Legal document OCR that respects Bates-stamped scans
A production labeled “searchable PDF” often is not. Confidentiality headers, Bates stamps, and a thin text layer can look like content to a naive extractor while the real pages are images. Reviewers then search a case and miss the exhibit that mattered.
exhibit.law’s document pipeline is built for that failure mode: measure usable text per page, escalate only when needed, and never claim coverage the reader did not earn.
Why average density lies
Judging a whole PDF by average characters-per-page fails in both directions. A scan with a repeated header and footer can clear a document-wide bar while every body page is blank to search. A typed cover pleading in front of scanned exhibits can pull the average up so the scans never get OCR.
exhibit.law decides per page, strips repeated boilerplate before measuring, and skips blank sheets before spending OCR or vision budget.
Local-first reading ladder
Documents climb a ladder: embedded text layer first, then free OCR, then a vision model only for pages that still fail. Under normal load that work stays on closed, self-hosted models on private infrastructure. Peak demand may use third-party model APIs under no-training terms — see the AI Disclosure.
When vision cannot read a page, the transcript marks it unread. Returning a weak OCR pass as if it succeeded would invent coverage someone will cite.
- Text layer accepted only when per-page density clears a real threshold
- OCR has its own acceptance floor — calibrated on real pages, not guessed
- Blank pages are skipped before paid or heavy rungs
- Unread pages are listed so reviewers know what was never read
Search and the timeline
Completed transcript text is indexed for case-scoped search alongside event titles and descriptions. Draft timeline events from documents still wait for human review; upload-time dates stay placeholders until someone sets a real moment.
Identical files dedupe inside one case and never share a media row across matters.
Try a scan-heavy PDF
Create a free account, open a matter, and upload a Bates-stamped production slice. Open the transcript to see method and coverage notes, then walk Review before drafts harden.
Related: discovery management and body-cam transcription.