1{
  2  "bench_version": 1,
  3  "track": "sequence",
  4  "model": "transformer_tiny",
  5  "metric_direction": "lower is better",
  6  "n_seeds": 8,
  7  "baseline": {
  8    "best_cfg": {
  9      "lr": 0.0015,
 10      "gamma": 1.0
 11    },
 12    "sweep": [
 13      {
 14        "cfg": {
 15          "lr": 0.0015,
 16          "gamma": 1.0
 17        },
 18        "mean": 0.21315788850188255
 19      },
 20      {
 21        "cfg": {
 22          "lr": 0.003,
 23          "gamma": 1.0
 24        },
 25        "mean": 0.2156946398317814
 26      },
 27      {
 28        "cfg": {
 29          "lr": 0.006,
 30          "gamma": 1.0
 31        },
 32        "mean": 0.25452903658151627
 33      }
 34    ],
 35    "full": {
 36      "mean": 0.2124029453843832,
 37      "std": 0.029564272137155054,
 38      "per_seed": [
 39        0.22719556093215942,
 40        0.1741580069065094,
 41        0.20896968245506287,
 42        0.24230830371379852,
 43        0.19200055301189423,
 44        0.1792486160993576,
 45        0.2660641670227051,
 46        0.2092786729335785
 47      ],
 48      "n": 8
 49    }
 50  },
 51  "idea": {
 52    "mean": 0.21240293979644775,
 53    "std": 0.02956426589305894,
 54    "per_seed": [
 55      0.22719556093215942,
 56      0.1741579920053482,
 57      0.20896968245506287,
 58      0.24230830371379852,
 59      0.19200055301189423,
 60      0.1792486310005188,
 61      0.2660641372203827,
 62      0.2092786580324173
 63    ],
 64    "n": 8
 65  },
 66  "comparison": {
 67    "delta_mean": -5.587935447692871e-09,
 68    "idea_wins": 3,
 69    "n_pairs": 8,
 70    "per_seed_diffs": [
 71      0.0,
 72      -1.4901161193847656e-08,
 73      0.0,
 74      0.0,
 75      0.0,
 76      1.4901161193847656e-08,
 77      -2.9802322387695312e-08,
 78      -1.4901161193847656e-08
 79    ],
 80    "p_value": 0.5058,
 81    "mde": 1.1414612684177589e-08,
 82    "mde_rel_pct": 5.374036910609076e-06,
 83    "verdict": "no measurable effect",
 84    "system_worked": false
 85  },
 86  "protocol_notes": "Matched sequence track; shared 2-block d=64 causal transformer, differing only in relative lag bias: unconstrained table versus 8-positive-exponential mixture.",
 87  "mechanism_signature": {
 88    "prediction": "trained mixture log-kernel bias decreases with lag; retained content logits can make total attention nonmonotone",
 89    "observed_bias_max_first_difference": -0.044586181640625,
 90    "observed_bias_min_second_difference": 0.0005350112915039062,
 91    "observed_attention_max_first_difference": 0.030105771829364122,
 92    "n_models": 8,
 93    "confirmed": true
 94  },
 95  "idea_sweep": [
 96    {
 97      "cfg": {
 98        "lr": 0.0015,
 99        "gamma": 1.0
100      },
101      "mean": 0.21315788477659225
102    },
103    {
104      "cfg": {
105        "lr": 0.003,
106        "gamma": 1.0
107      },
108      "mean": 0.21569465845823288
109    },
110    {
111      "cfg": {
112        "lr": 0.006,
113        "gamma": 1.0
114      },
115      "mean": 0.25452912598848343
116    }
117  ],
118  "idea_config": {
119    "lr": 0.0015,
120    "gamma": 1.0
121  },
122  "search_space": {
123    "baseline_grid": [
124      {
125        "lr": 0.0015,
126        "gamma": 1.0
127      },
128      {
129        "lr": 0.003,
130        "gamma": 1.0
131      },
132      {
133        "lr": 0.006,
134        "gamma": 1.0
135      }
136    ],
137    "idea_grid": [
138      {
139        "lr": 0.0015,
140        "gamma": 1.0
141      },
142      {
143        "lr": 0.003,
144        "gamma": 1.0
145      },
146      {
147        "lr": 0.006,
148        "gamma": 1.0
149      }
150    ]
151  }
152}