BDD-Certified Modular Equilibrium Network / 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": 0.006,
 10      "wd": 0.0
 11    },
 12    "sweep": [
 13      {
 14        "cfg": {
 15          "lr": 0.0015,
 16          "wd": 0.0
 17        },
 18        "mean": 0.00032828053372213617
 19      },
 20      {
 21        "cfg": {
 22          "lr": 0.0015,
 23          "wd": 0.0001
 24        },
 25        "mean": 0.0004147848885622807
 26      },
 27      {
 28        "cfg": {
 29          "lr": 0.003,
 30          "wd": 0.0
 31        },
 32        "mean": 4.9621185098658316e-05
 33      },
 34      {
 35        "cfg": {
 36          "lr": 0.003,
 37          "wd": 0.0001
 38        },
 39        "mean": 0.00020817978656850755
 40      },
 41      {
 42        "cfg": {
 43          "lr": 0.006,
 44          "wd": 0.0
 45        },
 46        "mean": 2.345215943932999e-05
 47      },
 48      {
 49        "cfg": {
 50          "lr": 0.006,
 51          "wd": 0.0001
 52        },
 53        "mean": 0.00010142060364159988
 54      }
 55    ],
 56    "full": {
 57      "mean": 2.282289460708853e-05,
 58      "std": 2.305330446600017e-06,
 59      "per_seed": [
 60        2.4922346710809506e-05,
 61        2.6788266040966846e-05,
 62        1.934224746946711e-05,
 63        2.2755777536076494e-05,
 64        2.3468126528314315e-05,
 65        2.385292282269802e-05,
 66        2.0721481632790528e-05,
 67        2.073198811558541e-05
 68      ],
 69      "n": 8
 70    }
 71  },
 72  "idea": {
 73    "mean": 4.757533133670222e-05,
 74    "std": 2.1643697598220244e-05,
 75    "per_seed": [
 76      9.76939991232939e-05,
 77      3.8501457311213017e-05,
 78      3.3765736588975415e-05,
 79      2.9315116989891976e-05,
 80      6.082628169679083e-05,
 81      5.370224607759155e-05,
 82      3.0736955523025244e-05,
 83      3.606085738283582e-05
 84    ],
 85    "n": 8
 86  },
 87  "comparison": {
 88    "delta_mean": 2.475243672961369e-05,
 89    "idea_wins": 0,
 90    "n_pairs": 8,
 91    "per_seed_diffs": [
 92      7.27716524124844e-05,
 93      1.171319127024617e-05,
 94      1.4423489119508304e-05,
 95      6.5593394538154826e-06,
 96      3.7358155168476515e-05,
 97      2.9849323254893534e-05,
 98      1.0015473890234716e-05,
 99      1.532886926725041e-05
100    ],
101    "p_value": 0.0081,
102    "mde": 1.8439870679175355e-05,
103    "mde_rel_pct": 80.79549503527105,
104    "verdict": "idea worse (significant)",
105    "system_worked": false
106  },
107  "mechanism_signature": {
108    "signature": {
109      "prediction": "idea lowers trained recurrent cross-block q and keeps it near <=0.8",
110      "baseline_q_mean": 2.9282636810198936,
111      "idea_q_mean": 0.7976552316857172,
112      "baseline_q_per_seed": [
113        2.7491974371903396,
114        2.8032834877024646,
115        2.7482760381642044,
116        3.334465893518113,
117        2.8231071925105216,
118        3.0133113893982655,
119        2.993095970408364,
120        2.9613720392668768
121      ],
122      "idea_q_per_seed": [
123        0.7953895769092711,
124        0.7993655229735372,
125        0.7991708584055264,
126        0.79743240619884,
127        0.7974510193922437,
128        0.798977436754247,
129        0.7961304629450436,
130        0.7973245699070284
131      ],
132      "observed_reduction": 2.1306084493341766,
133      "target_q": 0.8,
134      "confirmed": true
135    },
136    "idea_sweep": [
137      {
138        "cfg": {
139          "lr": 0.006,
140          "lam": 0.1
141        },
142        "result": {
143          "mean": 4.757533133670222e-05,
144          "std": 2.1643697598220244e-05,
145          "per_seed": [
146            9.76939991232939e-05,
147            3.8501457311213017e-05,
148            3.3765736588975415e-05,
149            2.9315116989891976e-05,
150            6.082628169679083e-05,
151            5.370224607759155e-05,
152            3.0736955523025244e-05,
153            3.606085738283582e-05
154          ],
155          "n": 8
156        }
157      },
158      {
159        "cfg": {
160          "lr": 0.0015,
161          "lam": 0.1
162        },
163        "result": {
164          "mean": 0.00037327232348616235,
165          "std": 3.6520547820565044e-05,
166          "per_seed": [
167            0.0003330762847326696,
168            0.00043403543531894684,
169            0.00033968265051953495,
170            0.000401037948904559,
171            0.00037536423769779503,
172            0.0003695053164847195,
173            0.0003257567877881229,
174            0.00040771992644295096
175          ],
176          "n": 8
177        }
178      },
179      {
180        "cfg": {
181          "lr": 0.006,
182          "lam": 0.1
183        },
184        "result": {
185          "mean": 4.757533133670222e-05,
186          "std": 2.1643697598220244e-05,
187          "per_seed": [
188            9.76939991232939e-05,
189            3.8501457311213017e-05,
190            3.3765736588975415e-05,
191            2.9315116989891976e-05,
192            6.082628169679083e-05,
193            5.370224607759155e-05,
194            3.0736955523025244e-05,
195            3.606085738283582e-05
196          ],
197          "n": 8
198        }
199      }
200    ],
201    "bdd_delta": 0.2
202  },
203  "protocol_notes": "Dynamics is structurally matched: controlled pendulum rollout and recurrent state coupling. Baseline and idea share rnn_small, data, Adam, epochs, batch, lr/weight-decay grids."
204}