        "seed": 29
      },
      {
        "candidate_pair_count": 512,
        "derived": {
          "honest_audit_accuracy": 0.4794921875,
          "honest_selection_accuracy": 0.59765625,
          "honest_test_accuracy": 0.48046875,
          "leaky_audit_accuracy": 0.4814453125,
          "leaky_reported_test_accuracy": 0.58984375,
          "leaky_vs_honest_test_gap": 0.109375,
          "optimism_gap": 0.1083984375
        },
        "honest": {
          "audit_correct": 491,
          "audit_total": 1024,
          "feature_index": 33,
          "polarity": 1,
          "selection_correct": 153,
          "selection_total": 256,
          "test_correct": 123,
          "test_total": 256
        },
        "k": 256,
        "leaky": {
          "audit_correct": 493,
          "audit_total": 1024,
          "feature_index": 136,
          "polarity": 1,
          "test_correct": 151,
          "test_total": 256
        },
        "seed": 29
      },
      {
        "candidate_pair_count": 1024,
        "derived": {
          "honest_audit_accuracy": 0.5166015625,
          "honest_selection_accuracy": 0.6015625,
          "honest_test_accuracy": 0.53515625,
          "leaky_audit_accuracy": 0.51953125,
          "leaky_reported_test_accuracy": 0.59375,
          "leaky_vs_honest_test_gap": 0.05859375,
          "optimism_gap": 0.07421875
        },
        "honest": {
          "audit_correct": 529,
          "audit_total": 1024,
          "feature_index": 451,
          "polarity": 1,
          "selection_correct": 154,
          "selection_total": 256,
          "test_correct": 137,
          "test_total": 256
        },
        "k": 512,
        "leaky": {
          "audit_correct": 532,
          "audit_total": 1024,
          "feature_index": 326,
          "polarity": 1,
          "test_correct": 152,
          "test_total": 256
        },
        "seed": 29
      },
      {
        "candidate_pair_count": 2048,
        "derived": {
          "honest_audit_accuracy": 0.4912109375,
          "honest_selection_accuracy": 0.609375,
          "honest_test_accuracy": 0.515625,
          "leaky_audit_accuracy": 0.51953125,
          "leaky_reported_test_accuracy": 0.59375,
          "leaky_vs_honest_test_gap": 0.078125,
          "optimism_gap": 0.07421875
        },
        "honest": {
          "audit_correct": 503,
          "audit_total": 1024,
          "feature_index": 576,
          "polarity": -1,
          "selection_correct": 156,
          "selection_total": 256,
          "test_correct": 132,
          "test_total": 256
        },
        "k": 1024,
        "leaky": {
          "audit_correct": 532,
          "audit_total": 1024,
          "feature_index": 326,
          "polarity": 1,
          "test_correct": 152,
          "test_total": 256
        },
        "seed": 29
      },
      {
        "candidate_pair_count": 2,
        "derived": {
          "honest_audit_accuracy": 0.5,
          "honest_selection_accuracy": 0.50390625,
          "honest_test_accuracy": 0.51953125,
          "leaky_audit_accuracy": 0.5,
          "leaky_reported_test_accuracy": 0.51953125,
          "leaky_vs_honest_test_gap": 0.0,
          "optimism_gap": 0.01953125
        },
        "honest": {
          "audit_correct": 512,
          "audit_total": 1024,
          "feature_index": 0,
          "polarity": 1,
          "selection_correct": 129,
          "selection_total": 256,
          "test_correct": 133,
          "test_total": 256
        },
        "k": 1,
        "leaky": {
          "audit_correct": 512,
          "audit_total": 1024,
          "feature_index": 0,
          "polarity": 1,
          "test_correct": 133,
          "test_total": 256
        },
        "seed": 30
      },
      {
        "candidate_pair_count": 4,
        "derived": {
          "honest_audit_accuracy": 0.5087890625,
          "honest_selection_accuracy": 0.5078125,
          "honest_test_accuracy": 0.47265625,
          "leaky_audit_accuracy": 0.4912109375,
          "leaky_reported_test_accuracy": 0.52734375,
          "leaky_vs_honest_test_gap": 0.0546875,
          "optimism_gap": 0.0361328125
        },
        "honest": {
          "audit_correct": 521,
          "audit_total": 1024,
          "feature_index": 1,
          "polarity": -1,
          "selection_correct": 130,
          "selection_total": 256,
          "test_correct": 121,
          "test_total": 256
        },
        "k": 2,
        "leaky": {
          "audit_correct": 503,
          "audit_total": 1024,
          "feature_index": 1,
          "polarity": 1,
          "test_correct": 135,
          "test_total": 256
        },
        "seed": 30
      },
      {
