Definition
Plain language
Software that reads text out of an image, like recognizing the words in a photo of a sign.
As stated in the literature
Optical Character Recognition — algorithms that extract text from pixel data, used as a constrained verifier channel in evidence-carrying agent architectures.
Why it matters: OCR turns visual interfaces into verifiable text, which lets agents and verifiers reason about screens they can only see as pixels.
For example, an agent uses OCR to read the order total from a screenshot of a checkout page.
Heard on the show
“An OCR engine that runs on actual pixels.”Episode 062 — Treating Hallucinations as Exploits: A Gate-Based Architecture for Agent Safety