Ground-truth cleanup needs separate evidence layers and adversarially complete review

Ground-truth cleanup for OCR/VLM evaluation should preserve four distinct layers: literal image transcription, canonical normalization, semantic field assignment, and quarantine. A deterministic builder should bind every decision to immutable image/model receipts, require the adjudication to be the complete canonical receipt export, keep independent and lead-review verdicts separate, and exclude ambiguous fields from scoring rather than guessing. Final-review coverage must include every changed or quarantined ledger decision; sampling is useful but cannot certify the release. Set overrides should be grounded in text actually observed by a blind reader or independent reviewer, and normalization must never substitute confusable characters.

This approach applies to image-to-structured-data benchmarks where existing labels are model-produced or inconsistently formatted. It improves auditability and prevents inflated accuracy, but it does not turn same-model review into human gold; correlated model bias and unreviewed non-target rows remain explicit limits.

Applies when: Cleaning OCR or VLM benchmark labels before trusting accuracy comparisons or production thresholds. Keywords: OCR ground truth label cleanup receipts quarantine adjudication normalization review