{
  "release": "h3-final-20261021T220053Z-wer-v1",
  "logical_cohort": "h3-final-20261021T220053Z",
  "prefix": "alignment-v1/h3-final-20261021T220053Z-wer-v1/",
  "policy": {
    "id": "h3_prompt_wer_nfkc_casefold_punctuation_whitespace_v1",
    "reference": "alignment.source_reference.text (existing de-tagged generation prompt)",
    "hypothesis": "alignment.transcription.text (actual Parakeet text with model control tags already removed)",
    "raw": "Unicode whitespace word tokens; case and punctuation retained",
    "normalized": "Unicode NFKC, casefold, replace Unicode category P punctuation with ASCII spaces, split/collapse Unicode whitespace",
    "numbers": "No number expansion, transliteration, accent removal or language-specific rewriting",
    "wer": "(substitutions + deletions + insertions) / reference_words; not clipped to 1",
    "cer": "Same normalized text, whitespace removed, Unicode code-point edit distance / reference_characters",
    "empty_reference": "null error rate; insertion counts retained when hypothesis exists",
    "empty_hypothesis": "Valid empty ASR output is scored as all deletions against a nonempty reference",
    "unavailable_asr": "Failed/skipped/missing ASR is unavailable, not silently treated as an empty recognition",
    "tokenization": "Whitespace WER for every language; not linguistic word segmentation for Japanese/Chinese/Thai/Cantonese. Normalized CER is also provided for all languages.",
    "interpretation": "Agreement with intended generation script, not human-verified ASR accuracy",
    "distance_implementation": "rapidfuzz.distance.Levenshtein.editops; unit substitution/deletion/insertion costs"
  },
  "policy_sha256": "007287f56db1b47fdff6e85b88790fc7ea52bd4dbbe06a7ea3b51af376494f8a",
  "shards": 50612,
  "clips": 10122224,
  "statuses": {
    "scored": 10122098,
    "unavailable": 126
  },
  "reasons": {
    "empty_normalized_reference": 126
  },
  "manifest_bytes": 732139487621,
  "embeddings_bytes": 22885234864,
  "raw": {
    "substitutions": 74045294,
    "deletions": 5173662,
    "insertions": 8731353,
    "reference_words": 289541529,
    "hypothesis_words": 293099220,
    "scored_clips": 10122098,
    "sum_clip_wer": 5343846.6731266705,
    "valid_edit_errors": 87946679,
    "valid_reference_words": 289541529,
    "micro_wer": 0.3037446106737939,
    "mean_clip_wer": 0.5279386420805915
  },
  "normalized": {
    "substitutions": 6426245,
    "deletions": 5648934,
    "insertions": 7922821,
    "reference_words": 302692171,
    "hypothesis_words": 304966058,
    "scored_clips": 10122098,
    "sum_clip_wer": 1411534.5652237942,
    "valid_edit_errors": 19994253,
    "valid_reference_words": 302692171,
    "micro_wer": 0.06605474113831639,
    "mean_clip_wer": 0.1394507902634211
  },
  "normalized_character": {
    "substitutions": 7932164,
    "deletions": 18021783,
    "insertions": 11975295,
    "reference_characters": 1139644149,
    "hypothesis_characters": 1133597661,
    "scored_clips": 10122098,
    "sum_clip_cer": 451275.15594872157,
    "valid_edit_errors": 37914883,
    "valid_reference_characters": 1139644149,
    "micro_cer": 0.03326905423352461,
    "mean_clip_cer": 0.044583164078111234
  }
}