Transient-risk certificate for Langevin training / 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": 0.003,
 10      "temperature": 0.0005
 11    },
 12    "sweep": [
 13      {
 14        "cfg": {
 15          "lr": 0.001,
 16          "temperature": 0.0005
 17        },
 18        "mean": 6.877625823020935
 19      },
 20      {
 21        "cfg": {
 22          "lr": 0.001,
 23          "temperature": 0.002
 24        },
 25        "mean": 6.878889203071594
 26      },
 27      {
 28        "cfg": {
 29          "lr": 0.003,
 30          "temperature": 0.0005
 31        },
 32        "mean": 6.73323667049408
 33      },
 34      {
 35        "cfg": {
 36          "lr": 0.003,
 37          "temperature": 0.002
 38        },
 39        "mean": 6.734855055809021
 40      },
 41      {
 42        "cfg": {
 43          "lr": 0.006,
 44          "temperature": 0.0005
 45        },
 46        "mean": 7.187614321708679
 47      },
 48      {
 49        "cfg": {
 50          "lr": 0.006,
 51          "temperature": 0.002
 52        },
 53        "mean": 7.431662321090698
 54      }
 55    ],
 56    "full": {
 57      "mean": 6.753931879997253,
 58      "std": 0.15501289580001476,
 59      "per_seed": [
 60        6.430861949920654,
 61        6.871230602264404,
 62        6.933614253997803,
 63        6.697239875793457,
 64        6.64493465423584,
 65        6.860416412353516,
 66        6.7199554443359375,
 67        6.873201847076416
 68      ],
 69      "n": 8
 70    }
 71  },
 72  "idea": {
 73    "mean": 6.753931879997253,
 74    "std": 0.15501289580001476,
 75    "per_seed": [
 76      6.430861949920654,
 77      6.871230602264404,
 78      6.933614253997803,
 79      6.697239875793457,
 80      6.64493465423584,
 81      6.860416412353516,
 82      6.7199554443359375,
 83      6.873201847076416
 84    ],
 85    "n": 8
 86  },
 87  "comparison": {
 88    "delta_mean": 0.0,
 89    "idea_wins": 0,
 90    "n_pairs": 8,
 91    "per_seed_diffs": [
 92      0.0,
 93      0.0,
 94      0.0,
 95      0.0,
 96      0.0,
 97      0.0,
 98      0.0,
 99      0.0
100    ],
101    "p_value": 1.0,
102    "mde": 0.0,
103    "mde_rel_pct": 0.0,
104    "verdict": "no measurable effect",
105    "system_worked": false
106  },
107  "idea_sweep": [
108    [
109      6.877625823020935,
110      {
111        "lr": 0.001,
112        "temperature": 0.0005
113      }
114    ],
115    [
116      6.878889203071594,
117      {
118        "lr": 0.001,
119        "temperature": 0.002
120      }
121    ],
122    [
123      6.73323667049408,
124      {
125        "lr": 0.003,
126        "temperature": 0.0005
127      }
128    ]
129  ],
130  "protocol_note": "Official bench tabular track; custom Langevin loop is the intervention.",
131  "mechanism_signature": {
132    "prediction": "observed unsafe frequency should not exceed certificate envelope",
133    "predicted_final_bound": 1.0,
134    "observed_mean_unsafe_frequency": 1.0,
135    "observed_post_stop_unsafe_frequency": 1.0,
136    "n_trained_models": 8,
137    "confirmed": true
138  }
139}