Contractive Floquet return map / bench_report.json

Mechanism confirmed, baseline not beaten

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": 0.006,
 10      "weight_decay": 0.0
 11    },
 12    "sweep": [
 13      {
 14        "cfg": {
 15          "lr": 0.0015,
 16          "weight_decay": 0.0
 17        },
 18        "mean": 0.001082472168491222
 19      },
 20      {
 21        "cfg": {
 22          "lr": 0.003,
 23          "weight_decay": 0.0
 24        },
 25        "mean": 0.0007095315813785419
 26      },
 27      {
 28        "cfg": {
 29          "lr": 0.006,
 30          "weight_decay": 0.0
 31        },
 32        "mean": 0.00044968911970499903
 33      }
 34    ],
 35    "full": {
 36      "mean": 0.0003913900891348021,
 37      "std": 0.00014032323399353638,
 38      "per_seed": [
 39        0.0004547900171019137,
 40        0.0005691215046681464,
 41        0.0002539993729442358,
 42        0.0005208455841057003,
 43        0.00024503248278051615,
 44        0.0003696440835483372,
 45        0.00018239273049402982,
 46        0.0005352949374355376
 47      ],
 48      "n": 8
 49    }
 50  },
 51  "idea": {
 52    "mean": 0.00038262269663391635,
 53    "std": 0.00014245699985162617,
 54    "per_seed": [
 55      0.0004878886684309691,
 56      0.0005433244514279068,
 57      0.0002431763568893075,
 58      0.0004911153810098767,
 59      0.00024032114015426487,
 60      0.0003892902168445289,
 61      0.00014719406317453831,
 62      0.0005186712951399386
 63    ],
 64    "n": 8,
 65    "best_cfg": {
 66      "lr": 0.006,
 67      "weight_decay": 0.0
 68    },
 69    "sweep": [
 70      {
 71        "cfg": {
 72          "lr": 0.0015,
 73          "weight_decay": 0.0
 74        },
 75        "mean": 0.0011307366366963834,
 76        "per_seed": [
 77          0.0008429070585407317,
 78          0.0012130566174164414,
 79          0.0007987393764778972,
 80          0.0014753397554159164,
 81          0.0007146105053834617,
 82          0.0010028763208538294,
 83          0.0012895666295662522,
 84          0.0017087968299165368
 85        ]
 86      },
 87      {
 88        "cfg": {
 89          "lr": 0.003,
 90          "weight_decay": 0.0
 91        },
 92        "mean": 0.0006998253520578146,
 93        "per_seed": [
 94          0.0005686159711331129,
 95          0.0008144777966663241,
 96          0.000506731157656759,
 97          0.0009455950348637998,
 98          0.0005022214027121663,
 99          0.0006177405593916774,
100          0.0007299749995581806,
101          0.0009132458944804966
102        ]
103      },
104      {
105        "cfg": {
106          "lr": 0.006,
107          "weight_decay": 0.0
108        },
109        "mean": 0.00038262269663391635,
110        "per_seed": [
111          0.0004878886684309691,
112          0.0005433244514279068,
113          0.0002431763568893075,
114          0.0004911153810098767,
115          0.00024032114015426487,
116          0.0003892902168445289,
117          0.00014719406317453831,
118          0.0005186712951399386
119        ]
120      }
121    ]
122  },
123  "comparison": {
124    "delta_mean": -8.767392500885762e-06,
125    "idea_wins": 6,
126    "n_pairs": 8,
127    "per_seed_diffs": [
128      3.309865132905543e-05,
129      -2.579705324023962e-05,
130      -1.0823016054928303e-05,
131      -2.9730203095823526e-05,
132      -4.71134262625128e-06,
133      1.9646133296191692e-05,
134      -3.5198667319491506e-05,
135      -1.6623642295598984e-05
136    ],
137    "p_value": 0.328,
138    "mde": 2.0163403192736034e-05,
139    "mde_rel_pct": 5.151740872465316,
140    "verdict": "no significant win",
141    "system_worked": false
142  },
143  "mechanism_signature": {
144    "math_sanity": {
145      "q": 0.8,
146      "e0": 0.37,
147      "delta": 0.013,
148      "max_bound_violation": 0.0,
149      "terminal_observed": 0.06537757171198205,
150      "terminal_bound": 0.06537757171198205,
151      "passed": true
152    },
153    "quantity": "trained one-step output gain under nearby input perturbation",
154    "predicted_q_upper_bound": 0.9,
155    "observed_baseline_mean_gain": 0.07031634113839227,
156    "observed_idea_mean_gain": 0.07054478773045154,
157    "observed_idea_max_gain": 0.07702433657582401,
158    "confirmed": true,
159    "n": 8,
160    "track_rationale": "Actuated pendulum rollout is the built-in stability/control task."
161  },
162  "math_sanity": {
163    "q": 0.8,
164    "e0": 0.37,
165    "delta": 0.013,
166    "max_bound_violation": 0.0,
167    "terminal_observed": 0.06537757171198205,
168    "terminal_bound": 0.06537757171198205,
169    "passed": true
170  }
171}