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      "radius": "constant"
 11    },
 12    "sweep": [
 13      {
 14        "cfg": {
 15          "lr": 0.001,
 16          "radius": "constant"
 17        },
 18        "mean": 0.14777550008147955
 19      },
 20      {
 21        "cfg": {
 22          "lr": 0.003,
 23          "radius": "constant"
 24        },
 25        "mean": 0.02788990270346403
 26      },
 27      {
 28        "cfg": {
 29          "lr": 0.01,
 30          "radius": "constant"
 31        },
 32        "mean": 0.010466416599228978
 33      }
 34    ],
 35    "full": {
 36      "mean": 0.009070940839592367,
 37      "std": 0.0043108428117023265,
 38      "per_seed": [
 39        0.01107354648411274,
 40        0.007685416843742132,
 41        0.006489071529358625,
 42        0.016617631539702415,
 43        0.014132681302726269,
 44        0.008140470832586288,
 45        0.002562329638749361,
 46        0.005866378545761108
 47      ],
 48      "n": 8
 49    }
 50  },
 51  "idea": {
 52    "mean": 0.009070940839592367,
 53    "std": 0.0043108428117023265,
 54    "per_seed": [
 55      0.01107354648411274,
 56      0.007685416843742132,
 57      0.006489071529358625,
 58      0.016617631539702415,
 59      0.014132681302726269,
 60      0.008140470832586288,
 61      0.002562329638749361,
 62      0.005866378545761108
 63    ],
 64    "n": 8,
 65    "selected_per_seed": [
 66      {
 67        "seed": 0,
 68        "best_lr": 0.01,
 69        "metric": 0.01107354648411274
 70      },
 71      {
 72        "seed": 1,
 73        "best_lr": 0.01,
 74        "metric": 0.007685416843742132
 75      },
 76      {
 77        "seed": 2,
 78        "best_lr": 0.01,
 79        "metric": 0.006489071529358625
 80      },
 81      {
 82        "seed": 3,
 83        "best_lr": 0.01,
 84        "metric": 0.016617631539702415
 85      },
 86      {
 87        "seed": 4,
 88        "best_lr": 0.01,
 89        "metric": 0.014132681302726269
 90      },
 91      {
 92        "seed": 5,
 93        "best_lr": 0.01,
 94        "metric": 0.008140470832586288
 95      },
 96      {
 97        "seed": 6,
 98        "best_lr": 0.01,
 99        "metric": 0.002562329638749361
100      },
101      {
102        "seed": 7,
103        "best_lr": 0.01,
104        "metric": 0.005866378545761108
105      }
106    ]
107  },
108  "comparison": {
109    "delta_mean": 0.0,
110    "idea_wins": 0,
111    "n_pairs": 8,
112    "per_seed_diffs": [
113      0.0,
114      0.0,
115      0.0,
116      0.0,
117      0.0,
118      0.0,
119      0.0,
120      0.0
121    ],
122    "p_value": 1.0,
123    "mde": 0.0,
124    "mde_rel_pct": 0.0,
125    "verdict": "no measurable effect",
126    "system_worked": false
127  },
128  "mechanism_signature": {
129    "prediction": "dense tube radius is affine in nearest-observation distance with slope Gamma",
130    "radius_slope_predicted_mean": 0.23295989949256185,
131    "radius_slope_observed_mean": 0.2329598994925619,
132    "radius_slope_abs_error_mean": 8.326672684688674e-17,
133    "tube_coverage_mean": 0.9995028409090909,
134    "constant_radius_coverage_mean": 0.9989109848484848,
135    "median_radius_mean": 0.6557790094055236,
136    "confirmed": true,
137    "trained_model_measurements": true,
138    "note": "The fixed dynamics bench predicts a single endpoint, so dense coverage is a diagnostic interpolation signature; endpoint MSE is the independent primary metric."
139  },
140  "protocol_notes": {
141    "baseline_grid": [
142      {
143        "lr": 0.001,
144        "radius": "constant"
145      },
146      {
147        "lr": 0.003,
148        "radius": "constant"
149      },
150      {
151        "lr": 0.01,
152        "radius": "constant"
153      }
154    ],
155    "idea_grid": [
156      {
157        "lr": 0.001,
158        "tube": "q+Gamma*delta"
159      },
160      {
161        "lr": 0.003,
162        "tube": "q+Gamma*delta"
163      },
164      {
165        "lr": 0.01,
166        "tube": "q+Gamma*delta"
167      }
168    ],
169    "paired_seeds": [
170      0,
171      1,
172      2,
173      3,
174      4,
175      5,
176      6,
177      7
178    ],
179    "epochs": 6,
180    "n_train": 500,
181    "n_test": 160,
182    "structural_match": "continuous-time trajectory uncertainty and stability -> dynamics"
183  }
184}