1{
  2  "bench_version": 1,
  3  "track": "dynamics",
  4  "model": "custom_bilinear_vs_additive",
  5  "metric_direction": "lower is better",
  6  "n_seeds": 8,
  7  "baseline": {
  8    "best_cfg": {
  9      "lr": 0.01,
 10      "epochs": 18
 11    },
 12    "sweep": [
 13      {
 14        "cfg": {
 15          "lr": 0.001,
 16          "epochs": 18
 17        },
 18        "mean": 0.019000880420207977
 19      },
 20      {
 21        "cfg": {
 22          "lr": 0.003,
 23          "epochs": 18
 24        },
 25        "mean": 0.006611686199903488
 26      },
 27      {
 28        "cfg": {
 29          "lr": 0.01,
 30          "epochs": 18
 31        },
 32        "mean": 0.0020916118810418993
 33      }
 34    ],
 35    "full": {
 36      "mean": 0.002013212229940109,
 37      "std": 0.0006065338578106211,
 38      "per_seed": [
 39        0.0018298298818990588,
 40        0.002559495158493519,
 41        0.0024440879933536053,
 42        0.0015330344904214144,
 43        0.0014931784244254231,
 44        0.0030939197167754173,
 45        0.002008654410019517,
 46        0.001143497764132917
 47      ],
 48      "n": 8
 49    }
 50  },
 51  "idea": {
 52    "mean": 0.0015174268191913143,
 53    "std": 0.00011583768851973981,
 54    "per_seed": [
 55      0.001512044109404087,
 56      0.0015710212755948305,
 57      0.001313605927862227,
 58      0.00142715685069561,
 59      0.0016863476485013962,
 60      0.0014500042889267206,
 61      0.0016646207077428699,
 62      0.001514613744802773
 63    ],
 64    "n": 8,
 65    "best_cfg": {
 66      "lr": 0.01,
 67      "epochs": 18
 68    },
 69    "grid": [
 70      {
 71        "cfg": {
 72          "lr": 0.001,
 73          "epochs": 18
 74        },
 75        "mean": 0.01489606557879597,
 76        "per_seed": [
 77          0.012612278573215008,
 78          0.022672472521662712,
 79          0.011998260393738747,
 80          0.015981927514076233,
 81          0.014354296028614044,
 82          0.016111958771944046,
 83          0.014130165800452232,
 84          0.011307165026664734
 85        ]
 86      },
 87      {
 88        "cfg": {
 89          "lr": 0.003,
 90          "epochs": 18
 91        },
 92        "mean": 0.005213575612287968,
 93        "per_seed": [
 94          0.0037002854514867067,
 95          0.008283769711852074,
 96          0.0034735731314867735,
 97          0.005930200219154358,
 98          0.004476846195757389,
 99          0.007469747681170702,
100          0.005013920832425356,
101          0.0033602616749703884
102        ]
103      },
104      {
105        "cfg": {
106          "lr": 0.01,
107          "epochs": 18
108        },
109        "mean": 0.0015174268191913143,
110        "per_seed": [
111          0.001512044109404087,
112          0.0015710212755948305,
113          0.001313605927862227,
114          0.00142715685069561,
115          0.0016863476485013962,
116          0.0014500042889267206,
117          0.0016646207077428699,
118          0.001514613744802773
119        ]
120      }
121    ]
122  },
123  "comparison": {
124    "delta_mean": -0.0004957854107487947,
125    "idea_wins": 6,
126    "n_pairs": 8,
127    "per_seed_diffs": [
128      -0.00031778577249497175,
129      -0.0009884738828986883,
130      -0.0011304820654913783,
131      -0.00010587763972580433,
132      0.00019316922407597303,
133      -0.0016439154278486967,
134      -0.0003440337022766471,
135      0.00037111598066985607
136    ],
137    "p_value": 0.0991,
138    "mde": 0.0005824596073258201,
139    "mde_rel_pct": 28.931853217638544,
140    "verdict": "no significant win",
141    "system_worked": false
142  },
143  "architecture": "matched recurrent latent transition; only bilinear interaction differs",
144  "parameter_note": "rank=4, hidden=48",
145  "mechanism_signature": {
146    "observed_action_flip_effect": 0.03430890548042953,
147    "prediction_error": 0.030489611206576228,
148    "ratio": 1.125265430472405,
149    "confirmed": true
150  }
151}