        "candidate_pair_count": 8,
        "derived": {
          "honest_audit_accuracy": 0.4990234375,
          "honest_selection_accuracy": 0.546875,
          "honest_test_accuracy": 0.48046875,
          "leaky_audit_accuracy": 0.4912109375,
          "leaky_reported_test_accuracy": 0.52734375,
          "leaky_vs_honest_test_gap": 0.046875,
          "optimism_gap": 0.0361328125
        },
        "honest": {
          "audit_correct": 511,
          "audit_total": 1024,
          "feature_index": 2,
          "polarity": -1,
          "selection_correct": 140,
          "selection_total": 256,
          "test_correct": 123,
          "test_total": 256
        },
        "k": 4,
        "leaky": {
          "audit_correct": 503,
          "audit_total": 1024,
          "feature_index": 1,
          "polarity": 1,
          "test_correct": 135,
          "test_total": 256
        },
        "seed": 30
      },
      {
        "candidate_pair_count": 16,
        "derived": {
          "honest_audit_accuracy": 0.505859375,
          "honest_selection_accuracy": 0.55859375,
          "honest_test_accuracy": 0.47265625,
          "leaky_audit_accuracy": 0.5009765625,
          "leaky_reported_test_accuracy": 0.54296875,
          "leaky_vs_honest_test_gap": 0.0703125,
          "optimism_gap": 0.0419921875
        },
        "honest": {
          "audit_correct": 518,
          "audit_total": 1024,
          "feature_index": 7,
          "polarity": 1,
          "selection_correct": 143,
          "selection_total": 256,
          "test_correct": 121,
          "test_total": 256
        },
        "k": 8,
        "leaky": {
          "audit_correct": 513,
          "audit_total": 1024,
          "feature_index": 6,
          "polarity": -1,
          "test_correct": 139,
          "test_total": 256
        },
        "seed": 30
      },
      {
        "candidate_pair_count": 32,
        "derived": {
          "honest_audit_accuracy": 0.4775390625,
          "honest_selection_accuracy": 0.58203125,
          "honest_test_accuracy": 0.5078125,
          "leaky_audit_accuracy": 0.4921875,
          "leaky_reported_test_accuracy": 0.546875,
          "leaky_vs_honest_test_gap": 0.0390625,
          "optimism_gap": 0.0546875
        },
        "honest": {
          "audit_correct": 489,
          "audit_total": 1024,
          "feature_index": 13,
          "polarity": -1,
          "selection_correct": 149,
          "selection_total": 256,
          "test_correct": 130,
          "test_total": 256
        },
        "k": 16,
        "leaky": {
          "audit_correct": 504,
          "audit_total": 1024,
          "feature_index": 9,
          "polarity": -1,
          "test_correct": 140,
          "test_total": 256
        },
        "seed": 30
      },
      {
        "candidate_pair_count": 64,
        "derived": {
          "honest_audit_accuracy": 0.4775390625,
          "honest_selection_accuracy": 0.58203125,
          "honest_test_accuracy": 0.5078125,
          "leaky_audit_accuracy": 0.4970703125,
          "leaky_reported_test_accuracy": 0.578125,
          "leaky_vs_honest_test_gap": 0.0703125,
          "optimism_gap": 0.0810546875
        },
        "honest": {
          "audit_correct": 489,
          "audit_total": 1024,
          "feature_index": 13,
          "polarity": -1,
          "selection_correct": 149,
          "selection_total": 256,
          "test_correct": 130,
          "test_total": 256
        },
        "k": 32,
        "leaky": {
          "audit_correct": 509,
          "audit_total": 1024,
          "feature_index": 24,
          "polarity": 1,
          "test_correct": 148,
          "test_total": 256
        },
        "seed": 30
      },
      {
        "candidate_pair_count": 128,
        "derived": {
          "honest_audit_accuracy": 0.4775390625,
          "honest_selection_accuracy": 0.58203125,
          "honest_test_accuracy": 0.5078125,
          "leaky_audit_accuracy": 0.4970703125,
          "leaky_reported_test_accuracy": 0.578125,
          "leaky_vs_honest_test_gap": 0.0703125,
          "optimism_gap": 0.0810546875
        },
        "honest": {
          "audit_correct": 489,
          "audit_total": 1024,
          "feature_index": 13,
          "polarity": -1,
          "selection_correct": 149,
          "selection_total": 256,
          "test_correct": 130,
          "test_total": 256
        },
        "k": 64,
        "leaky": {
          "audit_correct": 509,
          "audit_total": 1024,
          "feature_index": 24,
          "polarity": 1,
          "test_correct": 148,
          "test_total": 256
        },
        "seed": 30
      },
      {
        "candidate_pair_count": 256,
