Commutant-Gap Controlled Stochastic Training / bench_report.json

Failed on benchmark

Raw ⬇ ZIP
  1{
  2  "bench_version": 1,
  3  "track": "dynamics",
  4  "model": "rnn_small",
  5  "metric_direction": "lower is better",
  6  "n_seeds": 8,
  7  "baseline": {
  8    "best_cfg": {
  9      "lr": 0.003,
 10      "weight_decay": 0.0,
 11      "kappa": 0.0
 12    },
 13    "sweep": [
 14      {
 15        "cfg": {
 16          "lr": 0.001,
 17          "weight_decay": 0.0,
 18          "kappa": 0.0
 19        },
 20        "mean": 0.4700177311897278
 21      },
 22      {
 23        "cfg": {
 24          "lr": 0.001,
 25          "weight_decay": 0.0001,
 26          "kappa": 0.0
 27        },
 28        "mean": 0.47006765753030777
 29      },
 30      {
 31        "cfg": {
 32          "lr": 0.003,
 33          "weight_decay": 0.0,
 34          "kappa": 0.0
 35        },
 36        "mean": 0.02822397626005113
 37      },
 38      {
 39        "cfg": {
 40          "lr": 0.003,
 41          "weight_decay": 0.0001,
 42          "kappa": 0.0
 43        },
 44        "mean": 0.02829078072682023
 45      },
 46      {
 47        "cfg": {
 48          "lr": 0.01,
 49          "weight_decay": 0.0,
 50          "kappa": 0.0
 51        },
 52        "mean": 0.03205677215009928
 53      },
 54      {
 55        "cfg": {
 56          "lr": 0.01,
 57          "weight_decay": 0.0001,
 58          "kappa": 0.0
 59        },
 60        "mean": 0.03197645582258701
 61      }
 62    ],
 63    "full": {
 64      "mean": 0.03168170212302357,
 65      "std": 0.018327317088355676,
 66      "per_seed": [
 67        0.06044498085975647,
 68        0.014531154185533524,
 69        0.01185071561485529,
 70        0.026069054380059242,
 71        0.05299823358654976,
 72        0.023797301575541496,
 73        0.014214038848876953,
 74        0.04954813793301582
 75      ],
 76      "n": 8
 77    }
 78  },
 79  "idea": {
 80    "mean": 0.030110739055089653,
 81    "std": 0.018290232592247106,
 82    "per_seed": [
 83      0.062178634107112885,
 84      0.010466991923749447,
 85      0.017325958237051964,
 86      0.034337423741817474,
 87      0.055281877517700195,
 88      0.03069481998682022,
 89      0.01328994520008564,
 90      0.017310261726379395
 91    ],
 92    "n": 8,
 93    "best_cfg": {
 94      "lr": 0.003,
 95      "weight_decay": 0.0,
 96      "kappa": 0.002
 97    },
 98    "sweep": [
 99      {
100        "cfg": {
101          "lr": 0.003,
102          "weight_decay": 0.0,
103          "kappa": 0.002
104        },
105        "mean": 0.030110739055089653,
106        "per_seed": [
107          0.062178634107112885,
108          0.010466991923749447,
109          0.017325958237051964,
110          0.034337423741817474,
111          0.055281877517700195,
112          0.03069481998682022,
113          0.01328994520008564,
114          0.017310261726379395
115        ]
116      },
117      {
118        "cfg": {
119          "lr": 0.003,
120          "weight_decay": 0.0,
121          "kappa": 0.005
122        },
123        "mean": 0.030580853228457272,
124        "per_seed": [
125          0.06277664005756378,
126          0.010813195258378983,
127          0.017922919243574142,
128          0.03458359092473984,
129          0.056420501321554184,
130          0.030933773145079613,
131          0.013734272681176662,
132          0.017461933195590973
133        ]
134      },
135      {
136        "cfg": {
137          "lr": 0.003,
138          "weight_decay": 0.0,
139          "kappa": 0.01
140        },
141        "mean": 0.03123972751200199,
142        "per_seed": [
143          0.06403753161430359,
144          0.010984514839947224,
145          0.01874014548957348,
146          0.034955769777297974,
147          0.05827799811959267,
148          0.031186075881123543,
149          0.014287841506302357,
150          0.0174479428678751
151        ]
152      }
153    ]
154  },
155  "comparison": {
156    "delta_mean": -0.001570963067933917,
157    "idea_wins": 3,
158    "n_pairs": 8,
159    "per_seed_diffs": [
160      0.0017336532473564148,
161      -0.004064162261784077,
162      0.005475242622196674,
163      0.008268369361758232,
164      0.0022836439311504364,
165      0.006897518411278725,
166      -0.0009240936487913132,
167      -0.03223787620663643
168    ],
169    "p_value": 0.93175,
170    "mde": 0.010906916277571396,
171    "mde_rel_pct": 34.426547649550606,
172    "verdict": "no significant win",
173    "system_worked": false
174  },
175  "mechanism_signature": {
176    "math_sanity": {
177      "predicted_rate": 1.4,
178      "observed_rate": 1.288026525213435,
179      "relative_error": 0.07998105341897488,
180      "passed": true
181    },
182    "nn_scale": {
183      "trained_model_metric": 0.062178634107112885,
184      "predicted_rate": 0.008,
185      "observed_rate": 0.009309132871376703,
186      "relative_error": 0.16364160892208784,
187      "confirmed": true,
188      "mean_gap": 0.03696886791537205
189    }
190  },
191  "track_justification": "Dynamics is the structural match because the idea targets recurrent hidden-state stability and controlled relaxation."
192}