Evaluation

1 posts

94% of Fields Correct. Now What? A Two-Stage Architecture for Scanned Documents

Across 38 scans and 4,732 field assertions, the classic OCR pipeline passed 94.08% of fields, but only six documents passed every check. Page transcription, extraction and large-schema experiments show how to separate reading from interpretation, retain source evidence, and avoid treating model agreement as proof of correctness.

  • AI
  • OCR
  • VLM
  • Document Processing
November 16, 2026 · 20 min · 4231 words · Sam Peng
Scans become traceable text, then structured fields, followed by validation and review