        "derived": {
          "honest_audit_accuracy": 0.4775390625,
          "honest_selection_accuracy": 0.58203125,
          "honest_test_accuracy": 0.5078125,
          "leaky_audit_accuracy": 0.482421875,
          "leaky_reported_test_accuracy": 0.59765625,
          "leaky_vs_honest_test_gap": 0.08984375,
          "optimism_gap": 0.115234375
        },
        "honest": {
          "audit_correct": 489,
          "audit_total": 1024,
          "feature_index": 13,
          "polarity": -1,
          "selection_correct": 149,
          "selection_total": 256,
          "test_correct": 130,
          "test_total": 256
        },
        "k": 128,
        "leaky": {
          "audit_correct": 494,
          "audit_total": 1024,
          "feature_index": 121,
          "polarity": -1,
          "test_correct": 153,
          "test_total": 256
        },
        "seed": 30
      },
      {
        "candidate_pair_count": 512,
        "derived": {
          "honest_audit_accuracy": 0.4775390625,
          "honest_selection_accuracy": 0.58203125,
          "honest_test_accuracy": 0.5078125,
          "leaky_audit_accuracy": 0.482421875,
          "leaky_reported_test_accuracy": 0.59765625,
          "leaky_vs_honest_test_gap": 0.08984375,
          "optimism_gap": 0.115234375
        },
        "honest": {
          "audit_correct": 489,
          "audit_total": 1024,
          "feature_index": 13,
          "polarity": -1,
          "selection_correct": 149,
          "selection_total": 256,
          "test_correct": 130,
          "test_total": 256
        },
        "k": 256,
        "leaky": {
          "audit_correct": 494,
          "audit_total": 1024,
          "feature_index": 121,
          "polarity": -1,
          "test_correct": 153,
          "test_total": 256
        },
        "seed": 30
      },
      {
        "candidate_pair_count": 1024,
        "derived": {
          "honest_audit_accuracy": 0.5048828125,
          "honest_selection_accuracy": 0.59765625,
          "honest_test_accuracy": 0.4453125,
          "leaky_audit_accuracy": 0.482421875,
          "leaky_reported_test_accuracy": 0.59765625,
          "leaky_vs_honest_test_gap": 0.15234375,
          "optimism_gap": 0.115234375
        },
        "honest": {
          "audit_correct": 517,
          "audit_total": 1024,
          "feature_index": 406,
          "polarity": -1,
          "selection_correct": 153,
          "selection_total": 256,
          "test_correct": 114,
          "test_total": 256
        },
        "k": 512,
        "leaky": {
          "audit_correct": 494,
          "audit_total": 1024,
          "feature_index": 121,
          "polarity": -1,
          "test_correct": 153,
          "test_total": 256
        },
        "seed": 30
      },
      {
        "candidate_pair_count": 2048,
        "derived": {
          "honest_audit_accuracy": 0.4892578125,
          "honest_selection_accuracy": 0.60546875,
          "honest_test_accuracy": 0.4921875,
          "leaky_audit_accuracy": 0.482421875,
          "leaky_reported_test_accuracy": 0.59765625,
          "leaky_vs_honest_test_gap": 0.10546875,
          "optimism_gap": 0.115234375
        },
        "honest": {
          "audit_correct": 501,
          "audit_total": 1024,
          "feature_index": 701,
          "polarity": -1,
          "selection_correct": 155,
          "selection_total": 256,
          "test_correct": 126,
          "test_total": 256
        },
        "k": 1024,
        "leaky": {
          "audit_correct": 494,
          "audit_total": 1024,
          "feature_index": 121,
          "polarity": -1,
          "test_correct": 153,
          "test_total": 256
        },
        "seed": 30
      },
      {
        "candidate_pair_count": 2,
        "derived": {
          "honest_audit_accuracy": 0.466796875,
          "honest_selection_accuracy": 0.50390625,
          "honest_test_accuracy": 0.5,
          "leaky_audit_accuracy": 0.466796875,
          "leaky_reported_test_accuracy": 0.5,
          "leaky_vs_honest_test_gap": 0.0,
          "optimism_gap": 0.033203125
        },
        "honest": {
          "audit_correct": 478,
          "audit_total": 1024,
          "feature_index": 0,
          "polarity": 1,
          "selection_correct": 129,
          "selection_total": 256,
          "test_correct": 128,
          "test_total": 256
        },
        "k": 1,
        "leaky": {
          "audit_correct": 478,
          "audit_total": 1024,
          "feature_index": 0,
          "polarity": 1,
          "test_correct": 128,
          "test_total": 256
        },
        "seed": 31
      },
      {
        "candidate_pair_count": 4,
        "derived": {
