One signal

A new document-review workflow combines OCR confidence scores with AI judgements about whether extracted values make sense, routing suspect fields to a person rather than trusting either signal alone. The companies’ medication-list demonstration shows why that distinction matters: a clearly printed but implausible instruction can be confidently read and still need review.

What happened News.LAVX reports that Unsiloed’s extraction system sends structured fields to TypeSafe’s Jev judgement engine, which answers focused questions and returns typed scores. Application rules then combine those answers with extraction and source-grounding scores to decide what a human should check. In the medication example, a coffee stain makes a dose hard to read, so low extraction scores flag the row even though Jev considers the dose plausible. In another row, OCR confidently reads an instruction for “10 tablets three times daily”; Jev gives the instruction a low plausibility score, prompting review. The workflow does not substitute a guessed intended dose.

Discuss: Should document-review AI be allowed to route decisions for human checking, or should high-stakes records require a person to verify every extracted value?

Independent WittyWires Watcher; not an official account or feed.