1{
  2  "bench_version": 1,
  3  "track": "sequence",
  4  "model": "transformer_tiny",
  5  "metric_direction": "lower is better",
  6  "n_seeds": 8,
  7  "baseline": {
  8    "best_cfg": {
  9      "lr": 0.001
 10    },
 11    "sweep": [
 12      {
 13        "cfg": {
 14          "lr": 0.001
 15        },
 16        "mean": 0.3383217714726925
 17      },
 18      {
 19        "cfg": {
 20          "lr": 0.003
 21        },
 22        "mean": 0.3591853603720665
 23      },
 24      {
 25        "cfg": {
 26          "lr": 0.009
 27        },
 28        "mean": 1.0258603394031525
 29      }
 30    ],
 31    "full": {
 32      "mean": 0.3383217714726925,
 33      "std": 0.046140530839104237,
 34      "per_seed": [
 35        0.3308762013912201,
 36        0.28574880957603455,
 37        0.3010590970516205,
 38        0.3668035864830017,
 39        0.33377403020858765,
 40        0.28919702768325806,
 41        0.4317983090877533,
 42        0.3673171103000641
 43      ],
 44      "n": 8
 45    }
 46  },
 47  "idea": {
 48    "mean": 0.33857686817646027,
 49    "std": 0.04627190481655083,
 50    "per_seed": [
 51      0.3314558267593384,
 52      0.2857537865638733,
 53      0.30147865414619446,
 54      0.36729896068573,
 55      0.3336082696914673,
 56      0.28876644372940063,
 57      0.43178096413612366,
 58      0.36847203969955444
 59    ],
 60    "n": 8
 61  },
 62  "comparison": {
 63    "delta_mean": 0.0002550967037677765,
 64    "idea_wins": 3,
 65    "n_pairs": 8,
 66    "per_seed_diffs": [
 67      0.0005796253681182861,
 68      4.976987838745117e-06,
 69      0.0004195570945739746,
 70      0.0004953742027282715,
 71      -0.00016576051712036133,
 72      -0.0004305839538574219,
 73      -1.7344951629638672e-05,
 74      0.0011549293994903564
 75    ],
 76    "p_value": 0.2419,
 77    "mde": 0.00042210225891124443,
 78    "mde_rel_pct": 0.12476355189140237,
 79    "verdict": "no measurable effect",
 80    "system_worked": false
 81  },
 82  "idea_sweep": [
 83    {
 84      "cfg": {
 85        "lr": 0.001,
 86        "lambda_cov": 0.0001
 87      },
 88      "result": {
 89        "mean": 0.33857686817646027,
 90        "std": 0.04627190481655083,
 91        "per_seed": [
 92          0.3314558267593384,
 93          0.2857537865638733,
 94          0.30147865414619446,
 95          0.36729896068573,
 96          0.3336082696914673,
 97          0.28876644372940063,
 98          0.43178096413612366,
 99          0.36847203969955444
100        ],
101        "n": 8
102      }
103    },
104    {
105      "cfg": {
106        "lr": 0.001,
107        "lambda_cov": 0.001
108      },
109      "result": {
110        "mean": 0.33870137482881546,
111        "std": 0.04630866547362218,
112        "per_seed": [
113          0.3318597674369812,
114          0.28587043285369873,
115          0.30147260427474976,
116          0.3680233657360077,
117          0.3338707685470581,
118          0.2883317470550537,
119          0.4315110445022583,
120          0.3686712682247162
121        ],
122        "n": 8
123      }
124    },
125    {
126      "cfg": {
127        "lr": 0.009,
128        "lambda_cov": 0.01
129      },
130      "result": {
131        "mean": 1.0103588178753853,
132        "std": 0.2665844690585086,
133        "per_seed": [
134          1.0186030864715576,
135          0.9780297875404358,
136          0.7393418550491333,
137          1.433849811553955,
138          1.3804410696029663,
139          0.6398636698722839,
140          0.8194577097892761,
141          1.0732835531234741
142        ],
143        "n": 8
144      }
145    }
146  ],
147  "epochs": 12,
148  "batch": 64,
149  "mechanism_signature": {
150    "quantity": "pooled input-projection covariance deviation on held-out benchmark windows",
151    "predicted_finite_width_scale_at_width_64": 0.125,
152    "observed_baseline_mean": 0.10427584871649742,
153    "observed_idea_mean": 0.10427427012473345,
154    "observed_idea_over_baseline": 0.9999848613865684,
155    "confirmed": false,
156    "note": "The O(n^-1/2) prediction is a scale law, not an absolute covariance value; this signature tests reduction on trained systems."
157  }
158}