          "honest_audit_accuracy": 0.505859375,
          "honest_selection_accuracy": 0.515625,
          "honest_test_accuracy": 0.5390625,
          "leaky_audit_accuracy": 0.505859375,
          "leaky_reported_test_accuracy": 0.5390625,
          "leaky_vs_honest_test_gap": 0.0,
          "optimism_gap": 0.033203125
        },
        "honest": {
          "audit_correct": 518,
          "audit_total": 1024,
          "feature_index": 1,
          "polarity": -1,
          "selection_correct": 132,
          "selection_total": 256,
          "test_correct": 138,
          "test_total": 256
        },
        "k": 2,
        "leaky": {
          "audit_correct": 518,
          "audit_total": 1024,
          "feature_index": 1,
          "polarity": -1,
          "test_correct": 138,
          "test_total": 256
        },
        "seed": 31
      },
      {
        "candidate_pair_count": 8,
        "derived": {
          "honest_audit_accuracy": 0.5068359375,
          "honest_selection_accuracy": 0.5546875,
          "honest_test_accuracy": 0.54296875,
          "leaky_audit_accuracy": 0.5068359375,
          "leaky_reported_test_accuracy": 0.54296875,
          "leaky_vs_honest_test_gap": 0.0,
          "optimism_gap": 0.0361328125
        },
        "honest": {
          "audit_correct": 519,
          "audit_total": 1024,
          "feature_index": 3,
          "polarity": -1,
          "selection_correct": 142,
          "selection_total": 256,
          "test_correct": 139,
          "test_total": 256
        },
        "k": 4,
        "leaky": {
          "audit_correct": 519,
          "audit_total": 1024,
          "feature_index": 3,
          "polarity": -1,
          "test_correct": 139,
          "test_total": 256
        },
        "seed": 31
      },
      {
        "candidate_pair_count": 16,
        "derived": {
          "honest_audit_accuracy": 0.5068359375,
          "honest_selection_accuracy": 0.5546875,
          "honest_test_accuracy": 0.54296875,
          "leaky_audit_accuracy": 0.5185546875,
          "leaky_reported_test_accuracy": 0.55078125,
          "leaky_vs_honest_test_gap": 0.0078125,
          "optimism_gap": 0.0322265625
        },
        "honest": {
          "audit_correct": 519,
          "audit_total": 1024,
          "feature_index": 3,
          "polarity": -1,
          "selection_correct": 142,
          "selection_total": 256,
          "test_correct": 139,
          "test_total": 256
        },
        "k": 8,
        "leaky": {
          "audit_correct": 531,
          "audit_total": 1024,
          "feature_index": 6,
          "polarity": -1,
          "test_correct": 141,
          "test_total": 256
        },
        "seed": 31
      },
      {
        "candidate_pair_count": 32,
        "derived": {
          "honest_audit_accuracy": 0.5068359375,
          "honest_selection_accuracy": 0.5546875,
          "honest_test_accuracy": 0.54296875,
          "leaky_audit_accuracy": 0.5185546875,
          "leaky_reported_test_accuracy": 0.55078125,
          "leaky_vs_honest_test_gap": 0.0078125,
          "optimism_gap": 0.0322265625
        },
        "honest": {
          "audit_correct": 519,
          "audit_total": 1024,
          "feature_index": 3,
          "polarity": -1,
          "selection_correct": 142,
          "selection_total": 256,
          "test_correct": 139,
          "test_total": 256
        },
        "k": 16,
        "leaky": {
          "audit_correct": 531,
          "audit_total": 1024,
          "feature_index": 6,
          "polarity": -1,
          "test_correct": 141,
          "test_total": 256
        },
        "seed": 31
      },
      {
        "candidate_pair_count": 64,
        "derived": {
          "honest_audit_accuracy": 0.5068359375,
          "honest_selection_accuracy": 0.5546875,
          "honest_test_accuracy": 0.54296875,
          "leaky_audit_accuracy": 0.5185546875,
          "leaky_reported_test_accuracy": 0.55078125,
          "leaky_vs_honest_test_gap": 0.0078125,
          "optimism_gap": 0.0322265625
        },
        "honest": {
          "audit_correct": 519,
          "audit_total": 1024,
          "feature_index": 3,
          "polarity": -1,
          "selection_correct": 142,
          "selection_total": 256,
          "test_correct": 139,
          "test_total": 256
        },
        "k": 32,
        "leaky": {
          "audit_correct": 531,
          "audit_total": 1024,
          "feature_index": 6,
          "polarity": -1,
          "test_correct": 141,
          "test_total": 256
        },
        "seed": 31
      },
      {
        "candidate_pair_count": 128,
        "derived": {
          "honest_audit_accuracy": 0.5029296875,
