Convex-gradient robust augmenter / 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      "noise": 0.05
 11    },
 12    "sweep": [
 13      {
 14        "cfg": {
 15          "lr": 0.001,
 16          "noise": 0.0
 17        },
 18        "mean": 0.004072679847013205
 19      },
 20      {
 21        "cfg": {
 22          "lr": 0.001,
 23          "noise": 0.05
 24        },
 25        "mean": 0.004067454778123647
 26      },
 27      {
 28        "cfg": {
 29          "lr": 0.001,
 30          "noise": 0.12
 31        },
 32        "mean": 0.004125235078390688
 33      },
 34      {
 35        "cfg": {
 36          "lr": 0.003,
 37          "noise": 0.0
 38        },
 39        "mean": 0.0009398171678185463
 40      },
 41      {
 42        "cfg": {
 43          "lr": 0.003,
 44          "noise": 0.05
 45        },
 46        "mean": 0.0009230706054950133
 47      },
 48      {
 49        "cfg": {
 50          "lr": 0.003,
 51          "noise": 0.12
 52        },
 53        "mean": 0.0010494974412722513
 54      },
 55      {
 56        "cfg": {
 57          "lr": 0.006,
 58          "noise": 0.0
 59        },
 60        "mean": 0.00099106683046557
 61      },
 62      {
 63        "cfg": {
 64          "lr": 0.006,
 65          "noise": 0.05
 66        },
 67        "mean": 0.0010217856470262632
 68      },
 69      {
 70        "cfg": {
 71          "lr": 0.006,
 72          "noise": 0.12
 73        },
 74        "mean": 0.0012236582988407463
 75      }
 76    ],
 77    "full": {
 78      "mean": 0.0013185284115024842,
 79      "std": 0.00047476113201356173,
 80      "per_seed": [
 81        0.0008951174677349627,
 82        0.0010742346057668328,
 83        0.0008329956326633692,
 84        0.0008899347158148885,
 85        0.0012180092744529247,
 86        0.0017242580652236938,
 87        0.001678277039900422,
 88        0.00223540049046278
 89      ],
 90      "n": 8
 91    }
 92  },
 93  "idea": {
 94    "mean": 0.0013091382279526442,
 95    "std": 0.00047291421988416747,
 96    "per_seed": [
 97      0.0008530844934284687,
 98      0.0012209241976961493,
 99      0.0008412626339122653,
100      0.0008439973462373018,
101      0.0011196484556421638,
102      0.0017749746330082417,
103      0.0016231484478339553,
104      0.002196065615862608
105    ],
106    "n": 8
107  },
108  "comparison": {
109    "delta_mean": -9.390183549840003e-06,
110    "idea_wins": 5,
111    "n_pairs": 8,
112    "per_seed_diffs": [
113      -4.2032974306494e-05,
114      0.00014668959192931652,
115      8.267001248896122e-06,
116      -4.593736957758665e-05,
117      -9.836081881076097e-05,
118      5.0716567784547806e-05,
119      -5.512859206646681e-05,
120      -3.933487460017204e-05
121    ],
122    "p_value": 0.8087,
123    "mde": 6.445221887975781e-05,
124    "mde_rel_pct": 4.888193406944753,
125    "verdict": "no measurable effect",
126    "system_worked": false
127  },
128  "mechanism_signature": {
129    "math_sanity": {
130      "identity_kl": 0.0,
131      "identity_is_minimum": true,
132      "small_scale_kl_over_d_eps2": [
133        0.9966914683191119,
134        0.9934317595506632,
135        0.987054279199204
136      ],
137      "quadratic_ratio_mean": 0.9923925023563264,
138      "predicted_quadratic_limit": 1.0,
139      "dimension_linearity_check": 2.0,
140      "predicted_dimension_ratio": 2.0
141    },
142    "selected_idea_cfg": {
143      "lr": 0.003,
144      "scale": 1.0
145    },
146    "signature_cfg": {
147      "lr": 0.003,
148      "scale": 1.04
149    },
150    "idea_sweep": [
151      {
152        "cfg": {
153          "lr": 0.001,
154          "scale": 1.0
155        },
156        "mean": 0.004072679847013205
157      },
158      {
159        "cfg": {
160          "lr": 0.001,
161          "scale": 1.04
162        },
163        "mean": 0.003918257338227704
164      },
165      {
166        "cfg": {
167          "lr": 0.001,
168          "scale": 1.08
169        },
170        "mean": 0.003774982877075672
171      },
172      {
173        "cfg": {
174          "lr": 0.003,
175          "scale": 1.0
176        },
177        "mean": 0.0009398171678185463
178      },
179      {
180        "cfg": {
181          "lr": 0.003,
182          "scale": 1.04
183        },
184        "mean": 0.0009415723325219005
185      },
186      {
187        "cfg": {
188          "lr": 0.003,
189          "scale": 1.08
190        },
191        "mean": 0.0009481702145421878
192      },
193      {
194        "cfg": {
195          "lr": 0.006,
196          "scale": 1.0
197        },
198        "mean": 0.00099106683046557
199      },
200      {
201        "cfg": {
202          "lr": 0.006,
203          "scale": 1.04
204        },
205        "mean": 0.0009713326871860772
206      },
207      {
208        "cfg": {
209          "lr": 0.006,
210          "scale": 1.08
211        },
212        "mean": 0.0009424890304217115
213      }
214    ],
215    "map": "T(x)=grad(.5*s*||x||^2)=s*x",
216    "predicted_mean_displacement": 0.1655000495910646,
217    "observed_mean_displacement_model_inputs": 0.16549988090991974,
218    "displacement_relative_error": 1.0192211136694558e-06,
219    "predicted_gaussian_kl": 0.0379028843212495,
220    "observed_pushforward_kl_from_exact_gaussian_formula": 0.0379028843212495,
221    "confirmed": true,
222    "trained_model_test_mse": 0.0013091382279526442
223  },
224  "custom_track": null,
225  "track_justification": "Dynamics is structurally matched because the idea targets globally coherent perturbations of controlled state/action rollout distributions; rnn_small is shared."
226}