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      "weight_decay": 0.0001
 11    },
 12    "sweep": [
 13      {
 14        "cfg": {
 15          "lr": 0.0015,
 16          "weight_decay": 0.0
 17        },
 18        "mean": 0.005481100233737379
 19      },
 20      {
 21        "cfg": {
 22          "lr": 0.0015,
 23          "weight_decay": 0.0001
 24        },
 25        "mean": 0.005474042263813317
 26      },
 27      {
 28        "cfg": {
 29          "lr": 0.003,
 30          "weight_decay": 0.0
 31        },
 32        "mean": 0.003975238680141047
 33      },
 34      {
 35        "cfg": {
 36          "lr": 0.003,
 37          "weight_decay": 0.0001
 38        },
 39        "mean": 0.00394076865632087
 40      },
 41      {
 42        "cfg": {
 43          "lr": 0.006,
 44          "weight_decay": 0.0
 45        },
 46        "mean": 0.002146739774616435
 47      },
 48      {
 49        "cfg": {
 50          "lr": 0.006,
 51          "weight_decay": 0.0001
 52        },
 53        "mean": 0.002088654466206208
 54      }
 55    ],
 56    "full": {
 57      "mean": 0.0022117759071988985,
 58      "std": 0.0004126881098400608,
 59      "per_seed": [
 60        0.002800536109134555,
 61        0.001347046927548945,
 62        0.001825663261115551,
 63        0.0023813715670257807,
 64        0.0023169058840721846,
 65        0.0024658460170030594,
 66        0.0022247303277254105,
 67        0.002332107163965702
 68      ],
 69      "n": 8
 70    }
 71  },
 72  "idea": {
 73    "mean": 0.002208090227213688,
 74    "std": 0.00051193081082026,
 75    "per_seed": [
 76      0.002686932682991028,
 77      0.001070088124834001,
 78      0.002036568708717823,
 79      0.002309913281351328,
 80      0.0025295668747276068,
 81      0.0025805803015828133,
 82      0.0026061630342155695,
 83      0.0018449088092893362
 84    ],
 85    "n": 8,
 86    "best_cfg": {
 87      "lr": 0.006,
 88      "weight_decay": 0.0001,
 89      "alpha": 0.03
 90    },
 91    "trials": [
 92      {
 93        "cfg": {
 94          "lr": 0.0015,
 95          "weight_decay": 0.0001,
 96          "alpha": 0.03
 97        },
 98        "mean": 0.005427066018455662
 99      },
100      {
101        "cfg": {
102          "lr": 0.003,
103          "weight_decay": 0.0001,
104          "alpha": 0.03
105        },
106        "mean": 0.003149385840515606
107      },
108      {
109        "cfg": {
110          "lr": 0.006,
111          "weight_decay": 0.0001,
112          "alpha": 0.03
113        },
114        "mean": 0.002208090227213688
115      }
116    ]
117  },
118  "comparison": {
119    "delta_mean": -3.6856799852102995e-06,
120    "idea_wins": 4,
121    "n_pairs": 8,
122    "per_seed_diffs": [
123      -0.00011360342614352703,
124      -0.0002769588027149439,
125      0.00021090544760227203,
126      -7.145828567445278e-05,
127      0.0002126609906554222,
128      0.00011473428457975388,
129      0.00038143270649015903,
130      -0.00048719835467636585
131    ],
132    "p_value": 0.97695,
133    "mde": 0.00024075513247358545,
134    "mde_rel_pct": 10.88515033055449,
135    "verdict": "no measurable effect",
136    "system_worked": false
137  },
138  "paired_deltas": [
139    -0.00011360342614352703,
140    -0.0002769588027149439,
141    0.00021090544760227203,
142    -7.145828567445278e-05,
143    0.0002126609906554222,
144    0.00011473428457975388,
145    0.00038143270649015903,
146    -0.00048719835467636585
147  ],
148  "permutation_p": 0.97695,
149  "mechanism_signature": {
150    "q_observed": 0.010351310360661614,
151    "q_predicted": 0.010351310360661614,
152    "drift_energy_ratio": 1.0,
153    "orthogonality_residual": 0.0,
154    "prediction": "cumulative quadratic dissipation is capped at QMAX and skew drift is orthogonal to the instantaneous gradient",
155    "qmax": 0.5,
156    "confirmed": true
157  }
158}