{
  "bench_version": 1,
  "track": "quadratic_context_multitoken",
  "model": "local_mlp",
  "metric_direction": "lower is better",
  "n_seeds": 8,
  "baseline": {
    "best_cfg": {
      "lr": 0.003
    },
    "sweep": [
      {
        "cfg": {
          "lr": 0.001
        },
        "mean": 0.053421652875840664
      },
      {
        "cfg": {
          "lr": 0.003
        },
        "mean": 0.05309625621885061
      },
      {
        "cfg": {
          "lr": 0.009
        },
        "mean": 0.053741672076284885
      }
    ],
    "full": {
      "mean": 0.054212482180446386,
      "std": 0.007988832816369695,
      "per_seed": [
        0.06300174444913864,
        0.05289159715175629,
        0.04336535930633545,
        0.05312632396817207,
        0.06166467070579529,
        0.0440090112388134,
        0.0498502142727375,
        0.06579093635082245
      ],
      "n": 8
    },
    "all_full": [
      {
        "per_seed": [
          0.055687107145786285,
          0.05908738821744919,
          0.047396738082170486,
          0.051515378057956696,
          0.06218918785452843,
          0.057241544127464294,
          0.05223905295133591,
          0.0556969977915287
        ],
        "mean": 0.0551316742785275,
        "std": 0.004656944311225746,
        "lr": 0.001,
        "epochs": 18,
        "ratios": [
          0.6952053904533386,
          0.7031234502792358,
          0.6710729598999023,
          0.6854270696640015,
          0.7208964824676514,
          0.7045612335205078,
          0.6924821734428406,
          0.6988305449485779
        ]
      },
      {
        "per_seed": [
          0.06300174444913864,
          0.05289159715175629,
          0.04336535930633545,
          0.05312632396817207,
          0.06166467070579529,
          0.0440090112388134,
          0.0498502142727375,
          0.06579093635082245
        ],
        "mean": 0.054212482180446386,
        "std": 0.008540421519734782,
        "lr": 0.003,
        "epochs": 18,
        "ratios": [
          0.7323943376541138,
          0.6982712149620056,
          0.6692168116569519,
          0.702392041683197,
          0.7244734764099121,
          0.6718965172767639,
          0.6918891668319702,
          0.7385606169700623
        ]
      },
      {
        "per_seed": [
          0.0538078248500824,
          0.050708189606666565,
          0.05362054705619812,
          0.05683012679219246,
          0.05819164216518402,
          0.050066135823726654,
          0.04446486756205559,
          0.059522490948438644
        ],
        "mean": 0.053401478100568056,
        "std": 0.0049370686751292045,
        "lr": 0.009,
        "epochs": 18,
        "ratios": [
          0.7059165835380554,
          0.6909855008125305,
          0.6965495944023132,
          0.7143308520317078,
          0.7151844501495361,
          0.6841487884521484,
          0.6683564186096191,
          0.7204694151878357
        ]
      }
    ]
  },
  "idea": {
    "per_seed": [
      0.0006221848307177424,
      0.0005909778992645442,
      0.00076779950177297,
      0.0006266951677389443,
      0.000846916635055095,
      0.0007931980071589351,
      0.0008056167862378061,
      0.0007610649918206036
    ],
    "mean": 0.0007268067274708301,
    "std": 9.806469420502945e-05,
    "lr": 0.009,
    "epochs": 18,
    "ratios": [
      0.9858091473579407,
      0.9577128887176514,
      0.9639421105384827,
      0.9625136852264404,
      0.9858413338661194,
      0.969452440738678,
      0.95078045129776,
      0.9798619747161865
    ],
    "sweep": [
      {
        "per_seed": [
          0.005637110210955143,
          0.004224653821438551,
          0.0062506780959665775,
          0.005489944480359554,
          0.0034460758324712515,
          0.004024875815957785,
          0.003128137206658721,
          0.0033030833583325148
        ],
        "mean": 0.004438069852767512,
        "std": 0.0011972666719732973,
        "lr": 0.001,
        "epochs": 18,
        "ratios": [
          0.8957160115242004,
          0.9036473035812378,
          0.9012296199798584,
          0.8910174369812012,
          0.8902207612991333,
          0.8937434554100037,
          0.9013668894767761,
          0.9114634394645691
        ]
      },
      {
        "per_seed": [
          0.000794272986240685,
          0.0007921864162199199,
          0.001272958004847169,
          0.000900555111002177,
          0.0006123684579506516,
          0.0008760863565839827,
          0.0007426603115163743,
          0.0007867970853112638
        ],
        "mean": 0.0008472355912090279,
        "std": 0.00019291068730023744,
        "lr": 0.003,
        "epochs": 18,
        "ratios": [
          0.9768263101577759,
          0.9619719982147217,
          0.9295895099639893,
          0.9479344487190247,
          0.9476742744445801,
          0.9384970664978027,
          0.9460005164146423,
          0.9515169262886047
        ]
      },
      {
        "per_seed": [
          0.0006221848307177424,
          0.0005909778992645442,
          0.00076779950177297,
          0.0006266951677389443,
          0.000846916635055095,
          0.0007931980071589351,
          0.0008056167862378061,
          0.0007610649918206036
        ],
        "mean": 0.0007268067274708301,
        "std": 9.806469420502945e-05,
        "lr": 0.009,
        "epochs": 18,
        "ratios": [
          0.9858091473579407,
          0.9577128887176514,
          0.9639421105384827,
          0.9625136852264404,
          0.9858413338661194,
          0.969452440738678,
          0.95078045129776,
          0.9798619747161865
        ]
      }
    ]
  },
  "comparison": {
    "delta_mean": -0.053485675452975556,
    "idea_wins": 8,
    "n_pairs": 8,
    "per_seed_diffs": [
      -0.0623795596184209,
      -0.05230061925249174,
      -0.04259755980456248,
      -0.05249962880043313,
      -0.06081775407074019,
      -0.043215813231654465,
      -0.0490445974864997,
      -0.06502987135900185
    ],
    "p_value": 0.0081,
    "mde": 0.007152833233489385,
    "mde_rel_pct": 13.194070711761844,
    "verdict": "idea better (significant)",
    "system_worked": true
  },
  "mechanism_signature": {
    "predicted_shrink": 0.4,
    "observed_nn_output_ratio": 0.9694892540574074,
    "relative_error": 1.4237231351435182,
    "confirmed": false
  },
  "custom_track": {
    "name": "quadratic_context_multitoken",
    "file": "quadratic_context_track.py",
    "domain": "diffusion-sampling"
  },
  "seconds": 17.945859909057617
}