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      "momentum": 0.9
 11    },
 12    "sweep": [
 13      {
 14        "cfg": {
 15          "lr": 0.001,
 16          "momentum": 0.0
 17        },
 18        "mean": 0.5711148381233215
 19      },
 20      {
 21        "cfg": {
 22          "lr": 0.001,
 23          "momentum": 0.9
 24        },
 25        "mean": 0.0024616424925625324
 26      },
 27      {
 28        "cfg": {
 29          "lr": 0.003,
 30          "momentum": 0.0
 31        },
 32        "mean": 0.282429538667202
 33      },
 34      {
 35        "cfg": {
 36          "lr": 0.003,
 37          "momentum": 0.9
 38        },
 39        "mean": 0.0002945917294709943
 40      },
 41      {
 42        "cfg": {
 43          "lr": 0.01,
 44          "momentum": 0.0
 45        },
 46        "mean": 0.003463941771769896
 47      },
 48      {
 49        "cfg": {
 50          "lr": 0.01,
 51          "momentum": 0.9
 52        },
 53        "mean": 0.0002202998639404541
 54      }
 55    ],
 56    "full": {
 57      "mean": 0.00023555499956273707,
 58      "std": 9.620851592885491e-05,
 59      "per_seed": [
 60        0.00011628564243437722,
 61        0.0002339719794690609,
 62        0.00016565514670219272,
 63        0.0003652866871561855,
 64        0.0001685731258476153,
 65        0.00020532555936370045,
 66        0.0004151338362134993,
 67        0.00021420801931526512
 68      ],
 69      "n": 8
 70    }
 71  },
 72  "idea": {
 73    "mean": 0.0002354626767555601,
 74    "std": 9.609999631882683e-05,
 75    "per_seed": [
 76      0.0001164160858024843,
 77      0.00023479813535232097,
 78      0.00016612904437351972,
 79      0.00036248634569346905,
 80      0.0001680333080003038,
 81      0.00020627214689739048,
 82      0.00041672601946629584,
 83      0.0002128403284586966
 84    ],
 85    "n": 8
 86  },
 87  "comparison": {
 88    "delta_mean": -9.232280717696995e-08,
 89    "idea_wins": 3,
 90    "n_pairs": 8,
 91    "per_seed_diffs": [
 92      1.3044336810708046e-07,
 93      8.261558832600713e-07,
 94      4.738976713269949e-07,
 95      -2.8003414627164602e-06,
 96      -5.398178473114967e-07,
 97      9.465875336900353e-07,
 98      1.5921832527965307e-06,
 99      -1.3676908565685153e-06
100    ],
101    "p_value": 0.8746,
102    "mde": 1.1947225589678887e-06,
103    "mde_rel_pct": 0.50719473633999,
104    "verdict": "no measurable effect",
105    "system_worked": false
106  },
107  "mechanism_signature": {
108    "prediction": "EMA observer reduces slowly varying transition residual; measured on trained models",
109    "values": {
110      "raw_transition_rms": 0.0018187333487826674,
111      "observer_rms": 0.00027024194305492306,
112      "applied_direction_rms": 0.0211167361178205,
113      "predicted_residual_reduction": 0.8514120042744011,
114      "confirmed": true
115    },
116    "confirmed": true
117  },
118  "idea_sweep": [
119    {
120      "cfg": {
121        "lr": 0.003,
122        "momentum": 0.9,
123        "alpha": 0.02,
124        "clip": 0.05
125      },
126      "mean": 0.00032222304071183316
127    },
128    {
129      "cfg": {
130        "lr": 0.01,
131        "momentum": 0.9,
132        "alpha": 0.1,
133        "clip": 0.05
134      },
135      "mean": 0.0002354626767555601
136    },
137    {
138      "cfg": {
139        "lr": 0.01,
140        "momentum": 0.9,
141        "alpha": 0.3,
142        "clip": 0.05
143      },
144      "mean": 0.00023565105766465422
145    }
146  ],
147  "matched_structure_justification": "Dynamics is the control/stability track; both systems train identical rnn_small models on identical paired pendulum datasets."
148}