Two-sided conditioned DFA / bench_report.json

Failed on benchmark

Raw ⬇ ZIP
  1{
  2  "bench_version": 1,
  3  "track": "tabular",
  4  "model": "mlp_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": 17.2008535861969
 17      },
 18      {
 19        "cfg": {
 20          "lr": 0.003
 21        },
 22        "mean": 17.275779485702515
 23      },
 24      {
 25        "cfg": {
 26          "lr": 0.01
 27        },
 28        "mean": 18.10535430908203
 29      }
 30    ],
 31    "full": {
 32      "mean": 17.477807760238647,
 33      "std": 1.5610336422417328,
 34      "per_seed": [
 35        18.36859893798828,
 36        17.082815170288086,
 37        17.65996551513672,
 38        15.692034721374512,
 39        14.8045015335083,
 40        18.059467315673828,
 41        17.911405563354492,
 42        20.24367332458496
 43      ],
 44      "n": 8
 45    }
 46  },
 47  "idea": {
 48    "mean": 22.206230878829956,
 49    "std": 1.2034773425543663,
 50    "per_seed": [
 51      22.064266204833984,
 52      22.58884048461914,
 53      20.288148880004883,
 54      21.367250442504883,
 55      21.962797164916992,
 56      21.877431869506836,
 57      22.725826263427734,
 58      24.775285720825195
 59    ],
 60    "n": 8,
 61    "selected_cfg": {
 62      "lr": 0.01
 63    }
 64  },
 65  "comparison": {
 66    "delta_mean": 4.728423118591309,
 67    "idea_wins": 0,
 68    "n_pairs": 8,
 69    "per_seed_diffs": [
 70      3.695667266845703,
 71      5.506025314331055,
 72      2.628183364868164,
 73      5.675215721130371,
 74      7.158295631408691,
 75      3.817964553833008,
 76      4.814420700073242,
 77      4.531612396240234
 78    ],
 79    "p_value": 0.0081,
 80    "mde": 1.172236051535733,
 81    "mde_rel_pct": 6.70699705372962,
 82    "verdict": "idea worse (significant)",
 83    "system_worked": false
 84  },
 85  "mechanism_signature": {
 86    "mechanism_signature": {
 87      "prediction": "factorized Cholesky-preconditioned outer product equals matrix update",
 88      "per_seed": [
 89        {
 90          "factorization_abs_err": 0.0021419525146484375,
 91          "factorization_rel_err": 1.424667971150484e-05,
 92          "raw_update_norm": 99.73847961425781,
 93          "conditioned_update_norm": 471.4458923339844,
 94          "predicted_norm_ratio": 4.726820468902588,
 95          "trained_activation_trace": 3.3609566688537598,
 96          "trained_error_trace": 22.064266204833984,
 97          "confirmed": false
 98        },
 99        {
100          "factorization_abs_err": 0.0013217926025390625,
101          "factorization_rel_err": 9.378113645652775e-06,
102          "raw_update_norm": 119.99061584472656,
103          "conditioned_update_norm": 431.77825927734375,
104          "predicted_norm_ratio": 3.598433494567871,
105          "trained_activation_trace": 3.3852806091308594,
106          "trained_error_trace": 22.588836669921875,
107          "confirmed": true
108        },
109        {
110          "factorization_abs_err": 0.0012187957763671875,
111          "factorization_rel_err": 7.294798706425354e-06,
112          "raw_update_norm": 106.99031829833984,
113          "conditioned_update_norm": 435.8670654296875,
114          "predicted_norm_ratio": 4.073892593383789,
115          "trained_activation_trace": 3.3690006732940674,
116          "trained_error_trace": 20.288150787353516,
117          "confirmed": true
118        },
119        {
120          "factorization_abs_err": 0.0012369155883789062,
121          "factorization_rel_err": 8.22768470243318e-06,
122          "raw_update_norm": 88.57902526855469,
123          "conditioned_update_norm": 449.8793640136719,
124          "predicted_norm_ratio": 5.078847408294678,
125          "trained_activation_trace": 3.282832145690918,
126          "trained_error_trace": 21.367252349853516,
127          "confirmed": true
128        },
129        {
130          "factorization_abs_err": 0.0022106170654296875,
131          "factorization_rel_err": 1.4202377315086778e-05,
132          "raw_update_norm": 126.09120178222656,
133          "conditioned_update_norm": 424.4380798339844,
134          "predicted_norm_ratio": 3.366119623184204,
135          "trained_activation_trace": 3.368903636932373,
136          "trained_error_trace": 21.96279525756836,
137          "confirmed": false
138        },
139        {
140          "factorization_abs_err": 0.001743316650390625,
141          "factorization_rel_err": 9.334406058769673e-06,
142          "raw_update_norm": 113.74923706054688,
143          "conditioned_update_norm": 444.450927734375,
144          "predicted_norm_ratio": 3.907287120819092,
145          "trained_activation_trace": 3.3790221214294434,
146          "trained_error_trace": 21.877431869506836,
147          "confirmed": true
148        },
149        {
150          "factorization_abs_err": 0.002315521240234375,
151          "factorization_rel_err": 1.217806766362628e-05,
152          "raw_update_norm": 90.95088195800781,
153          "conditioned_update_norm": 474.05853271484375,
154          "predicted_norm_ratio": 5.212247848510742,
155          "trained_activation_trace": 3.336045265197754,
156          "trained_error_trace": 22.7258243560791,
157          "confirmed": false
158        },
159        {
160          "factorization_abs_err": 0.0021305084228515625,
161          "factorization_rel_err": 1.0202298653894104e-05,
162          "raw_update_norm": 137.70616149902344,
163          "conditioned_update_norm": 439.57916259765625,
164          "predicted_norm_ratio": 3.192153215408325,
165          "trained_activation_trace": 3.364262580871582,
166          "trained_error_trace": 24.775285720825195,
167          "confirmed": false
168        }
169      ],
170      "confirmed": false,
171      "mean_observed_norm_ratio": 4.144475221633911
172    },
173    "track_rationale": "Two-sided conditioning is an optimizer/update-rule idea; tabular is the prescribed optimizer track.",
174    "idea_sweep": [
175      {
176        "cfg": {
177          "lr": 0.001
178        },
179        "mean": 200.41951751708984
180      },
181      {
182        "cfg": {
183          "lr": 0.003
184        },
185        "mean": 104.71292400360107
186      },
187      {
188        "cfg": {
189          "lr": 0.01
190        },
191        "mean": 22.206230878829956
192      }
193    ],
194    "selected_idea_cfg": {
195      "lr": 0.01
196    },
197    "elapsed_sec": 17.28634238243103
198  }
199}