1{
  2  "bench_version": 1,
  3  "track": "sequence",
  4  "model": "modular_sequence_local",
  5  "metric_direction": "lower is better",
  6  "n_seeds": 8,
  7  "baseline": {
  8    "best_cfg": {
  9      "lr": 0.003,
 10      "epochs": 12,
 11      "batch": 128
 12    },
 13    "sweep": [
 14      {
 15        "cfg": {
 16          "lr": 0.001,
 17          "epochs": 12,
 18          "batch": 128
 19        },
 20        "mean": 0.15198808163404465
 21      },
 22      {
 23        "cfg": {
 24          "lr": 0.003,
 25          "epochs": 12,
 26          "batch": 128
 27        },
 28        "mean": 0.1417534351348877
 29      },
 30      {
 31        "cfg": {
 32          "lr": 0.01,
 33          "epochs": 12,
 34          "batch": 128
 35        },
 36        "mean": 0.15112163126468658
 37      }
 38    ],
 39    "full": {
 40      "mean": 0.13169990107417107,
 41      "std": 0.017087285342305542,
 42      "per_seed": [
 43        0.1741165667772293,
 44        0.12635251879692078,
 45        0.13050207495689392,
 46        0.13604258000850677,
 47        0.11734391003847122,
 48        0.11690718680620193,
 49        0.1270616501569748,
 50        0.1252727210521698
 51      ],
 52      "n": 8
 53    }
 54  },
 55  "idea": {
 56    "per_seed": [
 57      0.1741165667772293,
 58      0.12635251879692078,
 59      0.13050207495689392,
 60      0.13604258000850677,
 61      0.11734391003847122,
 62      0.11690718680620193,
 63      0.1270616501569748,
 64      0.1252727210521698
 65    ],
 66    "mean": 0.13169990107417107,
 67    "std": 0.017087285342305542,
 68    "n": 8
 69  },
 70  "comparison": {
 71    "delta_mean": 0.0,
 72    "idea_wins": 0,
 73    "n_pairs": 8,
 74    "per_seed_diffs": [
 75      0.0,
 76      0.0,
 77      0.0,
 78      0.0,
 79      0.0,
 80      0.0,
 81      0.0,
 82      0.0
 83    ],
 84    "p_value": 1.0,
 85    "mde": 0.0,
 86    "mde_rel_pct": 0.0,
 87    "verdict": "no measurable effect",
 88    "system_worked": false
 89  },
 90  "track_justification": "sequence is structurally matched: multi-token window correlations and a sequence-level model",
 91  "idea_sweep": [
 92    {
 93      "cfg": {
 94        "lr": 0.001,
 95        "epochs": 12,
 96        "batch": 128
 97      },
 98      "mean": 0.14950868487358093
 99    },
100    {
101      "cfg": {
102        "lr": 0.003,
103        "epochs": 12,
104        "batch": 128
105      },
106      "mean": 0.13169990107417107
107    },
108    {
109      "cfg": {
110        "lr": 0.01,
111        "epochs": 12,
112        "batch": 128
113      },
114      "mean": 0.14627121202647686
115    }
116  ],
117  "canonical_proposal": [
118    "g3",
119    "g1",
120    "global",
121    "g0",
122    "g2"
123  ],
124  "canonical_word": [
125    "g1",
126    "g3",
127    "global",
128    "g0",
129    "g2"
130  ],
131  "mechanism_signature": {
132    "prediction": "disjoint-support swaps preserve the trained network output; group/global swaps do not",
133    "predicted_commuting_max": 1e-05,
134    "observed_commuting_max": 0.0,
135    "predicted_noncommuting_positive": true,
136    "observed_noncommuting_median": 1.8506829738616943,
137    "confirmed": true,
138    "n_trained_models": 8
139  }
140}