Skew-Midpoint Neural Dynamics / bench_report.json

Mechanism confirmed, baseline not beaten

Raw ⬇ ZIP
  1{
  2  "bench_version": 1,
  3  "track": "dynamics",
  4  "model": "matched_recurrent_transition",
  5  "metric_direction": "lower is better",
  6  "n_seeds": 8,
  7  "baseline": {
  8    "best_cfg": {
  9      "lr": 0.003
 10    },
 11    "sweep": [
 12      {
 13        "cfg": {
 14          "lr": 0.001
 15        },
 16        "mean": 1.1671915558508772e-05
 17      },
 18      {
 19        "cfg": {
 20          "lr": 0.003
 21        },
 22        "mean": 8.121563496388262e-06
 23      },
 24      {
 25        "cfg": {
 26          "lr": 0.01
 27        },
 28        "mean": 9.811891231947811e-06
 29      }
 30    ],
 31    "full": {
 32      "mean": 6.580378681064758e-06,
 33      "std": 3.0327352566591053e-06,
 34      "per_seed": [
 35        4.271172201697482e-06,
 36        8.293503924505785e-06,
 37        7.3761525527515914e-06,
 38        1.254542530659819e-05,
 39        7.987388016772456e-06,
 40        6.073088570701657e-06,
 41        2.071602011710638e-06,
 42        4.024696863780264e-06
 43      ],
 44      "n": 8
 45    }
 46  },
 47  "idea": {
 48    "lr": 0.003,
 49    "mean": 5.4746383426618195e-06,
 50    "std": 5.993504897644281e-06,
 51    "per_seed": [
 52      1.7064038502212497e-06,
 53      2.001432039833162e-05,
 54      2.124709681083914e-06,
 55      2.3162219804362394e-06,
 56      7.144091796362773e-06,
 57      2.4120477064570878e-06,
 58      5.545186922972789e-07,
 59      7.524792636104394e-06
 60    ],
 61    "n": 8,
 62    "sweep": [
 63      {
 64        "lr": 0.001,
 65        "mean": 7.756197987873747e-06,
 66        "std": 4.2305065129479555e-06,
 67        "per_seed": [
 68          9.251196388504468e-06,
 69          1.7147776816273108e-05,
 70          5.101515398564516e-06,
 71          3.05935509459232e-06,
 72          8.358783816220239e-06,
 73          7.70948736317223e-06,
 74          8.440194505965337e-06,
 75          2.981274519697763e-06
 76        ],
 77        "n": 8
 78      },
 79      {
 80        "lr": 0.003,
 81        "mean": 5.4746383426618195e-06,
 82        "std": 5.993504897644281e-06,
 83        "per_seed": [
 84          1.7064038502212497e-06,
 85          2.001432039833162e-05,
 86          2.124709681083914e-06,
 87          2.3162219804362394e-06,
 88          7.144091796362773e-06,
 89          2.4120477064570878e-06,
 90          5.545186922972789e-07,
 91          7.524792636104394e-06
 92        ],
 93        "n": 8
 94      },
 95      {
 96        "lr": 0.01,
 97        "mean": 3.2431818851819116e-05,
 98        "std": 3.8885316263480605e-05,
 99        "per_seed": [
100          4.220135451760143e-06,
101          8.597759915573988e-06,
102          4.01485067413887e-06,
103          0.00010522234515519813,
104          2.6975987566402182e-05,
105          9.17537763598375e-05,
106          1.1658147741400171e-05,
107          7.0115479502419475e-06
108        ],
109        "n": 8
110      }
111    ],
112    "cfg": {
113      "lr": 0.003,
114      "nearby_tested": [
115        0.001,
116        0.003,
117        0.01
118      ]
119    }
120  },
121  "comparison": {
122    "delta_mean": -1.1057403384029385e-06,
123    "idea_wins": 6,
124    "n_pairs": 8,
125    "per_seed_diffs": [
126      -2.564768351476232e-06,
127      1.1720816473825835e-05,
128      -5.2514428716676775e-06,
129      -1.022920332616195e-05,
130      -8.432962204096839e-07,
131      -3.661040864244569e-06,
132      -1.5170833194133593e-06,
133      3.50009577232413e-06
134    ],
135    "p_value": 0.6253,
136    "mde": 5.428610796522651e-06,
137    "mde_rel_pct": 82.49693611316998,
138    "verdict": "no significant win",
139    "system_worked": false
140  },
141  "mechanism_signature": {
142    "prediction": "unforced quadratic energy remains constant under midpoint skew transition",
143    "trained_model_observed_max_energy_drift": 0.0001373291015625,
144    "trained_model_skew_frobenius_error": 0.0,
145    "predicted_energy_drift": 0.0,
146    "confirmed": true,
147    "sanity": {
148      "max_abs_energy_drift": 0.0006403923034667969,
149      "skew_error": 0.0,
150      "predicted": "zero up to floating point"
151    },
152    "note": "idea uses same input/output widths and hidden width; only recurrent transition differs"
153  },
154  "stage2_config": {
155    "epochs": 15,
156    "batch": 128,
157    "lr_union": [
158      0.001,
159      0.003,
160      0.01
161    ],
162    "seeds": [
163      0,
164      1,
165      2,
166      3,
167      4,
168      5,
169      6,
170      7
171    ]
172  }
173}