Limiter-Smoothing Bifurcation Guard / 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": 0.003,
 10      "radius": 0.05
 11    },
 12    "sweep": [
 13      {
 14        "cfg": {
 15          "lr": 0.0015,
 16          "radius": 0.05
 17        },
 18        "mean": 0.002696412382647395
 19      },
 20      {
 21        "cfg": {
 22          "lr": 0.003,
 23          "radius": 0.05
 24        },
 25        "mean": 0.0010492464643903077
 26      },
 27      {
 28        "cfg": {
 29          "lr": 0.006,
 30          "radius": 0.05
 31        },
 32        "mean": 0.0012117274600313976
 33      }
 34    ],
 35    "full": {
 36      "mean": 0.0012495772170950659,
 37      "std": 0.0005645325118199199,
 38      "per_seed": [
 39        0.000996186281554401,
 40        0.0008429015870206058,
 41        0.0008552631479687989,
 42        0.001502634841017425,
 43        0.001222004764713347,
 44        0.0009713814943097532,
 45        0.002642611740157008,
 46        0.0009636338800191879
 47      ],
 48      "n": 8
 49    }
 50  },
 51  "idea": {
 52    "mean": 0.0010236013476969674,
 53    "std": 0.0003525166306615029,
 54    "per_seed": [
 55      0.0007349217194132507,
 56      0.0010386385256424546,
 57      0.0008291155099868774,
 58      0.001807487802579999,
 59      0.0007241228595376015,
 60      0.000720260024536401,
 61      0.0013001237530261278,
 62      0.0010341405868530273
 63    ],
 64    "n": 8
 65  },
 66  "comparison": {
 67    "delta_mean": -0.00022597586939809844,
 68    "idea_wins": 5,
 69    "n_pairs": 8,
 70    "per_seed_diffs": [
 71      -0.00026126456214115024,
 72      0.00019573693862184882,
 73      -2.6147637981921434e-05,
 74      0.0003048529615625739,
 75      -0.0004978819051757455,
 76      -0.00025112146977335215,
 77      -0.0013424879871308804,
 78      7.050670683383942e-05
 79    ],
 80    "p_value": 0.2949,
 81    "mde": 0.00043675546709058414,
 82    "mde_rel_pct": 34.952259141369765,
 83    "verdict": "no significant win",
 84    "system_worked": false
 85  },
 86  "mechanism_signature": {
 87    "prediction": "smooth clipping can apply a different local gain and create larger gradients than exact projection near the boundary",
 88    "observed": {
 89      "cfg": {
 90        "lr": 0.006,
 91        "radius": 0.05
 92      },
 93      "baseline_seed0_metric": 0.0007390209357254207,
 94      "idea_seed0_metric": 0.0007349217194132507,
 95      "baseline": {
 96        "grad_norm_mean": 0.4881857982060562,
 97        "grad_norm_p95": 1.74818832874298,
 98        "applied_norm_mean": 0.050000001288329564,
 99        "applied_norm_p95": 0.05000000447034836,
100        "fraction_over_radius": 1.0
101      },
102      "idea": {
103        "grad_norm_mean": 0.48380100160526734,
104        "grad_norm_p95": 1.7462666749954214,
105        "applied_norm_mean": 0.046499995747581124,
106        "applied_norm_p95": 0.0499768827110529,
107        "fraction_over_radius": 0.9791666666666666
108      },
109      "smooth_applied_norm_exceeds_radius": false
110    },
111    "confirmed": false
112  },
113  "idea_grid": [
114    {
115      "cfg": {
116        "lr": 0.0015,
117        "radius": 0.05
118      },
119      "full": {
120        "mean": 0.0026488977746339515,
121        "std": 0.0009775027285094725,
122        "per_seed": [
123          0.002502700313925743,
124          0.0030538190621882677,
125          0.0011029867455363274,
126          0.0035616853274405003,
127          0.0022829172667115927,
128          0.0012881349539384246,
129          0.003845615778118372,
130          0.0035533227492123842
131        ],
132        "n": 8
133      }
134    },
135    {
136      "cfg": {
137        "lr": 0.003,
138        "radius": 0.05
139      },
140      "full": {
141        "mean": 0.0014651147939730436,
142        "std": 0.0007129411128583655,
143        "per_seed": [
144          0.001081180409528315,
145          0.0008826935081742704,
146          0.0008340234053321183,
147          0.0018998341402038932,
148          0.002233539242297411,
149          0.0008554045343771577,
150          0.0028406656347215176,
151          0.0010935774771496654
152        ],
153        "n": 8
154      }
155    },
156    {
157      "cfg": {
158        "lr": 0.006,
159        "radius": 0.05
160      },
161      "full": {
162        "mean": 0.0010236013476969674,
163        "std": 0.0003525166306615029,
164        "per_seed": [
165          0.0007349217194132507,
166          0.0010386385256424546,
167          0.0008291155099868774,
168          0.001807487802579999,
169          0.0007241228595376015,
170          0.000720260024536401,
171          0.0013001237530261278,
172          0.0010341405868530273
173        ],
174        "n": 8
175      }
176    }
177  ],
178  "math_sanity": [
179    {
180      "input_norm": 0.01,
181      "exact_norm": 0.01,
182      "smooth_norm": 0.009805806756909202,
183      "exact_bound_ok": true
184    },
185    {
186      "input_norm": 0.05,
187      "exact_norm": 0.05000000000000002,
188      "smooth_norm": 0.03535533905932738,
189      "exact_bound_ok": true
190    },
191    {
192      "input_norm": 0.2,
193      "exact_norm": 0.050000000744808065,
194      "smooth_norm": 0.048507125007266595,
195      "exact_bound_ok": false
196    },
197    {
198      "input_norm": 1.0,
199      "exact_norm": 0.05000000074500806,
200      "smooth_norm": 0.049937616943892225,
201      "exact_bound_ok": false
202    }
203  ],
204  "protocol": {
205    "epochs": 12,
206    "batch": 128,
207    "paired_seeds": [
208      0,
209      1,
210      2,
211      3,
212      4,
213      5,
214      6,
215      7
216    ],
217    "track_reason": "dynamics matches optimizer stability and bifurcation structure"
218  }
219}