          "honest_selection_accuracy": 0.55859375,
          "honest_test_accuracy": 0.44921875,
          "leaky_audit_accuracy": 0.533203125,
          "leaky_reported_test_accuracy": 0.59375,
          "leaky_vs_honest_test_gap": 0.14453125,
          "optimism_gap": 0.060546875
        },
        "honest": {
          "audit_correct": 515,
          "audit_total": 1024,
          "feature_index": 35,
          "polarity": -1,
          "selection_correct": 143,
          "selection_total": 256,
          "test_correct": 115,
          "test_total": 256
        },
        "k": 64,
        "leaky": {
          "audit_correct": 546,
          "audit_total": 1024,
          "feature_index": 46,
          "polarity": 1,
          "test_correct": 152,
          "test_total": 256
        },
        "seed": 31
      },
      {
        "candidate_pair_count": 256,
        "derived": {
          "honest_audit_accuracy": 0.50390625,
          "honest_selection_accuracy": 0.57421875,
          "honest_test_accuracy": 0.4375,
          "leaky_audit_accuracy": 0.533203125,
          "leaky_reported_test_accuracy": 0.59375,
          "leaky_vs_honest_test_gap": 0.15625,
          "optimism_gap": 0.060546875
        },
        "honest": {
          "audit_correct": 516,
          "audit_total": 1024,
          "feature_index": 126,
          "polarity": 1,
          "selection_correct": 147,
          "selection_total": 256,
          "test_correct": 112,
          "test_total": 256
        },
        "k": 128,
        "leaky": {
          "audit_correct": 546,
          "audit_total": 1024,
          "feature_index": 46,
          "polarity": 1,
          "test_correct": 152,
          "test_total": 256
        },
        "seed": 31
      },
      {
        "candidate_pair_count": 512,
        "derived": {
          "honest_audit_accuracy": 0.50390625,
          "honest_selection_accuracy": 0.57421875,
          "honest_test_accuracy": 0.4375,
          "leaky_audit_accuracy": 0.533203125,
          "leaky_reported_test_accuracy": 0.59375,
          "leaky_vs_honest_test_gap": 0.15625,
          "optimism_gap": 0.060546875
        },
        "honest": {
          "audit_correct": 516,
          "audit_total": 1024,
          "feature_index": 126,
          "polarity": 1,
          "selection_correct": 147,
          "selection_total": 256,
          "test_correct": 112,
          "test_total": 256
        },
        "k": 256,
        "leaky": {
          "audit_correct": 546,
          "audit_total": 1024,
          "feature_index": 46,
          "polarity": 1,
          "test_correct": 152,
          "test_total": 256
        },
        "seed": 31
      },
      {
        "candidate_pair_count": 1024,
        "derived": {
          "honest_audit_accuracy": 0.4921875,
          "honest_selection_accuracy": 0.58984375,
          "honest_test_accuracy": 0.48046875,
          "leaky_audit_accuracy": 0.486328125,
          "leaky_reported_test_accuracy": 0.59765625,
          "leaky_vs_honest_test_gap": 0.1171875,
          "optimism_gap": 0.111328125
        },
        "honest": {
          "audit_correct": 504,
          "audit_total": 1024,
          "feature_index": 263,
          "polarity": -1,
          "selection_correct": 151,
          "selection_total": 256,
          "test_correct": 123,
          "test_total": 256
        },
        "k": 512,
        "leaky": {
          "audit_correct": 498,
          "audit_total": 1024,
          "feature_index": 294,
          "polarity": -1,
          "test_correct": 153,
          "test_total": 256
        },
        "seed": 31
      },
      {
        "candidate_pair_count": 2048,
        "derived": {
          "honest_audit_accuracy": 0.4775390625,
          "honest_selection_accuracy": 0.59765625,
          "honest_test_accuracy": 0.484375,
          "leaky_audit_accuracy": 0.5126953125,
          "leaky_reported_test_accuracy": 0.62109375,
          "leaky_vs_honest_test_gap": 0.13671875,
          "optimism_gap": 0.1083984375
        },
        "honest": {
          "audit_correct": 489,
          "audit_total": 1024,
          "feature_index": 698,
          "polarity": 1,
          "selection_correct": 153,
          "selection_total": 256,
          "test_correct": 124,
          "test_total": 256
        },
        "k": 1024,
        "leaky": {
          "audit_correct": 525,
          "audit_total": 1024,
          "feature_index": 588,
          "polarity": 1,
          "test_correct": 159,
          "test_total": 256
        },
        "seed": 31
      },
      {
        "candidate_pair_count": 2,
        "derived": {
          "honest_audit_accuracy": 0.505859375,
