Finite-Excitation Latent Replay / bench_report.json

Failed on benchmark

Raw ⬇ ZIP
  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": 0.0012379240797599778
 19      },
 20      {
 21        "cfg": {
 22          "lr": 0.001,
 23          "weight_decay": 0.0001
 24        },
 25        "mean": 0.001226543157827109
 26      },
 27      {
 28        "cfg": {
 29          "lr": 0.003,
 30          "weight_decay": 0.0
 31        },
 32        "mean": 0.000877063677762635
 33      },
 34      {
 35        "cfg": {
 36          "lr": 0.003,
 37          "weight_decay": 0.0001
 38        },
 39        "mean": 0.0008647404611110687
 40      },
 41      {
 42        "cfg": {
 43          "lr": 0.01,
 44          "weight_decay": 0.0
 45        },
 46        "mean": 0.0003551548215909861
 47      },
 48      {
 49        "cfg": {
 50          "lr": 0.01,
 51          "weight_decay": 0.0001
 52        },
 53        "mean": 0.00048234833229798824
 54      }
 55    ],
 56    "full": {
 57      "mean": 0.0002390247773291776,
 58      "std": 0.00016356887817964183,
 59      "per_seed": [
 60        0.000364280364010483,
 61        0.0005520923878066242,
 62        0.0003466835187282413,
 63        0.00015756301581859589,
 64        0.00026062532560899854,
 65        0.00010608434968162328,
 66        3.411849320400506e-05,
 67        9.075076377484947e-05
 68      ],
 69      "n": 8
 70    }
 71  },
 72  "idea": {
 73    "mean": 0.7325335219502449,
 74    "std": 0.07627789337540042,
 75    "per_seed": [
 76      0.6195922493934631,
 77      0.8263598680496216,
 78      0.7263849377632141,
 79      0.8282962441444397,
 80      0.634881854057312,
 81      0.7560707926750183,
 82      0.681036651134491,
 83      0.7876455783843994
 84    ],
 85    "n": 8
 86  },
 87  "comparison": {
 88    "delta_mean": 0.7322944971729157,
 89    "idea_wins": 0,
 90    "n_pairs": 8,
 91    "per_seed_diffs": [
 92      0.6192279690294527,
 93      0.825807775661815,
 94      0.7260382542444859,
 95      0.8281386811286211,
 96      0.634621228731703,
 97      0.7559647083253367,
 98      0.681002532641287,
 99      0.7875548276206246
100    ],
101    "p_value": 0.0081,
102    "mde": 0.06817817350226507,
103    "mde_rel_pct": 28523.47537526295,
104    "verdict": "idea worse (significant)",
105    "system_worked": false
106  },
107  "mechanism_signature": {
108    "prediction": "parameter updates become active after q>gamma and loss then decreases",
109    "predicted_vs_observed": {
110      "predicted_positive_q_activation": true,
111      "observed_positive_q_rate_mean": 0.0,
112      "observed_activation_rate_mean": 0.0,
113      "observed_post_activation_loss_drop_mean": 0.0,
114      "activation_epoch_values": [
115        null,
116        null,
117        null,
118        null,
119        null,
120        null,
121        null,
122        null
123      ]
124    },
125    "confirmed": false
126  },
127  "idea_sweep": [
128    {
129      "cfg": {
130        "lr": 0.001,
131        "weight_decay": 0.0,
132        "gamma": 0.0,
133        "eps": 0.035,
134        "replay": 4
135      },
136      "mean": 0.7501583248376846
137    },
138    {
139      "cfg": {
140        "lr": 0.003,
141        "weight_decay": 0.0,
142        "gamma": 0.0,
143        "eps": 0.035,
144        "replay": 4
145      },
146      "mean": 0.7501583248376846
147    },
148    {
149      "cfg": {
150        "lr": 0.01,
151        "weight_decay": 0.0,
152        "gamma": 0.0,
153        "eps": 0.035,
154        "replay": 4
155      },
156      "mean": 0.7501583248376846
157    }
158  ],
159  "protocol": {
160    "seeds": [
161      0,
162      1,
163      2,
164      3,
165      4,
166      5,
167      6,
168      7
169    ],
170    "n_train": 900,
171    "n_test": 300,
172    "epochs": 14,
173    "structural_match": "controlled pendulum rollout / latent dynamics",
174    "custom_track": null
175  }
176}