Unexplained Topology Distillation / bench_report.json

Mechanism confirmed, baseline not beaten

Raw ⬇ ZIP
  1{
  2  "bench_version": 1,
  3  "track": "vision",
  4  "model": "cnn_small",
  5  "metric_direction": "lower is better",
  6  "n_seeds": 8,
  7  "baseline": {
  8    "best_cfg": {
  9      "lr": 0.003,
 10      "coef": 0.1
 11    },
 12    "sweep": [
 13      {
 14        "cfg": {
 15          "lr": 0.001,
 16          "coef": 0.03
 17        },
 18        "mean": 0.8412500023841858
 19      },
 20      {
 21        "cfg": {
 22          "lr": 0.003,
 23          "coef": 0.1
 24        },
 25        "mean": 0.8349999934434891
 26      },
 27      {
 28        "cfg": {
 29          "lr": 0.01,
 30          "coef": 0.3
 31        },
 32        "mean": 0.8937499821186066
 33      }
 34    ],
 35    "full": {
 36      "mean": 0.8243750035762787,
 37      "std": 0.04605956605822434,
 38      "per_seed": [
 39        0.7699999809265137,
 40        0.9049999713897705,
 41        0.8650000095367432,
 42        0.800000011920929,
 43        0.8149999976158142,
 44        0.8700000047683716,
 45        0.7850000262260437,
 46        0.7850000262260437
 47      ],
 48      "n": 8
 49    }
 50  },
 51  "idea": {
 52    "mean": 0.8062499985098839,
 53    "std": 0.03028510823634387,
 54    "per_seed": [
 55      0.7699999809265137,
 56      0.8550000190734863,
 57      0.8349999785423279,
 58      0.7549999952316284,
 59      0.8050000071525574,
 60      0.8050000071525574,
 61      0.8050000071525574,
 62      0.8199999928474426
 63    ],
 64    "n": 8,
 65    "unexplained_mass_per_seed": [
 66      0.029143333435058594,
 67      0.017951250076293945,
 68      0.0059850215911865234,
 69      0.02043771743774414,
 70      0.05716085433959961,
 71      0.006631910800933838,
 72      0.030822336673736572,
 73      0.006145954132080078
 74    ]
 75  },
 76  "comparison": {
 77    "delta_mean": -0.018125005066394806,
 78    "idea_wins": 5,
 79    "n_pairs": 8,
 80    "per_seed_diffs": [
 81      0.0,
 82      -0.04999995231628418,
 83      -0.030000030994415283,
 84      -0.04500001668930054,
 85      -0.009999990463256836,
 86      -0.06499999761581421,
 87      0.019999980926513672,
 88      0.034999966621398926
 89    ],
 90    "p_value": 0.1876,
 91    "mde": 0.029557209568880215,
 92    "mde_rel_pct": 3.585408271800579,
 93    "verdict": "no significant win",
 94    "system_worked": false
 95  },
 96  "mechanism_signature": {
 97    "prediction": "directional PCE should reduce teacher activation-event unexplained mass relative to feature-MSE baseline",
 98    "predicted_vs_observed": {
 99      "baseline_unexplained_mass": 0.12066004425287247,
100      "idea_unexplained_mass": 0.021784797310829163
101    },
102    "confirmed": true,
103    "measurement": "trained teacher/student CNN activation ranges on test inputs"
104  },
105  "sanity_check": {
106    "probability_valid": true,
107    "excess_identity_max_abs_error": 4.172325134277344e-06
108  }
109}