          "honest_selection_accuracy": 0.50390625,
          "honest_test_accuracy": 0.5234375,
          "leaky_audit_accuracy": 0.505859375,
          "leaky_reported_test_accuracy": 0.5234375,
          "leaky_vs_honest_test_gap": 0.0,
          "optimism_gap": 0.017578125
        },
        "honest": {
          "audit_correct": 518,
          "audit_total": 1024,
          "feature_index": 0,
          "polarity": 1,
          "selection_correct": 129,
          "selection_total": 256,
          "test_correct": 134,
          "test_total": 256
        },
        "k": 1,
        "leaky": {
          "audit_correct": 518,
          "audit_total": 1024,
          "feature_index": 0,
          "polarity": 1,
          "test_correct": 134,
          "test_total": 256
        },
        "seed": 32
      },
      {
        "candidate_pair_count": 4,
        "derived": {
          "honest_audit_accuracy": 0.505859375,
          "honest_selection_accuracy": 0.50390625,
          "honest_test_accuracy": 0.5234375,
          "leaky_audit_accuracy": 0.505859375,
          "leaky_reported_test_accuracy": 0.5234375,
          "leaky_vs_honest_test_gap": 0.0,
          "optimism_gap": 0.017578125
        },
        "honest": {
          "audit_correct": 518,
          "audit_total": 1024,
          "feature_index": 0,
          "polarity": 1,
          "selection_correct": 129,
          "selection_total": 256,
          "test_correct": 134,
          "test_total": 256
        },
        "k": 2,
        "leaky": {
          "audit_correct": 518,
          "audit_total": 1024,
          "feature_index": 0,
          "polarity": 1,
          "test_correct": 134,
          "test_total": 256
        },
        "seed": 32
      },
      {
        "candidate_pair_count": 8,
        "derived": {
          "honest_audit_accuracy": 0.51953125,
          "honest_selection_accuracy": 0.54296875,
          "honest_test_accuracy": 0.453125,
          "leaky_audit_accuracy": 0.48046875,
          "leaky_reported_test_accuracy": 0.546875,
          "leaky_vs_honest_test_gap": 0.09375,
          "optimism_gap": 0.06640625
        },
        "honest": {
          "audit_correct": 532,
          "audit_total": 1024,
          "feature_index": 2,
          "polarity": -1,
          "selection_correct": 139,
          "selection_total": 256,
          "test_correct": 116,
          "test_total": 256
        },
        "k": 4,
        "leaky": {
          "audit_correct": 492,
          "audit_total": 1024,
          "feature_index": 2,
          "polarity": 1,
          "test_correct": 140,
          "test_total": 256
        },
        "seed": 32
      },
      {
        "candidate_pair_count": 16,
        "derived": {
          "honest_audit_accuracy": 0.51953125,
          "honest_selection_accuracy": 0.54296875,
          "honest_test_accuracy": 0.453125,
          "leaky_audit_accuracy": 0.486328125,
          "leaky_reported_test_accuracy": 0.55078125,
          "leaky_vs_honest_test_gap": 0.09765625,
          "optimism_gap": 0.064453125
        },
        "honest": {
          "audit_correct": 532,
          "audit_total": 1024,
          "feature_index": 2,
          "polarity": -1,
          "selection_correct": 139,
          "selection_total": 256,
          "test_correct": 116,
          "test_total": 256
        },
        "k": 8,
        "leaky": {
          "audit_correct": 498,
          "audit_total": 1024,
          "feature_index": 6,
          "polarity": 1,
          "test_correct": 141,
          "test_total": 256
        },
        "seed": 32
      },
      {
        "candidate_pair_count": 32,
        "derived": {
          "honest_audit_accuracy": 0.470703125,
          "honest_selection_accuracy": 0.5546875,
          "honest_test_accuracy": 0.484375,
          "leaky_audit_accuracy": 0.5009765625,
          "leaky_reported_test_accuracy": 0.58203125,
          "leaky_vs_honest_test_gap": 0.09765625,
          "optimism_gap": 0.0810546875
        },
        "honest": {
          "audit_correct": 482,
          "audit_total": 1024,
          "feature_index": 14,
          "polarity": -1,
          "selection_correct": 142,
          "selection_total": 256,
          "test_correct": 124,
          "test_total": 256
        },
        "k": 16,
        "leaky": {
          "audit_correct": 513,
          "audit_total": 1024,
          "feature_index": 13,
          "polarity": 1,
          "test_correct": 149,
          "test_total": 256
        },
        "seed": 32
      },
      {
        "candidate_pair_count": 64,
        "derived": {
          "honest_audit_accuracy": 0.470703125,
          "honest_selection_accuracy": 0.5546875,
