Finite-Horizon Lyapunov Risk Monitor / 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": 0.003,
 10      "weight_decay": 0.0
 11    },
 12    "sweep": [
 13      {
 14        "cfg": {
 15          "lr": 0.001,
 16          "weight_decay": 0.0
 17        },
 18        "mean": 0.004407020111102611
 19      },
 20      {
 21        "cfg": {
 22          "lr": 0.003,
 23          "weight_decay": 0.0
 24        },
 25        "mean": 0.0013896534073865041
 26      }
 27    ],
 28    "full": {
 29      "mean": 0.0015848711991566233,
 30      "std": 0.0004409408404879072,
 31      "per_seed": [
 32        0.0009277964127250016,
 33        0.0014898052904754877,
 34        0.001270615728572011,
 35        0.0018703961977735162,
 36        0.0014201324665918946,
 37        0.001722313230857253,
 38        0.0014589702477678657,
 39        0.002518940018489957
 40      ],
 41      "n": 8
 42    }
 43  },
 44  "idea": {
 45    "mean": 0.0014701014006277546,
 46    "std": 0.0005499829937633325,
 47    "per_seed": [
 48      0.0008705478394404054,
 49      0.0014663755428045988,
 50      0.0011225328780710697,
 51      0.0015238631749525666,
 52      0.0013210882898420095,
 53      0.0013340504374355078,
 54      0.0012874765088781714,
 55      0.0028348765335977077
 56    ],
 57    "n": 8
 58  },
 59  "comparison": {
 60    "delta_mean": -0.0001147697985288687,
 61    "idea_wins": 7,
 62    "n_pairs": 8,
 63    "per_seed_diffs": [
 64      -5.7248573284596205e-05,
 65      -2.34297476708889e-05,
 66      -0.00014808285050094128,
 67      -0.00034653302282094955,
 68      -9.904417674988508e-05,
 69      -0.0003882627934217453,
 70      -0.00017149373888969421,
 71      0.0003159365151077509
 72    ],
 73    "p_value": 0.17025,
 74    "mde": 0.00018161700236817133,
 75    "mde_rel_pct": 11.459417173131632,
 76    "verdict": "no significant win",
 77    "system_worked": false
 78  },
 79  "mechanism_signature": {
 80    "prediction": "Gaussian p+=Phi(mu/sd) for finite-horizon FTLE",
 81    "predicted_positive_fraction": 6.103742223005928e-28,
 82    "observed_positive_fraction": 0.0,
 83    "absolute_error": 6.103742223005928e-28,
 84    "confirmed": true,
 85    "idea_sweep": [
 86      {
 87        "cfg": {
 88          "lr": 0.003,
 89          "weight_decay": 0.0,
 90          "rho": 0.05,
 91          "sigma": 0.04
 92        },
 93        "res": {
 94          "mean": 0.0014701014006277546,
 95          "std": 0.0005499829937633325,
 96          "per_seed": [
 97            0.0008705478394404054,
 98            0.0014663755428045988,
 99            0.0011225328780710697,
100            0.0015238631749525666,
101            0.0013210882898420095,
102            0.0013340504374355078,
103            0.0012874765088781714,
104            0.0028348765335977077
105          ],
106          "n": 8
107        }
108      },
109      {
110        "cfg": {
111          "lr": 0.003,
112          "weight_decay": 0.0,
113          "rho": 0.15,
114          "sigma": 0.04
115        },
116        "res": {
117          "mean": 0.0014701014006277546,
118          "std": 0.0005499829937633325,
119          "per_seed": [
120            0.0008705478394404054,
121            0.0014663755428045988,
122            0.0011225328780710697,
123            0.0015238631749525666,
124            0.0013210882898420095,
125            0.0013340504374355078,
126            0.0012874765088781714,
127            0.0028348765335977077
128          ],
129          "n": 8
130        }
131      },
132      {
133        "cfg": {
134          "lr": 0.003,
135          "weight_decay": 0.0,
136          "rho": 0.3,
137          "sigma": 0.04
138        },
139        "res": {
140          "mean": 0.0014701014006277546,
141          "std": 0.0005499829937633325,
142          "per_seed": [
143            0.0008705478394404054,
144            0.0014663755428045988,
145            0.0011225328780710697,
146            0.0015238631749525666,
147            0.0013210882898420095,
148            0.0013340504374355078,
149            0.0012874765088781714,
150            0.0028348765335977077
151          ],
152          "n": 8
153        }
154      }
155    ]
156  },
157  "official_architecture_note": "Matched GRUCell implementation exposes recurrent state for differentiable FTLE; baseline and idea share all parameters/optimizer/data."
158}