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      "weight_decay": 0.0,
 11      "epochs": 8,
 12      "adv_lr": 0.08,
 13      "lambda_lr": 0.2,
 14      "radius_scale": 0.1
 15    },
 16    "sweep": [
 17      {
 18        "cfg": {
 19          "lr": 0.001,
 20          "weight_decay": 0.0,
 21          "epochs": 8,
 22          "adv_lr": 0.08,
 23          "lambda_lr": 0.2,
 24          "radius_scale": 0.1
 25        },
 26        "mean": 0.41393087804317474
 27      },
 28      {
 29        "cfg": {
 30          "lr": 0.001,
 31          "weight_decay": 0.0001,
 32          "epochs": 8,
 33          "adv_lr": 0.08,
 34          "lambda_lr": 0.2,
 35          "radius_scale": 0.1
 36        },
 37        "mean": 0.41419199854135513
 38      },
 39      {
 40        "cfg": {
 41          "lr": 0.003,
 42          "weight_decay": 0.0,
 43          "epochs": 8,
 44          "adv_lr": 0.08,
 45          "lambda_lr": 0.2,
 46          "radius_scale": 0.1
 47        },
 48        "mean": 0.46493180841207504
 49      },
 50      {
 51        "cfg": {
 52          "lr": 0.003,
 53          "weight_decay": 0.0001,
 54          "epochs": 8,
 55          "adv_lr": 0.08,
 56          "lambda_lr": 0.2,
 57          "radius_scale": 0.1
 58        },
 59        "mean": 0.4638128876686096
 60      },
 61      {
 62        "cfg": {
 63          "lr": 0.006,
 64          "weight_decay": 0.0,
 65          "epochs": 8,
 66          "adv_lr": 0.08,
 67          "lambda_lr": 0.2,
 68          "radius_scale": 0.1
 69        },
 70        "mean": 0.667711153626442
 71      },
 72      {
 73        "cfg": {
 74          "lr": 0.006,
 75          "weight_decay": 0.0001,
 76          "epochs": 8,
 77          "adv_lr": 0.08,
 78          "lambda_lr": 0.2,
 79          "radius_scale": 0.1
 80        },
 81        "mean": 0.5989826694130898
 82      }
 83    ],
 84    "full": {
 85      "mean": 0.43296366557478905,
 86      "std": 0.0464950719362365,
 87      "per_seed": [
 88        0.4563036859035492,
 89        0.3589187264442444,
 90        0.3965584337711334,
 91        0.443942666053772,
 92        0.4800455868244171,
 93        0.38371741771698,
 94        0.5046233534812927,
 95        0.4395994544029236
 96      ],
 97      "n": 8
 98    }
 99  },
100  "idea": {
101    "mean": 0.4326546713709831,
102    "std": 0.046648733640043204,
103    "per_seed": [
104      0.45671361684799194,
105      0.3578817546367645,
106      0.3963860869407654,
107      0.44397932291030884,
108      0.4796009659767151,
109      0.3833600878715515,
110      0.50420743227005,
111      0.43910810351371765
112    ],
113    "n": 8
114  },
115  "comparison": {
116    "delta_mean": -0.00030899420380592346,
117    "idea_wins": 6,
118    "n_pairs": 8,
119    "per_seed_diffs": [
120      0.000409930944442749,
121      -0.0010369718074798584,
122      -0.000172346830368042,
123      3.6656856536865234e-05,
124      -0.00044462084770202637,
125      -0.0003573298454284668,
126      -0.0004159212112426758,
127      -0.0004913508892059326
128    ],
129    "p_value": 0.0797,
130    "mde": 0.00035339177440740587,
131    "mde_rel_pct": 0.0816215776301353,
132    "verdict": "no measurable effect",
133    "system_worked": false
134  },
135  "mechanism_signature": {
136    "math_and_nn_signature": {
137      "prediction": "translation S approximately k*delta^2",
138      "fitted_k": 0.7166269634451184,
139      "max_relative_error": 0.022699180966030196,
140      "S_values": [
141        0.0,
142        0.007007211446762085,
143        0.02810005284845829,
144        0.0634879618883133,
145        0.11350420862436295,
146        0.17858079075813293,
147        0.25921881198883057
148      ],
149      "debiased_identity": 0.0,
150      "confirmed": true,
151      "nn_observed_mean_sinkhorn": 9.942629664028146e-08,
152      "nn_observed_mean_lambda": 0.0,
153      "nn_predicted_radius": 0.035
154    },
155    "idea_sweep": [
156      {
157        "cfg": {
158          "lr": 0.001,
159          "weight_decay": 0.0,
160          "epochs": 8,
161          "adv_lr": 0.08,
162          "lambda_lr": 0.2,
163          "radius_scale": 0.1
164        },
165        "mean": 0.41374019533395767
166      },
167      {
168        "cfg": {
169          "lr": 0.003,
170          "weight_decay": 0.0,
171          "epochs": 8,
172          "adv_lr": 0.08,
173          "lambda_lr": 0.2,
174          "radius_scale": 0.1
175        },
176        "mean": 0.4662386029958725
177      },
178      {
179        "cfg": {
180          "lr": 0.001,
181          "weight_decay": 0.0,
182          "epochs": 8,
183          "adv_lr": 0.04,
184          "lambda_lr": 0.2,
185          "radius_scale": 0.07
186        },
187        "mean": 0.4139372706413269
188      }
189    ],
190    "custom_track": null,
191    "track_rationale": "Sequence forecasting has context-conditioned multi-token windows; residual trajectory augmentation acts on the generated sequence manifold."
192  }
193}