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      "weight_decay": 0.0001
 11    },
 12    "sweep": [
 13      {
 14        "cfg": {
 15          "lr": 0.001,
 16          "weight_decay": 0.0
 17        },
 18        "mean": 0.004298779065720737
 19      },
 20      {
 21        "cfg": {
 22          "lr": 0.001,
 23          "weight_decay": 0.0001
 24        },
 25        "mean": 0.0042995543335564435
 26      },
 27      {
 28        "cfg": {
 29          "lr": 0.003,
 30          "weight_decay": 0.0
 31        },
 32        "mean": 0.0034615940821822733
 33      },
 34      {
 35        "cfg": {
 36          "lr": 0.003,
 37          "weight_decay": 0.0001
 38        },
 39        "mean": 0.003432174737099558
 40      },
 41      {
 42        "cfg": {
 43          "lr": 0.01,
 44          "weight_decay": 0.0
 45        },
 46        "mean": 0.0014865802077110857
 47      },
 48      {
 49        "cfg": {
 50          "lr": 0.01,
 51          "weight_decay": 0.0001
 52        },
 53        "mean": 0.0014641905727330595
 54      }
 55    ],
 56    "full": {
 57      "mean": 0.0014106978487689048,
 58      "std": 0.00038184104551606325,
 59      "per_seed": [
 60        0.0009313806658610702,
 61        0.0018688973505049944,
 62        0.0010692982468754053,
 63        0.0019871860276907682,
 64        0.0009491611272096634,
 65        0.0014028067234903574,
 66        0.0016651722835376859,
 67        0.0014116803649812937
 68      ],
 69      "n": 8
 70    }
 71  },
 72  "idea": {
 73    "mean": 0.0014454916745307855,
 74    "std": 0.0009746414116167044,
 75    "per_seed": [
 76      0.003880624659359455,
 77      0.0016319446731358767,
 78      0.0007165370043367147,
 79      0.001440238207578659,
 80      0.0013295277021825314,
 81      0.0009956876747310162,
 82      0.000869154348038137,
 83      0.0007002191268838942
 84    ],
 85    "n": 8
 86  },
 87  "comparison": {
 88    "delta_mean": 3.4793825761880726e-05,
 89    "idea_wins": 6,
 90    "n_pairs": 8,
 91    "per_seed_diffs": [
 92      0.002949243993498385,
 93      -0.00023695267736911774,
 94      -0.00035276124253869057,
 95      -0.0005469478201121092,
 96      0.00038036657497286797,
 97      -0.00040711904875934124,
 98      -0.0007960179354995489,
 99      -0.0007114612380973995
100    ],
101    "p_value": 0.98335,
102    "mde": 0.0010298298032408945,
103    "mde_rel_pct": 73.00144422419102,
104    "verdict": "no significant win",
105    "system_worked": false
106  },
107  "idea_sweep": [
108    {
109      "cfg": {
110        "lr": 0.01,
111        "lambda_fr": 0.0
112      },
113      "mean": 0.001941849637660198
114    },
115    {
116      "cfg": {
117        "lr": 0.01,
118        "lambda_fr": 0.0001
119      },
120      "mean": 0.0019368745561223477
121    },
122    {
123      "cfg": {
124        "lr": 0.01,
125        "lambda_fr": 0.0005
126      },
127      "mean": 0.0019173361361026764
128    }
129  ],
130  "protocol_note": "baseline sweep used all shared learning rates and weight-decay values",
131  "mechanism_signature": {
132    "quantity": "trained-model local discrete resolvent peak",
133    "predicted_effect": "frequency penalty lowers peak",
134    "baseline_peak_mean": 1.1064540445804596,
135    "idea_peak_mean": 1.0916701406240463,
136    "reduction_fraction": 0.013361516484870337,
137    "confirmed": true
138  }
139}