benchmark-results.json

{
  "manual_audit": {
    "samples": 100,
    "source_storage_exact": 99,
    "fp8_storage_exact": 99,
    "regression_points": 0,
    "failed_samples": [
      {
        "index": 77,
        "expected": "ЕЩЁ",
        "observed": "ЁЩЁ"
      }
    ]
  },
  "byte_identity": {
    "matching_generated_png_sha256": 100,
    "samples": 100
  },
  "source_storage_diagnostics": {
    "ocr_exact": 77,
    "mean_cer": 0.1142,
    "mean_glyph_similarity": 0.9564
  },
  "fp8_storage_diagnostics": {
    "ocr_exact": 77,
    "mean_cer": 0.1142,
    "mean_glyph_similarity": 0.9564
  },
  "edge_retry_same_seed": {
    "expected": "ЕЩЁ",
    "seed": 3076,
    "scales": [
      0.9,
      1.0,
      1.1
    ],
    "observed": [
      "ЁЩЁ",
      "ЁЩЁ",
      "ЁЩЁ"
    ]
  },
  "filled_retry_same_seed": {
    "expected": "ЕЩЁ",
    "seed": 3076,
    "scale": 0.7,
    "observed": "ЁЩЁ"
  }
}