1{
  2  "bench_version": 1,
  3  "track": "dynamics",
  4  "model": "residual_rnn_shared",
  5  "metric_direction": "lower is better",
  6  "n_seeds": 8,
  7  "baseline": {
  8    "best_cfg": {
  9      "lr": 0.006,
 10      "dt": 0.25,
 11      "m": 0
 12    },
 13    "sweep": [
 14      {
 15        "cfg": {
 16          "lr": 0.0015,
 17          "dt": 0.25,
 18          "m": 0
 19        },
 20        "mean": 0.0012726996355922893
 21      },
 22      {
 23        "cfg": {
 24          "lr": 0.003,
 25          "dt": 0.25,
 26          "m": 0
 27        },
 28        "mean": 0.0005508951580850407
 29      },
 30      {
 31        "cfg": {
 32          "lr": 0.006,
 33          "dt": 0.25,
 34          "m": 0
 35        },
 36        "mean": 0.0002624734515848104
 37      }
 38    ],
 39    "full": {
 40      "mean": 0.00025840229682216886,
 41      "std": 9.502614806341116e-05,
 42      "per_seed": [
 43        0.00021290902805048972,
 44        0.00015823994181118906,
 45        0.00023945324937812984,
 46        0.00043929158709943295,
 47        0.00016651710029691458,
 48        0.0003356700181029737,
 49        0.000337076693540439,
 50        0.00017806075629778206
 51      ],
 52      "n": 8
 53    }
 54  },
 55  "idea": {
 56    "mean": 0.00041295221308246255,
 57    "std": 0.00010711652432851176,
 58    "per_seed": [
 59      0.000396863033529371,
 60      0.00026599428383633494,
 61      0.0004898399929516017,
 62      0.0006024055765010417,
 63      0.00039211593684740365,
 64      0.00029558956157416105,
 65      0.0005119965644553304,
 66      0.00034881275496445596
 67    ],
 68    "n": 8
 69  },
 70  "comparison": {
 71    "delta_mean": 0.00015454991626029368,
 72    "idea_wins": 1,
 73    "n_pairs": 8,
 74    "per_seed_diffs": [
 75      0.0001839540054788813,
 76      0.00010775434202514589,
 77      0.0002503867435734719,
 78      0.0001631139894016087,
 79      0.00022559883655048907,
 80      -4.008045652881265e-05,
 81      0.00017491987091489136,
 82      0.0001707519986666739
 83    ],
 84    "p_value": 0.0163,
 85    "mde": 7.473365296214683e-05,
 86    "mde_rel_pct": 28.92143525085543,
 87    "verdict": "idea worse (significant)",
 88    "system_worked": false
 89  },
 90  "idea_sweep": [
 91    {
 92      "cfg": {
 93        "lr": 0.006,
 94        "dt": 0.25,
 95        "m": 0
 96      },
 97      "result": {
 98        "mean": 0.00041295221308246255,
 99        "std": 0.00010711652432851176,
100        "per_seed": [
101          0.000396863033529371,
102          0.00026599428383633494,
103          0.0004898399929516017,
104          0.0006024055765010417,
105          0.00039211593684740365,
106          0.00029558956157416105,
107          0.0005119965644553304,
108          0.00034881275496445596
109        ],
110        "n": 8
111      }
112    },
113    {
114      "cfg": {
115        "lr": 0.0015,
116        "dt": 0.25,
117        "m": 1
118      },
119      "result": {
120        "mean": 0.011022639810107648,
121        "std": 0.0020031189115314024,
122        "per_seed": [
123          0.011321489699184895,
124          0.008100657723844051,
125          0.01179476361721754,
126          0.01137582492083311,
127          0.011110322549939156,
128          0.010274713858962059,
129          0.015252543613314629,
130          0.008950802497565746
131        ],
132        "n": 8
133      }
134    },
135    {
136      "cfg": {
137        "lr": 0.006,
138        "dt": 0.25,
139        "m": 1
140      },
141      "result": {
142        "mean": 0.0010292112237948459,
143        "std": 0.00026172876255944693,
144        "per_seed": [
145          0.0011666412465274334,
146          0.00048204700578935444,
147          0.001413801102899015,
148          0.0010645093861967325,
149          0.0009550628019496799,
150          0.0008412864408455789,
151          0.0012185709783807397,
152          0.0010917708277702332
153        ],
154        "n": 8
155      }
156    }
157  ],
158  "custom_track": null,
159  "selection_note": "m=1 idea; baseline sweep and idea sweep use identical lr union and epochs",
160  "mechanism_signature": {
161    "quantity": "trained radial derivative vs -a*r^(m+1)",
162    "predicted_mean": -4.693806644385415,
163    "observed_mean": -4.693806647272514,
164    "mean_abs_error": 5.941758198397501e-07,
165    "relative_abs_error": 1.4333428362199376e-07,
166    "correlation": 0.9999999999999342,
167    "n_samples": 22400,
168    "hidden_norm_mean": 2.9794825143207397,
169    "confirmed": true
170  }
171}