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      "weight_decay": 0.0
 11    },
 12    "sweep": [
 13      {
 14        "cfg": {
 15          "lr": 0.001,
 16          "weight_decay": 0.0
 17        },
 18        "mean": 3.534919142111903e-05
 19      },
 20      {
 21        "cfg": {
 22          "lr": 0.001,
 23          "weight_decay": 0.0001
 24        },
 25        "mean": 0.00018523921971791424
 26      },
 27      {
 28        "cfg": {
 29          "lr": 0.003,
 30          "weight_decay": 0.0
 31        },
 32        "mean": 7.330269568228687e-06
 33      },
 34      {
 35        "cfg": {
 36          "lr": 0.003,
 37          "weight_decay": 0.0001
 38        },
 39        "mean": 6.71145257911121e-05
 40      },
 41      {
 42        "cfg": {
 43          "lr": 0.01,
 44          "weight_decay": 0.0
 45        },
 46        "mean": 3.9207230884130695e-06
 47      },
 48      {
 49        "cfg": {
 50          "lr": 0.01,
 51          "weight_decay": 0.0001
 52        },
 53        "mean": 2.6831472723642946e-05
 54      }
 55    ],
 56    "full": {
 57      "mean": 3.3532741667841037e-06,
 58      "std": 1.1020743584708325e-06,
 59      "per_seed": [
 60        4.035734946228331e-06,
 61        3.1895601750875358e-06,
 62        2.5228919184883125e-06,
 63        5.934705313848099e-06,
 64        3.3925441584869986e-06,
 65        2.830740641002194e-06,
 66        2.4708413093321724e-06,
 67        2.449174871799187e-06
 68      ],
 69      "n": 8
 70    }
 71  },
 72  "idea": {
 73    "mean": 0.5757362097501755,
 74    "std": 0.020123708021359012,
 75    "per_seed": [
 76      0.5845423936843872,
 77      0.5802162289619446,
 78      0.5462422966957092,
 79      0.6178359985351562,
 80      0.5681619644165039,
 81      0.5554149746894836,
 82      0.5804445147514343,
 83      0.5730313062667847
 84    ],
 85    "n": 8
 86  },
 87  "comparison": {
 88    "delta_mean": 0.5757328564760087,
 89    "idea_wins": 0,
 90    "n_pairs": 8,
 91    "per_seed_diffs": [
 92      0.584538357949441,
 93      0.5802130394017695,
 94      0.5462397738037907,
 95      0.6178300638298424,
 96      0.5681585718723454,
 97      0.5554121439488426,
 98      0.580442043910125,
 99      0.5730288570919129
100    ],
101    "p_value": 0.0081,
102    "mde": 0.017987483920840156,
103    "mde_rel_pct": 536415.5457079945,
104    "verdict": "idea worse (significant)",
105    "system_worked": false
106  },
107  "idea_grid": [
108    {
109      "cfg": {
110        "lr": 0.01,
111        "weight_decay": 0.0
112      },
113      "mean": 0.5757362097501755
114    },
115    {
116      "cfg": {
117        "lr": 0.003,
118        "weight_decay": 0.0
119      },
120      "mean": 0.575765073299408
121    },
122    {
123      "cfg": {
124        "lr": 0.001,
125        "weight_decay": 0.0
126      },
127      "mean": 0.5757729709148407
128    }
129  ],
130  "math_sanity_check": {
131    "mu": 0.3,
132    "Lf": 0.2,
133    "theoretical_bound": -0.09999999999999998,
134    "max_residual_fd_lipschitz": 0.1999997138561903,
135    "max_one_sided_ratio": -0.09999965212070401,
136    "violation_fraction": 5e-06,
137    "confirmed": false
138  },
139  "custom_track": null,
140  "mechanism_signature": {
141    "mu": 0.3,
142    "Lf": 0.2,
143    "idea_observed_residual_fd_lipschitz_max": 0.1959163397550583,
144    "idea_observed_one_sided_ratio_max": -0.10408368706703186,
145    "baseline_observed_one_sided_ratio_max": 0.009726762771606445,
146    "predicted_bound": -0.09999999999999998,
147    "confirmed": true
148  },
149  "idea_best_cfg": {
150    "lr": 0.01,
151    "weight_decay": 0.0
152  },
153  "protocol_note": "Dynamics is structurally matched: controlled pendulum windows contain (theta, omega, action), and the intervention acts on the action coordinate while retaining the shared GRU core."
154}