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": 0.2947123125195503
 17      },
 18      {
 19        "cfg": {
 20          "lr": 0.003
 21        },
 22        "mean": 0.3950112909078598
 23      },
 24      {
 25        "cfg": {
 26          "lr": 0.006
 27        },
 28        "mean": 0.6112398952245712
 29      }
 30    ],
 31    "full": {
 32      "mean": 0.3066449798643589,
 33      "std": 0.032607185412887076,
 34      "per_seed": [
 35        0.34020674228668213,
 36        0.267076700925827,
 37        0.2600124776363373,
 38        0.31155332922935486,
 39        0.3084578514099121,
 40        0.3220774829387665,
 41        0.36002540588378906,
 42        0.28374984860420227
 43      ],
 44      "n": 8
 45    }
 46  },
 47  "idea": {
 48    "mean": 0.30605993419885635,
 49    "std": 0.03247841729165149,
 50    "per_seed": [
 51      0.3398566246032715,
 52      0.26776447892189026,
 53      0.2589544951915741,
 54      0.3121564984321594,
 55      0.3084918260574341,
 56      0.3192761540412903,
 57      0.35937950015068054,
 58      0.28259989619255066
 59    ],
 60    "n": 8
 61  },
 62  "comparison": {
 63    "delta_mean": -0.0005850456655025482,
 64    "idea_wins": 5,
 65    "n_pairs": 8,
 66    "per_seed_diffs": [
 67      -0.00035011768341064453,
 68      0.0006877779960632324,
 69      -0.0010579824447631836,
 70      0.0006031692028045654,
 71      3.3974647521972656e-05,
 72      -0.0028013288974761963,
 73      -0.0006459057331085205,
 74      -0.0011499524116516113
 75    ],
 76    "p_value": 0.18685,
 77    "mde": 0.0009443470728781248,
 78    "mde_rel_pct": 0.3079610412327136,
 79    "verdict": "no measurable effect",
 80    "system_worked": false
 81  },
 82  "mechanism_signature": {
 83    "prediction": "dual summary should have low linear overlap with primary summary (abs corr < 0.2), while Gale residual is numerical roundoff",
 84    "predicted_abs_summary_corr_max": 0.2,
 85    "observed_abs_summary_corr_mean": 0.1512247179634869,
 86    "observed_abs_summary_corr_per_seed": [
 87      0.07177071273326874,
 88      0.0552356131374836,
 89      0.29892387986183167,
 90      0.023424027487635612,
 91      0.1484474241733551,
 92      0.4907960593700409,
 93      0.1091001033782959,
 94      0.012099923565983772
 95    ],
 96    "observed_null_residual_mean": 9.662536939458732e-07,
 97    "observed_gate_mean": 0.4977433420717716,
 98    "confirmed": true
 99  },
100  "idea_sweep": [
101    {
102      "cfg": {
103        "lr": 0.001
104      },
105      "mean": 0.2946830242872238
106    },
107    {
108      "cfg": {
109        "lr": 0.003
110      },
111      "mean": 0.4010865166783333
112    },
113    {
114      "cfg": {
115        "lr": 0.006
116      },
117      "mean": 0.5471484512090683
118    }
119  ],
120  "core_math_check": {
121    "sizes": [
122      [
123        3,
124        7
125      ],
126      [
127        8,
128        32
129      ],
130      [
131        12,
132        32
133      ]
134    ],
135    "residuals": [
136      8.81410249758507e-16,
137      4.8764372361745095e-15,
138      5.559546421959475e-15
139    ],
140    "max_residual": 5.559546421959475e-15,
141    "passed": true
142  },
143  "setup": {
144    "epochs": 8,
145    "n_train": 1200,
146    "n_test": 400,
147    "a": 8,
148    "window": 32,
149    "track_justification": "sequence is structurally matched because Gale mixing operates across multi-token windows"
150  }
151}