{
  "verdict": "trade-off",
  "S1": {
    "meanTvdDiff": {
      "estimate": -0.0008,
      "lo": -0.0013,
      "hi": -0.0004,
      "n": 100,
      "excludesZero": true
    },
    "topChoiceChanges": {
      "self": 0,
      "v1": 0.01
    },
    "checks": {
      "tvd_lower_with_interval_excluding_0": true,
      "top_choice_change_rate_not_higher": true
    },
    "pass": true
  },
  "S2": {
    "checks": {
      "top1_not_lower_by_more_than_1pt": true,
      "tvd_not_higher_by_more_than_0.01": true,
      "sheet_mae_not_higher_by_more_than_0.01": true,
      "juror_mae_not_higher_by_more_than_0.01": false
    },
    "pass": false,
    "rule": "point estimates of self-1 minus v1 against the pre-registered margins; intervals reported"
  },
  "S3": {
    "complianceDiff": {
      "estimate": 0,
      "lo": 0,
      "hi": 0,
      "n": 400,
      "excludesZero": false
    },
    "pass": true,
    "rule": "point estimate not lower than v1 by more than 0.02"
  },
  "labels": {
    "teacher": "checkpoints/v1 (SHA-256 27f210cb321eda7e9dd93fd3c493d8c3b467e510353dbc4d013853801006429c), served temperatures applied",
    "seed": 20260923,
    "K": 4,
    "states": 104516,
    "questions": 134371,
    "questions_labelled": 134371,
    "questions_dropped_option_fell_off": 0,
    "decisions_with_state_keys_reordered": 50090,
    "decisions": 100000,
    "by_origin": {
      "train/character_sheet": {
        "questions": 34120,
        "mean_view_tvd": 0.01560600949380359,
        "top_choice_changes_across_views": 0.04410902696365768,
        "mean_tvd_teacher_vs_logged_view": 0.009994324169365664,
        "mean_tvd_geometric_vs_arithmetic": 0.00011499688713676738
      },
      "train/juror_vote": {
        "questions": 251,
        "mean_view_tvd": 0.04519767298364386,
        "top_choice_changes_across_views": 0.41434262948207173,
        "mean_tvd_teacher_vs_logged_view": 0.04713359268087667,
        "mean_tvd_geometric_vs_arithmetic": 0.0003379074399279831
      },
      "train/villager_decision": {
        "questions": 50000,
        "mean_view_tvd": 0.03891233383848433,
        "top_choice_changes_across_views": 0.08512,
        "mean_tvd_teacher_vs_logged_view": 0.030636755755822215,
        "mean_tvd_geometric_vs_arithmetic": 0.001559835992927409
      },
      "v1-run/villager_decision": {
        "questions": 50000,
        "mean_view_tvd": 0.04444454146863189,
        "top_choice_changes_across_views": 0.10508,
        "mean_tvd_teacher_vs_logged_view": 0.03494843816432027,
        "mean_tvd_geometric_vs_arithmetic": 0.0019032297151949855
      }
    },
    "gpu_minutes_forward": 57.76,
    "wall_minutes": 58.83,
    "sequences_forwarded": 537484,
    "labels_sha256": "ebc7e1ab65bce2cfb9530f88368504c3e1e40c578cbcafa696228995b24f6e67",
    "args": {
      "decisions_v1run": 50000,
      "decisions_train": 50000,
      "limit": 0,
      "chunk": 16000,
      "batch_tokens": 12288,
      "out": "/home/elberacasa/Projects/jevtown/layatown/selfimprove/cache/labels.pt"
    }
  },
  "training": {
    "data": "round 1 self-consistency labels from v1 (K=4 views, mean), no Jev labels; see docs/SELF-IMPROVE.md",
    "labels_sha256": "ebc7e1ab65bce2cfb9530f88368504c3e1e40c578cbcafa696228995b24f6e67",
    "decisions": 100000,
    "sheet_answers": 34120,
    "juror_votes": 251,
    "steps": 6225,
    "minutes": 75.1,
    "temperature_fit_minutes": 0.8,
    "peak_gpu_gib": 4.05,
    "args": {
      "out": "checkpoints/self-1",
      "labels": "/home/elberacasa/Projects/jevtown/layatown/selfimprove/cache/labels.pt",
      "freeze": 20,
      "max_tokens": 6144,
      "lr_encoder": 1e-05,
      "lr_head": 3e-05,
      "juror_repeat": 20,
      "target": "mean",
      "seed": 7,
      "max_steps": 0
    }
  },
  "temperature": [
    1.021849513053894,
    1.0,
    1.0171219110488892
  ]
}