1{
  2  "bench_version": 1,
  3  "track": "conditional_sequence_diffusion",
  4  "model": "score_mlp",
  5  "metric_direction": "lower is better",
  6  "n_seeds": 8,
  7  "baseline": {
  8    "best_cfg": {
  9      "lr": 0.003,
 10      "epochs": 12,
 11      "steps": 16
 12    },
 13    "sweep": [
 14      {
 15        "cfg": {
 16          "lr": 0.001,
 17          "epochs": 12,
 18          "steps": 8
 19        },
 20        "mean": 0.3794306069612503
 21      },
 22      {
 23        "cfg": {
 24          "lr": 0.001,
 25          "epochs": 12,
 26          "steps": 16
 27        },
 28        "mean": 0.36941855028271675
 29      },
 30      {
 31        "cfg": {
 32          "lr": 0.003,
 33          "epochs": 12,
 34          "steps": 8
 35        },
 36        "mean": 0.3344487212598324
 37      },
 38      {
 39        "cfg": {
 40          "lr": 0.003,
 41          "epochs": 12,
 42          "steps": 16
 43        },
 44        "mean": 0.32668623328208923
 45      },
 46      {
 47        "cfg": {
 48          "lr": 0.01,
 49          "epochs": 12,
 50          "steps": 8
 51        },
 52        "mean": 0.3527134172618389
 53      },
 54      {
 55        "cfg": {
 56          "lr": 0.01,
 57          "epochs": 12,
 58          "steps": 16
 59        },
 60        "mean": 0.3493003807961941
 61      }
 62    ],
 63    "full": {
 64      "mean": 0.32668623328208923,
 65      "std": 0.0183314657731986,
 66      "per_seed": [
 67        0.34015029668807983,
 68        0.3335983455181122,
 69        0.34249189496040344,
 70        0.2935677170753479,
 71        0.3214184641838074,
 72        0.3177299499511719,
 73        0.35397353768348694,
 74        0.3105596601963043
 75      ],
 76      "n": 8
 77    }
 78  },
 79  "idea": {
 80    "mean": 0.3268098309636116,
 81    "std": 0.018441350747342378,
 82    "per_seed": [
 83      0.3401668965816498,
 84      0.33357036113739014,
 85      0.3425343632698059,
 86      0.2938295602798462,
 87      0.32191580533981323,
 88      0.3179206848144531,
 89      0.3546198308467865,
 90      0.30992114543914795
 91    ],
 92    "n": 8,
 93    "trials": [
 94      {
 95        "cfg": {
 96          "lr": 0.001,
 97          "epochs": 12,
 98          "steps": 8
 99        },
100        "mean": 0.3800293132662773
101      },
102      {
103        "cfg": {
104          "lr": 0.001,
105          "epochs": 12,
106          "steps": 16
107        },
108        "mean": 0.3697975389659405
109      },
110      {
111        "cfg": {
112          "lr": 0.003,
113          "epochs": 12,
114          "steps": 8
115        },
116        "mean": 0.33464548736810684
117      },
118      {
119        "cfg": {
120          "lr": 0.003,
121          "epochs": 12,
122          "steps": 16
123        },
124        "mean": 0.3268098309636116
125      },
126      {
127        "cfg": {
128          "lr": 0.01,
129          "epochs": 12,
130          "steps": 8
131        },
132        "mean": 0.35246242582798004
133      },
134      {
135        "cfg": {
136          "lr": 0.01,
137          "epochs": 12,
138          "steps": 16
139        },
140        "mean": 0.34918101504445076
141      }
142    ]
143  },
144  "comparison": {
145    "delta_mean": 0.00012359768152236938,
146    "idea_wins": 2,
147    "n_pairs": 8,
148    "per_seed_diffs": [
149      1.659989356994629e-05,
150      -2.79843807220459e-05,
151      4.246830940246582e-05,
152      0.000261843204498291,
153      0.0004973411560058594,
154      0.00019073486328125,
155      0.0006462931632995605,
156      -0.0006385147571563721
157    ],
158    "p_value": 0.34835,
159    "mde": 0.00032512406414669,
160    "mde_rel_pct": 0.09952181360086565,
161    "verdict": "no measurable effect",
162    "system_worked": false,
163    "permutation_pvalue": 0.34835
164  },
165  "mechanism_signature": {
166    "density_mean": 0.005077726362893979,
167    "density_max": 0.005121961701661348,
168    "width_inverse_sqrt_corr": 0.9998106126499892,
169    "observed_cost_cv": 7.354513075912003e-05,
170    "predicted_cost_equalization": "low CV under sqrt allocation",
171    "confirmed": true
172  },
173  "custom_track": {
174    "name": "conditional_sequence_diffusion",
175    "file": "custom_diffusion_track.py",
176    "domain": "sequence-level diffusion sampling"
177  }
178}