{
  "model_config": {
    "vocab_size": 14,
    "d_model": 64,
    "n_heads": 4,
    "n_layers": 1,
    "max_seq_len": 12
  },
  "sample_size": 10,
  "seed": 123,
  "batch_size": 5,
  "device": "cpu",
  "reconstruction_error": 9.947598300641403e-14,
  "sample_accuracy": 1.0,
  "components": [
    "residual",
    "head_0",
    "head_1",
    "head_2",
    "head_3",
    "total"
  ],
  "diagnostic_examples": [
    [
      3,
      7,
      2,
      5,
      1
    ],
    [
      9,
      1,
      2,
      3,
      4
    ],
    [
      5,
      4,
      5,
      1,
      0
    ],
    [
      0,
      0,
      0,
      0,
      0
    ],
    [
      0,
      1,
      0,
      0,
      0
    ]
  ],
  "method": "At [ANS], residual logits plus each head's value-weighted output projected through its W_O slice and the unembed exactly reconstruct the model logits.",
  "primary_metric": "For each input, use the model's final strongest wrong digit and compute component[correct_digit] - component[final_wrong_digit]."
}