Smooth Spectral Muon / bench_report.json

Mechanism confirmed, baseline not beaten

Raw ⬇ ZIP
  1{
  2  "bench_version": 1,
  3  "track": "tabular",
  4  "model": "mlp_tiny",
  5  "metric_direction": "lower is better",
  6  "n_seeds": 8,
  7  "baseline": {
  8    "best_cfg": {
  9      "lr": 0.006,
 10      "beta1": 0.85,
 11      "weight_decay": 0.0
 12    },
 13    "sweep": [
 14      {
 15        "cfg": {
 16          "lr": 0.0015,
 17          "beta1": 0.85,
 18          "weight_decay": 0.0
 19        },
 20        "mean": 12.617201328277588
 21      },
 22      {
 23        "cfg": {
 24          "lr": 0.0015,
 25          "beta1": 0.95,
 26          "weight_decay": 0.0
 27        },
 28        "mean": 15.512621641159058
 29      },
 30      {
 31        "cfg": {
 32          "lr": 0.003,
 33          "beta1": 0.85,
 34          "weight_decay": 0.0
 35        },
 36        "mean": 8.035151362419128
 37      },
 38      {
 39        "cfg": {
 40          "lr": 0.003,
 41          "beta1": 0.95,
 42          "weight_decay": 0.0
 43        },
 44        "mean": 12.083711862564087
 45      },
 46      {
 47        "cfg": {
 48          "lr": 0.006,
 49          "beta1": 0.85,
 50          "weight_decay": 0.0
 51        },
 52        "mean": 6.865306854248047
 53      },
 54      {
 55        "cfg": {
 56          "lr": 0.006,
 57          "beta1": 0.95,
 58          "weight_decay": 0.0
 59        },
 60        "mean": 8.458353757858276
 61      }
 62    ],
 63    "full": {
 64      "mean": 7.004796802997589,
 65      "std": 0.3739293172768815,
 66      "per_seed": [
 67        6.71429443359375,
 68        7.094020843505859,
 69        6.866542816162109,
 70        6.786369323730469,
 71        6.342711448669434,
 72        7.308849334716797,
 73        7.395407676696777,
 74        7.530178546905518
 75      ],
 76      "n": 8
 77    }
 78  },
 79  "idea": {
 80    "mean": 21.87750768661499,
 81    "std": 1.4977844581209827,
 82    "per_seed": [
 83      22.2957763671875,
 84      21.944984436035156,
 85      22.89360809326172,
 86      19.71537971496582,
 87      19.675622940063477,
 88      22.763729095458984,
 89      21.369779586791992,
 90      24.361181259155273
 91    ],
 92    "n": 8
 93  },
 94  "comparison": {
 95    "delta_mean": 14.872710883617401,
 96    "idea_wins": 0,
 97    "n_pairs": 8,
 98    "per_seed_diffs": [
 99      15.58148193359375,
100      14.850963592529297,
101      16.02706527709961,
102      12.929010391235352,
103      13.332911491394043,
104      15.454879760742188,
105      13.974371910095215,
106      16.831002712249756
107    ],
108    "p_value": 0.0081,
109    "mde": 1.1379574709779237,
110    "mde_rel_pct": 16.24540301427378,
111    "verdict": "idea worse (significant)",
112    "system_worked": false
113  },
114  "mechanism_signature": {
115    "mechanism_signature": {
116      "predicted_response_mean": 0.3883231416550379,
117      "observed_response_mean": 0.3883247250989132,
118      "response_mae": 1.7781029070577825e-06,
119      "update_norm_cv": 0.08146856078007694,
120      "n_observations": 600,
121      "prediction": "observed singular response follows t/sqrt(t^2+1), t=s/sqrt(epsilon)",
122      "confirmed": true
123    },
124    "idea_grid": [
125      {
126        "cfg": {
127          "lr": 0.0015,
128          "beta": 0.9,
129          "c": 0.001
130        },
131        "result": {
132          "mean": 105.54728984832764,
133          "std": 5.39382412585861,
134          "per_seed": [
135            114.30227661132812,
136            109.72885131835938,
137            105.97409057617188,
138            96.68315124511719,
139            103.86304473876953,
140            104.20171356201172,
141            99.61021423339844,
142            110.01497650146484
143          ],
144          "n": 8
145        }
146      },
147      {
148        "cfg": {
149          "lr": 0.003,
150          "beta": 0.9,
151          "c": 0.001
152        },
153        "result": {
154          "mean": 43.18799352645874,
155          "std": 3.9588922932446455,
156          "per_seed": [
157            49.57056427001953,
158            46.739295959472656,
159            42.695228576660156,
160            37.679901123046875,
161            39.12871551513672,
162            43.38615036010742,
163            39.62131118774414,
164            46.68278121948242
165          ],
166          "n": 8
167        }
168      },
169      {
170        "cfg": {
171          "lr": 0.006,
172          "beta": 0.9,
173          "c": 0.001
174        },
175        "result": {
176          "mean": 21.87750768661499,
177          "std": 1.4977844581209827,
178          "per_seed": [
179            22.2957763671875,
180            21.944984436035156,
181            22.89360809326172,
182            19.71537971496582,
183            19.675622940063477,
184            22.763729095458984,
185            21.369779586791992,
186            24.361181259155273
187          ],
188          "n": 8
189        }
190      }
191    ],
192    "track_rationale": "Optimizer modification belongs on Friedman#1 tabular MLP; architecture and data are shared."
193  }
194}