Shared-Private Matrix-Weighted Expert Layers / 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": 0.01,
 10      "wd": 0.0001,
 11      "epochs": 12,
 12      "batch": 128,
 13      "rank": 8,
 14      "coupling": 0.0
 15    },
 16    "sweep": [
 17      {
 18        "cfg": {
 19          "lr": 0.001,
 20          "wd": 0.0001,
 21          "epochs": 12,
 22          "batch": 128,
 23          "rank": 8,
 24          "coupling": 0.0
 25        },
 26        "mean": 0.01910123834386468
 27      },
 28      {
 29        "cfg": {
 30          "lr": 0.003,
 31          "wd": 0.0001,
 32          "epochs": 12,
 33          "batch": 128,
 34          "rank": 8,
 35          "coupling": 0.0
 36        },
 37        "mean": 0.005347788450308144
 38      },
 39      {
 40        "cfg": {
 41          "lr": 0.01,
 42          "wd": 0.0001,
 43          "epochs": 12,
 44          "batch": 128,
 45          "rank": 8,
 46          "coupling": 0.0
 47        },
 48        "mean": 0.0039304872043430805
 49      }
 50    ],
 51    "full": {
 52      "mean": 0.003924444579752162,
 53      "std": 0.0009507399174062226,
 54      "per_seed": [
 55        0.005481751170009375,
 56        0.003171606920659542,
 57        0.0028356127440929413,
 58        0.004232977982610464,
 59        0.004473012872040272,
 60        0.0031893684063106775,
 61        0.005031306762248278,
 62        0.0029799197800457478
 63      ],
 64      "n": 8
 65    }
 66  },
 67  "idea": {
 68    "mean": 0.0036855603102594614,
 69    "std": 0.0008949623394657411,
 70    "per_seed": [
 71      0.0049509322270751,
 72      0.0027915509417653084,
 73      0.0026541445404291153,
 74      0.0039238822646439075,
 75      0.0043672905303537846,
 76      0.0031351130455732346,
 77      0.004849778488278389,
 78      0.002811790443956852
 79    ],
 80    "n": 8
 81  },
 82  "comparison": {
 83    "delta_mean": -0.0002388842694927007,
 84    "idea_wins": 8,
 85    "n_pairs": 8,
 86    "per_seed_diffs": [
 87      -0.0005308189429342747,
 88      -0.0003800559788942337,
 89      -0.000181468203663826,
 90      -0.00030909571796655655,
 91      -0.0001057223416864872,
 92      -5.425536073744297e-05,
 93      -0.0001815282739698887,
 94      -0.0001681293360888958
 95    ],
 96    "p_value": 0.0081,
 97    "mde": 0.00013163873291981864,
 98    "mde_rel_pct": 3.3543277333816226,
 99    "verdict": "idea better (significant)",
100    "system_worked": true
101  },
102  "idea_sweep": [
103    {
104      "cfg": {
105        "lr": 0.01,
106        "wd": 0.0001,
107        "epochs": 12,
108        "batch": 128,
109        "rank": 8,
110        "coupling": 0.0
111      },
112      "mean": 0.003924444579752162
113    },
114    {
115      "cfg": {
116        "lr": 0.01,
117        "wd": 0.0001,
118        "epochs": 12,
119        "batch": 128,
120        "rank": 8,
121        "coupling": 0.03
122      },
123      "mean": 0.0036855603102594614
124    },
125    {
126      "cfg": {
127        "lr": 0.01,
128        "wd": 0.0001,
129        "epochs": 12,
130        "batch": 128,
131        "rank": 8,
132        "coupling": 0.1
133      },
134      "mean": 0.003690520068630576
135    }
136  ],
137  "device": "cpu",
138  "mechanism_signature": {
139    "predicted": "shared disagreement decreases while private variance remains nonzero",
140    "observed_idea_shared_disagreement_mean": 0.020418916479684412,
141    "observed_baseline_shared_disagreement_mean": 0.06511001940816641,
142    "observed_idea_private_variance_mean": 0.08402373269200325,
143    "observed_baseline_private_variance_mean": 0.08042469155043364,
144    "confirmed": true
145  },
146  "custom_track": null
147}