Phase-Margin Residual Jacobians / bench_report.json

Mechanism confirmed, baseline not beaten

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": 0.001,
 10      "epochs": 12
 11    },
 12    "sweep": [
 13      {
 14        "cfg": {
 15          "lr": 0.001,
 16          "epochs": 12
 17        },
 18        "per_seed": [
 19          0.029528150334954262,
 20          0.04754278063774109,
 21          0.033632513135671616,
 22          0.031397681683301926,
 23          0.029850929975509644,
 24          0.12557671964168549,
 25          0.022646039724349976,
 26          0.05035697668790817
 27        ],
 28        "mean": 0.04631647397764027,
 29        "std": 0.031218425382336913,
 30        "n": 8
 31      },
 32      {
 33        "cfg": {
 34          "lr": 0.003,
 35          "epochs": 12
 36        },
 37        "per_seed": [
 38          0.05427622050046921,
 39          0.028770923614501953,
 40          0.41897544264793396,
 41          0.025073563680052757,
 42          0.07675475627183914,
 43          0.16374778747558594,
 44          0.06638805568218231,
 45          0.09212644398212433
 46        ],
 47        "mean": 0.1157641492318362,
 48        "std": 0.12161998162703432,
 49        "n": 8
 50      },
 51      {
 52        "cfg": {
 53          "lr": 0.01,
 54          "epochs": 12
 55        },
 56        "per_seed": [
 57          0.68760746717453,
 58          0.04096516966819763,
 59          0.026457389816641808,
 60          0.32609719038009644,
 61          0.01830579899251461,
 62          0.08879950642585754,
 63          0.2804693281650543,
 64          0.6182840466499329
 65        ],
 66        "mean": 0.26087323715910316,
 67        "std": 0.25155552050299584,
 68        "n": 8
 69      }
 70    ],
 71    "harness_tuning": {
 72      "best_cfg": {
 73        "lr": 0.001,
 74        "epochs": 12
 75      },
 76      "sweep": [
 77        {
 78          "cfg": {
 79            "lr": 0.001,
 80            "epochs": 12
 81          },
 82          "mean": 0.03552528144791722
 83        },
 84        {
 85          "cfg": {
 86            "lr": 0.003,
 87            "epochs": 12
 88          },
 89          "mean": 0.13177403761073947
 90        },
 91        {
 92          "cfg": {
 93            "lr": 0.01,
 94            "epochs": 12
 95          },
 96          "mean": 0.2702818042598665
 97        }
 98      ],
 99      "full": {
100        "mean": 0.04631647397764027,
101        "std": 0.031218425382336913,
102        "per_seed": [
103          0.029528150334954262,
104          0.04754278063774109,
105          0.033632513135671616,
106          0.031397681683301926,
107          0.029850929975509644,
108          0.12557671964168549,
109          0.022646039724349976,
110          0.05035697668790817
111        ],
112        "n": 8
113      }
114    },
115    "full": {
116      "per_seed": [
117        0.029528150334954262,
118        0.04754278063774109,
119        0.033632513135671616,
120        0.031397681683301926,
121        0.029850929975509644,
122        0.12557671964168549,
123        0.022646039724349976,
124        0.05035697668790817
125      ],
126      "mean": 0.04631647397764027,
127      "std": 0.031218425382336913,
128      "n": 8
129    }
130  },
131  "idea": {
132    "per_seed": [
133      0.029486622661352158,
134      0.047586631029844284,
135      0.03362056985497475,
136      0.03137165307998657,
137      0.02987625077366829,
138      0.11335577070713043,
139      0.022612692788243294,
140      0.05035630613565445
141    ],
142    "mean": 0.04478331212885678,
143    "std": 0.0273708875835822,
144    "n": 8
145  },
146  "comparison": {
147    "delta_mean": -0.001533161848783493,
148    "idea_wins": 6,
149    "n_pairs": 8,
150    "per_seed_diffs": [
151      -4.152767360210419e-05,
152      4.385039210319519e-05,
153      -1.1943280696868896e-05,
154      -2.6028603315353394e-05,
155      2.532079815864563e-05,
156      -0.012220948934555054,
157      -3.3346936106681824e-05,
158      -6.705522537231445e-07
159    ],
160    "p_value": 0.2886,
161    "mde": 0.003611027360643652,
162    "mde_rel_pct": 7.796421123047731,
163    "verdict": "no significant win",
164    "system_worked": false
165  },
166  "mechanism_signature": {
167    "idea_sweep": [
168      {
169        "cfg": {
170          "lr": 0.001,
171          "epochs": 12
172        },
173        "per_seed": [
174          0.029486622661352158,
175          0.047586631029844284,
176          0.03362056985497475,
177          0.03137165307998657,
178          0.02987625077366829,
179          0.11335577070713043,
180          0.022612692788243294,
181          0.05035630613565445
182        ],
183        "mean": 0.04478331212885678,
184        "std": 0.0273708875835822,
185        "n": 8
186      },
187      {
188        "cfg": {
189          "lr": 0.003,
190          "epochs": 12
191        },
192        "per_seed": [
193          0.054160747677087784,
194          0.14119376242160797,
195          0.014620370231568813,
196          0.03189314156770706,
197          0.04598933085799217,
198          0.3756285011768341,
199          0.04325627163052559,
200          0.07682377099990845
201        ],
202        "mean": 0.097945737070404,
203        "std": 0.110870329571024,
204        "n": 8
205      },
206      {
207        "cfg": {
208          "lr": 0.01,
209          "epochs": 12
210        },
211        "per_seed": [
212          0.9418538212776184,
213          0.0459330677986145,
214          0.025276441127061844,
215          0.38526275753974915,
216          0.034148477017879486,
217          0.09575814008712769,
218          0.09502670913934708,
219          6.104557991027832
220        ],
221        "mean": 0.9659771756269038,
222        "std": 1.963999860256041,
223        "n": 8
224      }
225    ],
226    "paired_deltas_idea_minus_baseline": [
227      -4.152767360210419e-05,
228      4.385039210319519e-05,
229      -1.1943280696868896e-05,
230      -2.6028603315353394e-05,
231      2.532079815864563e-05,
232      -0.012220948934555054,
233      -3.3346936106681824e-05,
234      -6.705522537231445e-07
235    ],
236    "permutation_pvalue": 0.2886,
237    "signature": {
238      "prediction": "phase certificate excess is suppressed by the Jacobian penalty",
239      "predicted_excess_values": [
240        0.8990192516358979,
241        0.941527019313205,
242        0.8427875144990571,
243        0.8680509431871064
244      ],
245      "observed_block_gain_proxy": [
246        1.048457384109497,
247        1.0909651517868042,
248        0.9922256469726562,
249        1.0174890756607056
250      ],
251      "trained_test_mse": 0.029486622661352158,
252      "confirmed": true
253    },
254    "track_selection": "dynamics: actuated pendulum stability/control has recurrent propagation structure."
255  },
256  "custom_track": null
257}