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    "sweep": {
  9      "best_cfg": {
 10        "lr": 0.003,
 11        "rounds": 4
 12      },
 13      "sweep": [
 14        {
 15          "cfg": {
 16            "lr": 0.001,
 17            "rounds": 4
 18          },
 19          "mean": 0.062015012837946415
 20        },
 21        {
 22          "cfg": {
 23            "lr": 0.003,
 24            "rounds": 4
 25          },
 26          "mean": 0.02783486293628812
 27        },
 28        {
 29          "cfg": {
 30            "lr": 0.01,
 31            "rounds": 4
 32          },
 33          "mean": 0.031387015245854855
 34        }
 35      ],
 36      "full": {
 37        "mean": 0.03436466213315725,
 38        "std": 0.012433180922138218,
 39        "per_seed": [
 40          0.028132088482379913,
 41          0.009857414290308952,
 42          0.046546123921871185,
 43          0.02681129425764084,
 44          0.028106732293963432,
 45          0.046415776014328,
 46          0.045219533145427704,
 47          0.043828334659338
 48        ],
 49        "n": 8
 50      }
 51    },
 52    "best_cfg": {
 53      "lr": 0.003,
 54      "rounds": 4
 55    },
 56    "full": {
 57      "mean": 0.03436466213315725,
 58      "std": 0.012433180922138218,
 59      "per_seed": [
 60        0.028132088482379913,
 61        0.009857414290308952,
 62        0.046546123921871185,
 63        0.02681129425764084,
 64        0.028106732293963432,
 65        0.046415776014328,
 66        0.045219533145427704,
 67        0.043828334659338
 68      ],
 69      "n": 8
 70    }
 71  },
 72  "idea": {
 73    "mean": 0.0322596225887537,
 74    "std": 0.01605094247073741,
 75    "per_seed": [
 76      0.008163033053278923,
 77      0.04895344749093056,
 78      0.02211311273276806,
 79      0.046254031360149384,
 80      0.026844650506973267,
 81      0.018853165209293365,
 82      0.028555940836668015,
 83      0.05833959951996803
 84    ],
 85    "n": 8
 86  },
 87  "comparison": {
 88    "delta_mean": -0.002105039544403553,
 89    "idea_wins": 5,
 90    "n_pairs": 8,
 91    "per_seed_diffs": [
 92      -0.01996905542910099,
 93      0.039096033200621605,
 94      -0.024433011189103127,
 95      0.019442737102508545,
 96      -0.0012620817869901657,
 97      -0.027562610805034637,
 98      -0.01666359230875969,
 99      0.014511264860630035
100    ],
101    "p_value": 0.80015,
102    "mde": 0.020276508456807906,
103    "mde_rel_pct": 59.003951146791046,
104    "verdict": "no significant win",
105    "system_worked": false
106  },
107  "protocol_note": "Matched dynamics task and freshly trained identical rnn_small architecture; only the inference stopping rule differs.",
108  "mechanism_signature": {
109    "quantity": "NN observed disagreement contraction",
110    "predicted_mean_ratio": 0.29535447060123415,
111    "observed_mean_ratio": 1.0081205979778152,
112    "absolute_error": 0.7127661273765811,
113    "n_trajectories": 8,
114    "confirmed": false
115  },
116  "idea_sweep": [
117    {
118      "cfg": {
119        "lr": 0.001,
120        "rounds": 4
121      },
122      "mean": 0.06841936893761158
123    },
124    {
125      "cfg": {
126        "lr": 0.003,
127        "rounds": 4
128      },
129      "mean": 0.034365283441729844
130    },
131    {
132      "cfg": {
133        "lr": 0.01,
134        "rounds": 4
135      },
136      "mean": 0.0322564683156088
137    }
138  ],
139  "idea_best_cfg": {
140    "lr": 0.01,
141    "rounds": 4
142  },
143  "idea_auxiliary_rounds": [
144    {
145      "rounds": 4,
146      "changes": [
147        0.026034874841570854,
148        0.025428546592593193,
149        0.02514001540839672
150      ]
151    },
152    {
153      "rounds": 4,
154      "changes": [
155        0.0409514456987381,
156        0.042881350964307785,
157        0.04315655678510666
158      ]
159    },
160    {
161      "rounds": 4,
162      "changes": [
163        0.03505569323897362,
164        0.03519539535045624,
165        0.03512731194496155
166      ]
167    },
168    {
169      "rounds": 4,
170      "changes": [
171        0.042443498969078064,
172        0.043852899223566055,
173        0.04409702122211456
174      ]
175    },
176    {
177      "rounds": 4,
178      "changes": [
179        0.03753160312771797,
180        0.038068048655986786,
181        0.03855843096971512
182      ]
183    },
184    {
185      "rounds": 4,
186      "changes": [
187        0.03447379544377327,
188        0.034757938235998154,
189        0.034541524946689606
190      ]
191    },
192    {
193      "rounds": 4,
194      "changes": [
195        0.03614768758416176,
196        0.03720034658908844,
197        0.03734774887561798
198      ]
199    },
200    {
201      "rounds": 4,
202      "changes": [
203        0.04730474203824997,
204        0.04924433305859566,
205        0.04874498397111893
206      ]
207    }
208  ]
209}