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    },
 11    "sweep": [
 12      {
 13        "cfg": {
 14          "lr": 0.001
 15        },
 16        "mean": 0.5088173821568489
 17      },
 18      {
 19        "cfg": {
 20          "lr": 0.003
 21        },
 22        "mean": 0.06073589529842138
 23      },
 24      {
 25        "cfg": {
 26          "lr": 0.006
 27        },
 28        "mean": 0.023900731466710567
 29      }
 30    ],
 31    "full": {
 32      "mean": 0.019088300527073443,
 33      "std": 0.00558386015869873,
 34      "per_seed": [
 35        0.018445175141096115,
 36        0.023440849035978317,
 37        0.02936750277876854,
 38        0.024349398910999298,
 39        0.01488791685551405,
 40        0.012780136428773403,
 41        0.01545807532966137,
 42        0.013977349735796452
 43      ],
 44      "n": 8
 45    }
 46  },
 47  "idea": {
 48    "best_cfg": {
 49      "lr": 0.006
 50    },
 51    "sweep": [
 52      {
 53        "cfg": {
 54          "lr": 0.001
 55        },
 56        "mean": 0.3442936111241579,
 57        "std": 0.11055745967503162,
 58        "per_seed": [
 59          0.4150988757610321,
 60          0.572791337966919,
 61          0.3577822744846344,
 62          0.3750651478767395,
 63          0.2974511682987213,
 64          0.17869944870471954,
 65          0.254057914018631,
 66          0.30340272188186646
 67        ],
 68        "n": 8
 69      },
 70      {
 71        "cfg": {
 72          "lr": 0.003
 73        },
 74        "mean": 0.03436855808831751,
 75        "std": 0.01405565796915205,
 76        "per_seed": [
 77          0.05125405639410019,
 78          0.041576966643333435,
 79          0.04229721054434776,
 80          0.055242300033569336,
 81          0.021142112091183662,
 82          0.016799848526716232,
 83          0.020402031019330025,
 84          0.026233939453959465
 85        ],
 86        "n": 8
 87      },
 88      {
 89        "cfg": {
 90          "lr": 0.006
 91        },
 92        "mean": 0.015928541193716228,
 93        "std": 0.004679734568777004,
 94        "per_seed": [
 95          0.0178981926292181,
 96          0.013649515807628632,
 97          0.024342793971300125,
 98          0.021126413717865944,
 99          0.012977873906493187,
100          0.008715411648154259,
101          0.01320882048457861,
102          0.015509307384490967
103        ],
104        "n": 8
105      }
106    ],
107    "mean": 0.015928541193716228,
108    "std": 0.004679734568777004,
109    "per_seed": [
110      0.0178981926292181,
111      0.013649515807628632,
112      0.024342793971300125,
113      0.021126413717865944,
114      0.012977873906493187,
115      0.008715411648154259,
116      0.01320882048457861,
117      0.015509307384490967
118    ],
119    "n": 8
120  },
121  "comparison": {
122    "delta_mean": -0.003159759333357215,
123    "idea_wins": 7,
124    "n_pairs": 8,
125    "per_seed_diffs": [
126      -0.0005469825118780136,
127      -0.009791333228349686,
128      -0.005024708807468414,
129      -0.003222985193133354,
130      -0.0019100429490208626,
131      -0.0040647247806191444,
132      -0.00224925484508276,
133      0.0015319576486945152
134    ],
135    "p_value": 0.02315,
136    "mde": 0.002818017003243901,
137    "mde_rel_pct": 14.763058655992097,
138    "verdict": "idea better (significant)",
139    "system_worked": true
140  },
141  "mechanism_signature": {
142    "collective_detectability": {
143      "window": 2,
144      "predicted_lambda_min_positive": true,
145      "observed_mean_fused_information": 1.459951639175415,
146      "observed_positive_fraction": 1.0,
147      "tolerance": "positivity and >0.99 positive fraction",
148      "confirmed": true
149    }
150  },
151  "protocol_notes": {
152    "structural_match": "dynamics: recurrent pendulum rollout and latent-state stability",
153    "paired_seeds": 8,
154    "epochs": 12,
155    "batch": 128,
156    "shared_lr_union": [
157      0.001,
158      0.003,
159      0.006
160    ],
161    "baseline_method_knob": "equal arithmetic mean (fixed by definition)",
162    "system_parity": "separately trained identical two-agent GRU encoders and predictor; only fusion differs"
163  }
164}