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      0.006,
 10      0.0
 11    ],
 12    "sweep": [
 13      {
 14        "cfg": [
 15          0.001,
 16          0.0
 17        ],
 18        "mean": 0.0015310924500226974
 19      },
 20      {
 21        "cfg": [
 22          0.001,
 23          0.0001
 24        ],
 25        "mean": 0.0015161814808379859
 26      },
 27      {
 28        "cfg": [
 29          0.003,
 30          0.0
 31        ],
 32        "mean": 0.0008559521666029468
 33      },
 34      {
 35        "cfg": [
 36          0.003,
 37          0.0001
 38        ],
 39        "mean": 0.0008401070081163198
 40      },
 41      {
 42        "cfg": [
 43          0.006,
 44          0.0
 45        ],
 46        "mean": 0.0005687399534508586
 47      },
 48      {
 49        "cfg": [
 50          0.006,
 51          0.0001
 52        ],
 53        "mean": 0.000572658478631638
 54      }
 55    ],
 56    "full": {
 57      "mean": 0.0005194859513721894,
 58      "std": 0.0001693891077925739,
 59      "per_seed": [
 60        0.0005914251087233424,
 61        0.0007362987380474806,
 62        0.00035295303678140044,
 63        0.0005942829302512109,
 64        0.0002958264376502484,
 65        0.0005069610779173672,
 66        0.00032333534909412265,
 67        0.0007548049325123429
 68      ],
 69      "n": 8
 70    }
 71  },
 72  "idea": {
 73    "mean": 0.0005190527626837138,
 74    "std": 0.00016528570799671346,
 75    "per_seed": [
 76      0.0005434696213342249,
 77      0.0007388722151517868,
 78      0.00032320327591151,
 79      0.000590407638810575,
 80      0.00030803284607827663,
 81      0.0005485645961016417,
 82      0.0003513387346174568,
 83      0.0007485331734642386
 84    ],
 85    "n": 8
 86  },
 87  "comparison": {
 88    "delta_mean": -4.3318868847563863e-07,
 89    "idea_wins": 4,
 90    "n_pairs": 8,
 91    "per_seed_diffs": [
 92      -4.795548738911748e-05,
 93      2.573477104306221e-06,
 94      -2.974976086989045e-05,
 95      -3.8752914406359196e-06,
 96      1.2206408428028226e-05,
 97      4.1603518184274435e-05,
 98      2.8003385523334146e-05,
 99      -6.271759048104286e-06
100    ],
101    "p_value": 0.9775,
102    "mde": 2.429226669015372e-05,
103    "mde_rel_pct": 4.67621244154673,
104    "verdict": "no measurable effect",
105    "system_worked": false,
106    "idea_cfg_sweep": [
107      {
108        "cfg": [
109          0.006,
110          0.005
111        ],
112        "result": {
113          "mean": 0.0005225717504799832,
114          "std": 0.0001655050714685202,
115          "per_seed": [
116            0.0005708694807253778,
117            0.0007390371174551547,
118            0.0003488508809823543,
119            0.0006015055696479976,
120            0.00029773524147458375,
121            0.0005109026096761227,
122            0.0003550440014805645,
123            0.0007566291023977101
124          ],
125          "n": 8
126        }
127      },
128      {
129        "cfg": [
130          0.006,
131          0.01
132        ],
133        "result": {
134          "mean": 0.0005221629617153667,
135          "std": 0.00016535253075975801,
136          "per_seed": [
137            0.0005585983162745833,
138            0.0007406040094792843,
139            0.0003366354794707149,
140            0.0005960776470601559,
141            0.0003025302430614829,
142            0.0005268516251817346,
143            0.00035928949364461005,
144            0.0007567168795503676
145          ],
146          "n": 8
147        }
148      },
149      {
150        "cfg": [
151          0.006,
152          0.02
153        ],
154        "result": {
155          "mean": 0.0005190527626837138,
156          "std": 0.00016528570799671346,
157          "per_seed": [
158            0.0005434696213342249,
159            0.0007388722151517868,
160            0.00032320327591151,
161            0.000590407638810575,
162            0.00030803284607827663,
163            0.0005485645961016417,
164            0.0003513387346174568,
165            0.0007485331734642386
166          ],
167          "n": 8
168        }
169      }
170    ]
171  },
172  "mechanism_signature": {
173    "signature": {
174      "prediction": "PK training should reduce recurrent ensemble spread and phase KL to P_x relative to unregularized rnn_small",
175      "observed_baseline": {
176        "pairwise_hidden_distance": 1.643260881304741,
177        "phase_kl_to_Px": 0.6458364666976848,
178        "phase_kl_to_uniform": 0.6656140625776023
179      },
180      "observed_idea": {
181        "pairwise_hidden_distance": 1.6309331208467484,
182        "phase_kl_to_Px": 2.571629603478935,
183        "phase_kl_to_uniform": 2.7556800191465998
184      },
185      "confirmed": false
186    },
187    "idea_cfg": [
188      0.006,
189      0.02
190    ],
191    "baseline_grid": [
192      [
193        0.001,
194        0.0
195      ],
196      [
197        0.001,
198        0.0001
199      ],
200      [
201        0.003,
202        0.0
203      ],
204      [
205        0.003,
206        0.0001
207      ],
208      [
209        0.006,
210        0.0
211      ],
212      [
213        0.006,
214        0.0001
215      ]
216    ],
217    "idea_grid": [
218      [
219        0.006,
220        0.005
221      ],
222      [
223        0.006,
224        0.01
225      ],
226      [
227        0.006,
228        0.02
229      ]
230    ]
231  }
232}