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": 0.006,
 10      "weight_decay": 0.0001,
 11      "epochs": 12,
 12      "attr_weight": 0.0,
 13      "target_lambda": -0.08
 14    },
 15    "sweep": [
 16      {
 17        "cfg": {
 18          "lr": 0.0015,
 19          "weight_decay": 0.0,
 20          "epochs": 12,
 21          "attr_weight": 0.0,
 22          "target_lambda": -0.08
 23        },
 24        "mean": 0.004987949621863663
 25      },
 26      {
 27        "cfg": {
 28          "lr": 0.0015,
 29          "weight_decay": 0.0001,
 30          "epochs": 12,
 31          "attr_weight": 0.0,
 32          "target_lambda": -0.08
 33        },
 34        "mean": 0.00498378137126565
 35      },
 36      {
 37        "cfg": {
 38          "lr": 0.003,
 39          "weight_decay": 0.0,
 40          "epochs": 12,
 41          "attr_weight": 0.0,
 42          "target_lambda": -0.08
 43        },
 44        "mean": 0.00358188062091358
 45      },
 46      {
 47        "cfg": {
 48          "lr": 0.003,
 49          "weight_decay": 0.0001,
 50          "epochs": 12,
 51          "attr_weight": 0.0,
 52          "target_lambda": -0.08
 53        },
 54        "mean": 0.0035488905268721282
 55      },
 56      {
 57        "cfg": {
 58          "lr": 0.006,
 59          "weight_decay": 0.0,
 60          "epochs": 12,
 61          "attr_weight": 0.0,
 62          "target_lambda": -0.08
 63        },
 64        "mean": 0.0019017510931007564
 65      },
 66      {
 67        "cfg": {
 68          "lr": 0.006,
 69          "weight_decay": 0.0001,
 70          "epochs": 12,
 71          "attr_weight": 0.0,
 72          "target_lambda": -0.08
 73        },
 74        "mean": 0.001846620871219784
 75      }
 76    ],
 77    "full": {
 78      "mean": 0.0020994337101001292,
 79      "std": 0.0003699326898898649,
 80      "per_seed": [
 81        0.0021909112110733986,
 82        0.0014349172124639153,
 83        0.0016000064788386226,
 84        0.0021606485825031996,
 85        0.0026131158228963614,
 86        0.002182674128562212,
 87        0.002174227498471737,
 88        0.0024389687459915876
 89      ],
 90      "n": 8
 91    }
 92  },
 93  "idea": {
 94    "mean": 0.0019659589452203363,
 95    "std": 0.0007528780037898819,
 96    "per_seed": [
 97      0.002175788627937436,
 98      0.0017074780771508813,
 99      0.0010215449146926403,
100      0.001942452392540872,
101      0.002001833636313677,
102      0.0019844763446599245,
103      0.0012087502982467413,
104      0.003685347270220518
105    ],
106    "n": 8
107  },
108  "comparison": {
109    "delta_mean": -0.00013347476487979293,
110    "idea_wins": 6,
111    "n_pairs": 8,
112    "per_seed_diffs": [
113      -1.5122583135962486e-05,
114      0.00027256086468696594,
115      -0.0005784615641459823,
116      -0.00021819618996232748,
117      -0.0006112821865826845,
118      -0.00019819778390228748,
119      -0.0009654772002249956,
120      0.0012463785242289305
121    ],
122    "p_value": 0.59995,
123    "mde": 0.0005664705121806638,
124    "mde_rel_pct": 26.982062327352402,
125    "verdict": "no significant win",
126    "system_worked": false
127  },
128  "mechanism_signature": {
129    "prediction": "trained hidden pullback diameter contracts exponentially under shared replay",
130    "config": {
131      "lr": 0.006,
132      "weight_decay": 0.0001,
133      "epochs": 12,
134      "attr_weight": 0.001,
135      "target_lambda": -0.08
136    },
137    "predicted_log_diameter_step": -0.08,
138    "observed_log_diameter_slope": -0.4773993907385317,
139    "final_diameter": 0.034216802567243576,
140    "confirmed": false
141  },
142  "idea_sweep": [
143    {
144      "cfg": {
145        "lr": 0.006,
146        "weight_decay": 0.0001,
147        "epochs": 12,
148        "attr_weight": 0.001,
149        "target_lambda": -0.08
150      },
151      "result": {
152        "mean": 0.0019659589452203363,
153        "std": 0.0007528780037898819,
154        "per_seed": [
155          0.002175788627937436,
156          0.0017074780771508813,
157          0.0010215449146926403,
158          0.001942452392540872,
159          0.002001833636313677,
160          0.0019844763446599245,
161          0.0012087502982467413,
162          0.003685347270220518
163        ],
164        "n": 8
165      }
166    },
167    {
168      "cfg": {
169        "lr": 0.006,
170        "weight_decay": 0.0001,
171        "epochs": 12,
172        "attr_weight": 0.01,
173        "target_lambda": -0.08
174      },
175      "result": {
176        "mean": 0.0019659589452203363,
177        "std": 0.0007528780037898819,
178        "per_seed": [
179          0.002175788627937436,
180          0.0017074780771508813,
181          0.0010215449146926403,
182          0.001942452392540872,
183          0.002001833636313677,
184          0.0019844763446599245,
185          0.0012087502982467413,
186          0.003685347270220518
187        ],
188        "n": 8
189      }
190    },
191    {
192      "cfg": {
193        "lr": 0.006,
194        "weight_decay": 0.0001,
195        "epochs": 12,
196        "attr_weight": 0.03,
197        "target_lambda": -0.04
198      },
199      "result": {
200        "mean": 0.0019659589452203363,
201        "std": 0.0007528780037898819,
202        "per_seed": [
203          0.002175788627937436,
204          0.0017074780771508813,
205          0.0010215449146926403,
206          0.001942452392540872,
207          0.002001833636313677,
208          0.0019844763446599245,
209          0.0012087502982467413,
210          0.003685347270220518
211        ],
212        "n": 8
213      }
214    }
215  ],
216  "track_justification": "Dynamics is structurally matched: the idea diagnoses stability and contraction of recurrent latent trajectories."
217}