          "honest_test_accuracy": 0.484375,
          "leaky_audit_accuracy": 0.533203125,
          "leaky_reported_test_accuracy": 0.5859375,
          "leaky_vs_honest_test_gap": 0.1015625,
          "optimism_gap": 0.052734375
        },
        "honest": {
          "audit_correct": 482,
          "audit_total": 1024,
          "feature_index": 14,
          "polarity": -1,
          "selection_correct": 142,
          "selection_total": 256,
          "test_correct": 124,
          "test_total": 256
        },
        "k": 32,
        "leaky": {
          "audit_correct": 546,
          "audit_total": 1024,
          "feature_index": 29,
          "polarity": 1,
          "test_correct": 150,
          "test_total": 256
        },
        "seed": 32
      },
      {
        "candidate_pair_count": 128,
        "derived": {
          "honest_audit_accuracy": 0.4853515625,
          "honest_selection_accuracy": 0.5703125,
          "honest_test_accuracy": 0.48046875,
          "leaky_audit_accuracy": 0.533203125,
          "leaky_reported_test_accuracy": 0.5859375,
          "leaky_vs_honest_test_gap": 0.10546875,
          "optimism_gap": 0.052734375
        },
        "honest": {
          "audit_correct": 497,
          "audit_total": 1024,
          "feature_index": 33,
          "polarity": -1,
          "selection_correct": 146,
          "selection_total": 256,
          "test_correct": 123,
          "test_total": 256
        },
        "k": 64,
        "leaky": {
          "audit_correct": 546,
          "audit_total": 1024,
          "feature_index": 29,
          "polarity": 1,
          "test_correct": 150,
          "test_total": 256
        },
        "seed": 32
      },
      {
        "candidate_pair_count": 256,
        "derived": {
          "honest_audit_accuracy": 0.482421875,
          "honest_selection_accuracy": 0.5859375,
          "honest_test_accuracy": 0.4765625,
          "leaky_audit_accuracy": 0.533203125,
          "leaky_reported_test_accuracy": 0.5859375,
          "leaky_vs_honest_test_gap": 0.109375,
          "optimism_gap": 0.052734375
        },
        "honest": {
          "audit_correct": 494,
          "audit_total": 1024,
          "feature_index": 87,
          "polarity": 1,
          "selection_correct": 150,
          "selection_total": 256,
          "test_correct": 122,
          "test_total": 256
        },
        "k": 128,
        "leaky": {
          "audit_correct": 546,
          "audit_total": 1024,
          "feature_index": 29,
          "polarity": 1,
          "test_correct": 150,
          "test_total": 256
        },
        "seed": 32
      },
      {
        "candidate_pair_count": 512,
        "derived": {
          "honest_audit_accuracy": 0.482421875,
          "honest_selection_accuracy": 0.5859375,
          "honest_test_accuracy": 0.4765625,
          "leaky_audit_accuracy": 0.533203125,
          "leaky_reported_test_accuracy": 0.5859375,
          "leaky_vs_honest_test_gap": 0.109375,
          "optimism_gap": 0.052734375
        },
        "honest": {
          "audit_correct": 494,
          "audit_total": 1024,
          "feature_index": 87,
          "polarity": 1,
          "selection_correct": 150,
          "selection_total": 256,
          "test_correct": 122,
          "test_total": 256
        },
        "k": 256,
        "leaky": {
          "audit_correct": 546,
          "audit_total": 1024,
          "feature_index": 29,
          "polarity": 1,
          "test_correct": 150,
          "test_total": 256
        },
        "seed": 32
      },
      {
        "candidate_pair_count": 1024,
        "derived": {
          "honest_audit_accuracy": 0.5009765625,
          "honest_selection_accuracy": 0.59765625,
          "honest_test_accuracy": 0.56640625,
          "leaky_audit_accuracy": 0.505859375,
          "leaky_reported_test_accuracy": 0.60546875,
          "leaky_vs_honest_test_gap": 0.0390625,
          "optimism_gap": 0.099609375
        },
        "honest": {
          "audit_correct": 513,
          "audit_total": 1024,
          "feature_index": 414,
          "polarity": -1,
          "selection_correct": 153,
          "selection_total": 256,
          "test_correct": 145,
          "test_total": 256
        },
        "k": 512,
        "leaky": {
          "audit_correct": 518,
          "audit_total": 1024,
          "feature_index": 509,
          "polarity": -1,
          "test_correct": 155,
          "test_total": 256
        },
        "seed": 32
      },
      {
        "candidate_pair_count": 2048,
        "derived": {
          "honest_audit_accuracy": 0.5009765625,
          "honest_selection_accuracy": 0.59765625,
