Spectral Burn-In and Retrieval Switch / 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": 0.01
 10    },
 11    "sweep": [
 12      {
 13        "cfg": {
 14          "lr": 0.001
 15        },
 16        "mean": 0.02141550718806684
 17      },
 18      {
 19        "cfg": {
 20          "lr": 0.003
 21        },
 22        "mean": 0.0037718631356256083
 23      },
 24      {
 25        "cfg": {
 26          "lr": 0.01
 27        },
 28        "mean": 0.0034299326362088323
 29      }
 30    ],
 31    "full": {
 32      "mean": 0.002775895278318785,
 33      "std": 0.0017084856346187775,
 34      "per_seed": [
 35        0.0014615225372835994,
 36        0.004334237892180681,
 37        0.0012712624156847596,
 38        0.006652707699686289,
 39        0.0020110509358346462,
 40        0.0024756770581007004,
 41        0.0018914247630164027,
 42        0.0021092789247632027
 43      ],
 44      "n": 8
 45    }
 46  },
 47  "idea": {
 48    "best_config": {
 49      "lr": 0.01
 50    },
 51    "per_seed": [
 52      0.006583117414265871,
 53      0.009368831291794777,
 54      0.007072345819324255,
 55      0.01220707781612873,
 56      0.0064491755329072475,
 57      0.006343720480799675,
 58      0.006881004199385643,
 59      0.0053885444067418575
 60    ],
 61    "mean": 0.007536727120168507,
 62    "std": 0.0020598694618333747,
 63    "n": 8,
 64    "idea_grid": [
 65      {
 66        "lr": 0.001
 67      },
 68      {
 69        "lr": 0.003
 70      },
 71      {
 72        "lr": 0.01
 73      }
 74    ]
 75  },
 76  "comparison": {
 77    "delta_mean": 0.004760831841849722,
 78    "idea_wins": 0,
 79    "n_pairs": 8,
 80    "per_seed_diffs": [
 81      0.005121594876982272,
 82      0.005034593399614096,
 83      0.005801083403639495,
 84      0.005554370116442442,
 85      0.004438124597072601,
 86      0.0038680434226989746,
 87      0.00498957943636924,
 88      0.003279265481978655
 89    ],
 90    "p_value": 0.0081,
 91    "mde": 0.000711292400072098,
 92    "mde_rel_pct": 25.623891709016156,
 93    "verdict": "idea worse (significant)",
 94    "system_worked": false
 95  },
 96  "mechanism_signature": {
 97    "prediction": "strong contraction and small residual keep observer; otherwise retrieve",
 98    "observed": {
 99      "median_r": 0.9880073070526123,
100      "mean_switch_rate": 1.0,
101      "mean_q_threshold": 0.10375677701085806
102    },
103    "per_seed": [
104      {
105        "r_est": 0.9952841997146606,
106        "q_threshold": 0.11885927617549896,
107        "switch_rate": 1.0,
108        "observer_mse": 0.0014615217223763466,
109        "retrieval_mse": 0.006583117414265871
110      },
111      {
112        "r_est": 0.980730414390564,
113        "q_threshold": 0.13610804080963135,
114        "switch_rate": 1.0,
115        "observer_mse": 0.004334237426519394,
116        "retrieval_mse": 0.009368831291794777
117      },
118      {
119        "r_est": 1.0263248682022095,
120        "q_threshold": 0.07391086220741272,
121        "switch_rate": 1.0,
122        "observer_mse": 0.0012712617171928287,
123        "retrieval_mse": 0.007072345819324255
124      },
125      {
126        "r_est": 0.9759218096733093,
127        "q_threshold": 0.12216280400753021,
128        "switch_rate": 1.0,
129        "observer_mse": 0.006652706768363714,
130        "retrieval_mse": 0.01220707781612873
131      },
132      {
133        "r_est": 1.007511854171753,
134        "q_threshold": 0.10744541883468628,
135        "switch_rate": 1.0,
136        "observer_mse": 0.0020110502373427153,
137        "retrieval_mse": 0.0064491755329072475
138      },
139      {
140        "r_est": 0.9763659238815308,
141        "q_threshold": 0.11470293998718262,
142        "switch_rate": 1.0,
143        "observer_mse": 0.0024756789207458496,
144        "retrieval_mse": 0.006343720480799675
145      },
146      {
147        "r_est": 0.9624899625778198,
148        "q_threshold": 0.08983396738767624,
149        "switch_rate": 1.0,
150        "observer_mse": 0.0018914258107542992,
151        "retrieval_mse": 0.006881004199385643
152      },
153      {
154        "r_est": 0.9965905547142029,
155        "q_threshold": 0.0670309066772461,
156        "switch_rate": 1.0,
157        "observer_mse": 0.002109278691932559,
158        "retrieval_mse": 0.0053885444067418575
159      }
160    ],
161    "confirmed": true
162  },
163  "protocol_note": "Built-in dynamics is the required stability/control track. Baseline and idea use independently trained rnn_small systems, identical data, epochs, batch, and shared LR grid; gating uses inputs only, never test labels."
164}