1{
  2  "bench_version": 1,
  3  "track": "multitoken_diffusion",
  4  "model": "custom_local_vs_rg_pyramid",
  5  "metric_direction": "lower is better",
  6  "n_seeds": 8,
  7  "baseline": {
  8    "best_cfg": {
  9      "lr": 0.006,
 10      "epochs": 12,
 11      "hidden": 32
 12    },
 13    "sweep": [
 14      {
 15        "cfg": {
 16          "lr": 0.001,
 17          "epochs": 12,
 18          "hidden": 32
 19        },
 20        "mean": 0.5760585516691208
 21      },
 22      {
 23        "cfg": {
 24          "lr": 0.003,
 25          "epochs": 12,
 26          "hidden": 32
 27        },
 28        "mean": 0.5570157170295715
 29      },
 30      {
 31        "cfg": {
 32          "lr": 0.006,
 33          "epochs": 12,
 34          "hidden": 32
 35        },
 36        "mean": 0.552870586514473
 37      }
 38    ],
 39    "full": {
 40      "mean": 0.5488443523645401,
 41      "std": 0.02526189879738744,
 42      "per_seed": [
 43        0.544949471950531,
 44        0.5276203751564026,
 45        0.5781564116477966,
 46        0.5607560873031616,
 47        0.5518556833267212,
 48        0.5262078642845154,
 49        0.5900437831878662,
 50        0.5111651420593262
 51      ],
 52      "n": 8
 53    }
 54  },
 55  "idea": {
 56    "mean": 0.5474011525511742,
 57    "std": 0.025686412604992322,
 58    "per_seed": [
 59      0.547066330909729,
 60      0.5185491442680359,
 61      0.5782044529914856,
 62      0.5547541379928589,
 63      0.5519959330558777,
 64      0.5247784852981567,
 65      0.5898048281669617,
 66      0.5140559077262878
 67    ],
 68    "n": 8
 69  },
 70  "comparison": {
 71    "delta_mean": -0.0014431998133659363,
 72    "idea_wins": 4,
 73    "n_pairs": 8,
 74    "per_seed_diffs": [
 75      0.002116858959197998,
 76      -0.0090712308883667,
 77      4.8041343688964844e-05,
 78      -0.006001949310302734,
 79      0.00014024972915649414,
 80      -0.0014293789863586426,
 81      -0.00023895502090454102,
 82      0.00289076566696167
 83    ],
 84    "p_value": 0.40545,
 85    "mde": 0.0034127227373531473,
 86    "mde_rel_pct": 0.6218015586113622,
 87    "verdict": "no measurable effect",
 88    "system_worked": false
 89  },
 90  "mechanism_signature": {
 91    "prediction": "coarse pooled conditioning transmits distant-token influence beyond a radius-1 local field",
 92    "predicted_vs_observed": {
 93      "predicted": "idea distant sensitivity > baseline",
 94      "observed_idea_mean_distant_sensitivity": 0.0,
 95      "observed_baseline_mean_distant_sensitivity": 0.0
 96    },
 97    "trained_model_observations": [
 98      {
 99        "params": 4673,
100        "mean_output_sensitivity": 0.017446190118789673,
101        "distant_sensitivity_0_to_7": 0.0,
102        "near_sensitivity_0_to_1": 0.04621781036257744,
103        "train_metric": 0.5470663905143738
104      },
105      {
106        "params": 4673,
107        "mean_output_sensitivity": 0.014791442081332207,
108        "distant_sensitivity_0_to_7": 0.0,
109        "near_sensitivity_0_to_1": 0.02925388514995575,
110        "train_metric": 0.5185490846633911
111      },
112      {
113        "params": 4673,
114        "mean_output_sensitivity": 0.02319423481822014,
115        "distant_sensitivity_0_to_7": 0.0,
116        "near_sensitivity_0_to_1": 0.038679420948028564,
117        "train_metric": 0.5782045125961304
118      },
119      {
120        "params": 4673,
121        "mean_output_sensitivity": 0.020275751128792763,
122        "distant_sensitivity_0_to_7": 0.0,
123        "near_sensitivity_0_to_1": 0.04100608453154564,
124        "train_metric": 0.5547541975975037
125      },
126      {
127        "params": 4673,
128        "mean_output_sensitivity": 0.015608072280883789,
129        "distant_sensitivity_0_to_7": 0.0,
130        "near_sensitivity_0_to_1": 0.023431016132235527,
131        "train_metric": 0.5519958734512329
132      },
133      {
134        "params": 4673,
135        "mean_output_sensitivity": 0.01546372752636671,
136        "distant_sensitivity_0_to_7": 0.0,
137        "near_sensitivity_0_to_1": 0.03454026207327843,
138        "train_metric": 0.5247784852981567
139      },
140      {
141        "params": 4673,
142        "mean_output_sensitivity": 0.020734351128339767,
143        "distant_sensitivity_0_to_7": 0.0,
144        "near_sensitivity_0_to_1": 0.04659464955329895,
145        "train_metric": 0.5898047685623169
146      },
147      {
148        "params": 4673,
149        "mean_output_sensitivity": 0.017892511561512947,
150        "distant_sensitivity_0_to_7": 0.0,
151        "near_sensitivity_0_to_1": 0.038680300116539,
152        "train_metric": 0.5140558481216431
153      }
154    ],
155    "confirmed": false
156  },
157  "idea_sweep": [
158    {
159      "cfg": {
160        "lr": 0.001,
161        "epochs": 12,
162        "hidden": 32
163      },
164      "result": {
165        "mean": 0.5708411112427711,
166        "std": 0.01710534526819156,
167        "per_seed": [
168          0.5729581713676453,
169          0.55045485496521,
170          0.594555139541626,
171          0.5791563987731934,
172          0.5713737607002258,
173          0.5566545128822327,
174          0.5941076874732971,
175          0.547468364238739
176        ],
177        "n": 8
178      }
179    },
180    {
181      "cfg": {
182        "lr": 0.003,
183        "epochs": 12,
184        "hidden": 32
185      },
186      "result": {
187        "mean": 0.5542507246136665,
188        "std": 0.020406232018671196,
189        "per_seed": [
190          0.5557947158813477,
191          0.5313525199890137,
192          0.5770509243011475,
193          0.5587972402572632,
194          0.5542395710945129,
195          0.5387598276138306,
196          0.590461015701294,
197          0.5275499820709229
198        ],
199        "n": 8
200      }
201    },
202    {
203      "cfg": {
204        "lr": 0.006,
205        "epochs": 12,
206        "hidden": 32
207      },
208      "result": {
209        "mean": 0.5474011525511742,
210        "std": 0.025686412604992322,
211        "per_seed": [
212          0.547066330909729,
213          0.5185491442680359,
214          0.5782044529914856,
215          0.5547541379928589,
216          0.5519959330558777,
217          0.5247784852981567,
218          0.5898048281669617,
219          0.5140559077262878
220        ],
221        "n": 8
222      }
223    }
224  ],
225  "custom_track": {
226    "name": "multitoken_diffusion",
227    "file": "/home/maxwelhelp/all/math2nn/bench/custom_tracks/multitoken_diffusion.py",
228    "domain": "diffusion-sampling"
229  },
230  "architecture": {
231    "baseline": "local radius-1 convolutional velocity field",
232    "idea": "dyadic pooled coarse field plus local fine residual conditioning",
233    "baseline_params": 3361,
234    "idea_params": 4673
235  }
236}