Most AI conversations start at the model. In document-heavy organizations, trust breaks earlier. OCR misreads a clause. Chunking loses the page. Search returns the right document but the wrong section, and the citation points nowhere. By the time the model runs, the foundation has already shifted.
How OEP fits
- OCR quality as a business layer. Detection, recognition, layout, and tables run on the device, with quality gates, repair passes, and review queues. Measured, not assumed.
- Layout is business logic. Sections, tables, forms, and figures survive extraction as structure. Tables come back as cells with rows, columns, and merged spans, not flattened text. A clause keeps its place in the document.
- Page anchors on everything. Every extracted object knows its source page and region. Evidence is a pointer you can follow, not a paraphrase you must trust.
- Confidence you can act on. Each field carries a state, read it, check it, or send it to review, fused from the recognizer score, image quality, a format check, and a second read. One page number is never the whole story.
- Identifiers that check themselves. National ID numbers, vehicle VINs, passport machine-readable zones, and barcodes are validated by their own check digits, so a confident but wrong read fails instead of passing quietly.
- A chain of custody. The original is never overwritten. Every correction step is hashed, so you can show what was done to a page and what was not. Conservative, evidence-safe corrections stay separate from readability enhancements.
- Provenance for PDFs. The pipeline tells a born-digital PDF from a scan, flags when a text layer sits over a page-sized image (a likely OCR overlay whose text may not match the rendered page), and reports a signature as present or as changed after signing. It does not pronounce a document authentic or forged.
- Long and wide documents. A receipt or page too large for one frame is captured in overlapping shots and joined by matching content, with any seam it cannot confirm flagged rather than guessed.
- Figures, in the mode you choose. Citation mode preserves the original figure, cropped and hash-pinned, never redrawn. Hybrid pairs preservation with cryptographic pinning for anything where authenticity matters: seals, stamps, and signatures. Regenerative mode, where a vector backend is configured, redraws a diagram and keeps it only when it validates against the original, preserving the original otherwise. With no vector backend configured, every mode preserves rather than redraws.
- Structured extraction by document type. Receipts, invoices, contracts, warranties: per-type field schemas as governed, validated pack content.
What exists today
An OCR lane built and proven on real legal and educational documents. The evidence machinery above, the confidence states, the field validators, the hashed chain of custody, the figure-preservation modes, and the PDF inspection, is built and tested in the shared engine. ScanVecta is the consumer expression. The same pipeline takes enterprise collections. See ScanVecta →
What we won’t tell you
No OCR is perfect. Ours ships confidence scores, review queues, and honest gaps instead of pretending. See how we bound our claims.