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      "delay": 3,
 11      "epsilon": 0.5
 12    },
 13    "sweep": [
 14      {
 15        "cfg": {
 16          "lr": 0.001,
 17          "delay": 0,
 18          "epsilon": 0.05
 19        },
 20        "mean": 0.6047389209270477
 21      },
 22      {
 23        "cfg": {
 24          "lr": 0.003,
 25          "delay": 1,
 26          "epsilon": 0.18
 27        },
 28        "mean": 0.37101951241493225
 29      },
 30      {
 31        "cfg": {
 32          "lr": 0.01,
 33          "delay": 3,
 34          "epsilon": 0.5
 35        },
 36        "mean": 0.014778749784454703
 37      }
 38    ],
 39    "full": {
 40      "mean": 0.012637637846637517,
 41      "std": 0.008144052987350348,
 42      "per_seed": [
 43        0.004888118244707584,
 44        0.009219375438988209,
 45        0.01350479293614626,
 46        0.03150271251797676,
 47        0.006999698933213949,
 48        0.008468696847558022,
 49        0.008145930245518684,
 50        0.01837177760899067
 51      ],
 52      "n": 8
 53    }
 54  },
 55  "idea": {
 56    "mean": 0.17860826198011637,
 57    "std": 0.049080342215062335,
 58    "per_seed": [
 59      0.12278071790933609,
 60      0.16866335272789001,
 61      0.180322527885437,
 62      0.2821338176727295,
 63      0.13234758377075195,
 64      0.1678561568260193,
 65      0.14913593232631683,
 66      0.22562600672245026
 67    ],
 68    "n": 8
 69  },
 70  "comparison": {
 71    "delta_mean": 0.16597062413347885,
 72    "idea_wins": 0,
 73    "n_pairs": 8,
 74    "per_seed_diffs": [
 75      0.1178925996646285,
 76      0.1594439772889018,
 77      0.16681773494929075,
 78      0.25063110515475273,
 79      0.125347884837538,
 80      0.15938745997846127,
 81      0.14099000208079815,
 82      0.2072542291134596
 83    ],
 84    "p_value": 0.0081,
 85    "mde": 0.0367800689889995,
 86    "mde_rel_pct": 291.0359470285464,
 87    "verdict": "idea worse (significant)",
 88    "system_worked": false
 89  },
 90  "mechanism_signature": {
 91    "prediction": "triggered updates reduce transmissions while retaining a positive inter-event gap under delayed actuation",
 92    "trained_model_signature": {
 93      "trained_event_count": 1,
 94      "total_steps": 48,
 95      "communication_fraction": 0.020833333333333332,
 96      "min_inter_event_gap": null,
 97      "predicted_gap_proxy": "positive for nonzero threshold",
 98      "observed_positive_gap": true,
 99      "mean_drift_norm": 0.24016375679639168,
100      "confirmed": true
101    },
102    "idea_sweep": [
103      {
104        "cfg": {
105          "lr": 0.001,
106          "delay": 0,
107          "epsilon": 0.05
108        },
109        "mean": 0.6575141847133636
110      },
111      {
112        "cfg": {
113          "lr": 0.003,
114          "delay": 1,
115          "epsilon": 0.18
116        },
117        "mean": 0.526682935655117
118      },
119      {
120        "cfg": {
121          "lr": 0.01,
122          "delay": 3,
123          "epsilon": 0.5
124        },
125        "mean": 0.17860826198011637
126      }
127    ],
128    "parameter_counts": {
129      "baseline": 13313,
130      "idea": 13313
131    },
132    "protocol_note": "baseline and idea use identical rnn_small architecture, SGD lr/delay union, epochs, batches, and paired seeds"
133  },
134  "protocol": {
135    "seeds": [
136      0,
137      1,
138      2,
139      3,
140      4,
141      5,
142      6,
143      7
144    ],
145    "epochs": 12,
146    "n_train": 400,
147    "n_test": 200,
148    "batch": 128,
149    "grid": [
150      {
151        "lr": 0.001,
152        "delay": 0,
153        "epsilon": 0.05
154      },
155      {
156        "lr": 0.003,
157        "delay": 1,
158        "epsilon": 0.18
159      },
160      {
161        "lr": 0.01,
162        "delay": 3,
163        "epsilon": 0.5
164      }
165    ]
166  }
167}