          "honest_test_accuracy": 0.56640625,
          "leaky_audit_accuracy": 0.4931640625,
          "leaky_reported_test_accuracy": 0.6171875,
          "leaky_vs_honest_test_gap": 0.05078125,
          "optimism_gap": 0.1240234375
        },
        "honest": {
          "audit_correct": 513,
          "audit_total": 1024,
          "feature_index": 414,
          "polarity": -1,
          "selection_correct": 153,
          "selection_total": 256,
          "test_correct": 145,
          "test_total": 256
        },
        "k": 1024,
        "leaky": {
          "audit_correct": 505,
          "audit_total": 1024,
          "feature_index": 600,
          "polarity": 1,
          "test_correct": 158,
          "test_total": 256
        },
        "seed": 32
      }
    ],
    "seed_count": 32
  },
  "protocol": {
    "audit_examples": 1024,
    "audit_policy": "All feature, polarity, and K choices are frozen before audit evaluation. One final reporting stage then records audit counts for every predeclared candidate and audit accuracies for the frozen choices. The audit is disclosed, retired, and never reused for tuning.",
    "noise_features": 1024,
    "selection_examples": 256,
    "test_examples": 256,
    "tie_rule": "accuracy descending, feature index ascending, positive polarity before inverse polarity"
  },
  "provenance": {
    "artifact_review": {
      "assurance": "self-asserted-local",
      "decision": "approve",
      "label": "Artifact checked",
      "receipt": "research/test-guided-feature-selection-null/artifact-review.agent.json",
      "receipt_sha256": "c0a49e9b87dc3d8927c9d124da3d2c2f2b7951eacf37780ed8308b6ca4559b0f",
      "reviewer": {
        "id": "codex-test-reuse-artifact-review-agent",
        "type": "agent"
      },
      "scope": "artifact-check",
      "warning": "This receipt does not cryptographically authenticate reviewer identity."
    },
    "bundle_digest": "76a9879c53d5a7e5936f132195cd9bffc39733b89cf57f0829223604569e3d3d",
    "bundle_path": "research/test-guided-feature-selection-null/bundle",
    "commit_dirty": false,
    "covered_evidence_paths": 393,
    "evidence_labels": [
      "Executed",
      "Artifact checked"
    ],
    "label_warning": "Two trials by one harness are exact same-environment repeat execution, not independent scientific reproduction.",
    "limitations": [
      "No seed was rerun for this artifact check or learning-lab projection.",
      "Not independent reproduction, replication, peer review, external attestation, authenticated identity, safety certification, or human approval.",
      "Network and memory isolation were not enforced at the operating-system boundary.",
      "The bundle digest is not externally signed or authenticated."
    ],
    "protocol_sha256": "d820eda7098e070fc672e9d287785e13c5874dbe015ef139d46c73bcaa7decbb",
    "repeat_status": "exact-same-environment-repeat",
    "run_artifacts_checked": 384,
    "scientific_label": "Executed",
    "source_commit": "0d89488110f4d7356a448d7d89cd8521296fb92d",
    "source_files_checked": 4,
    "source_sha256": "b3922c5612b859c4a123db48a5536ea3a0593fe3a35b3b260ef6630d3a10d9b6",
    "summary_sha256": "dc68e31bac7a5d7aeaab96b352a54f3d0a84f2e47c3d0da3952312667912ce81",
    "trial_a_manifest_sha256": "ea7f4e953e0ade10f3233fd9b2abedda6383596fe28d6250aa2a88374fd6d1a7",
    "trial_b_manifest_sha256": "1ba9b588c499f9ad661d6ee2e952b490855705d34fa286fae879acff5e8b222a",
    "trial_fingerprint": "b99d21823c3318603e27497ad7fe0fa04cb9d1b71d2002d270fedb6e7101f20b"
  },
  "reset": {
    "audit_revealed": false,
    "candidate_feature_index": 0,
    "export_enabled": false,
    "prediction": null,
    "seed": 1
  },
  "schema": "cybernative.test-set-reuse-lab-data/v1",
  "solution": {
    "correct_prediction_id": "audit-can-differ",
    "explanation": "The independent audit was not used to choose the feature or polarity, so sampling variation can make its accuracy differ from the reused test result. It is not required to match the test result or equal exactly 50 percent in a finite sample.",
    "options": [
      {
        "id": "audit-must-match",
        "label": "The audit must match the reported test accuracy."
      },
      {
        "id": "audit-can-differ",
        "label": "The independent audit can differ; its exact value cannot be inferred in advance."
      },
      {
        "id": "audit-exactly-half",
        "label": "The audit must equal exactly 50 percent."
      }
    ],
    "scope_warning": "This retained result is not a universal estimate, independent reproduction or replication, peer review, authenticated identity, external attestation, safety certification, or human approval."
  },
  "summary_all_criteria_passed": true
}