Hallucination-guarded OCR for AI agents

The problem: modern OCR fails silently. A vision-language model reading blurry handwriting will not return an error - it returns fluent, confident text that is simply wrong, sometimes in the wrong language entirely. A human spots this instantly. An agent pipeline ingests it and acts on it.

We measured this on real documents: on one handwritten Japanese note, a frontier OCR model output fluent Devanagari (Hindi) - zero overlap with the document - while a traditional OCR engine read it correctly. On a printed tax receipt, the same model subtly rewrote a legal notice into different (wrong) wording. Neither failure raises an exception. Both poison downstream automation.

What Mart402 does about it

Every OCR result - POST /v1/ocr (single image, flat $0.05) and POST /v1/parse (PDF) - carries machine-actionable quality signals (consensus on both; ocr_quality on parse OCR-route results):

{"ok": true, "route": "mistral",
 "ocr_quality": {"confidence": 0.9, "hallucination_suspected": false,
                 "flags": [], "chars": 1483},
 "consensus": {"verdict": "high_consensus", "confidence": 0.86, "agreement": 0.91,
               "flags": [], "best_engine": "mistral"},
 "markdown": "..."}

Hallucination detection: script-mismatch (wrong writing system for the document language) and excessive repetition - the two dominant silent-failure modes - are flagged, and confidence drops to 0.15.
Dual-engine consensus: a second, independent engine with orthogonal failure modes (traditional OCR vs VLM) cross-checks the read. On detected hallucination the output is replaced with the verified read and flagged.
Honest verdicts: high_consensus, partial_consensus, low_consensus_review_recommended or all_engines_suspected_hallucination. When engines disagree we say so and show the agreement score. We do not sell overconfidence - an agent that gets burned once never comes back.

How an agent should consume this

Signal	Action
confidence ≥ 0.7, no flags	Use the text.
hallucination_suspected	Discard; use consensus best_engine text (already applied) or re-scan.
low_consensus_review_recommended / all_engines_suspected_hallucination	Queue for human review; don't act on extracted numbers.

Try it

Free sandbox (same API, Base Sepolia; OCR returns a recorded sample there): mart402.dev. Production, via x402: single images at a flat $0.05/image (POST /v1/ocr); PDFs via free quote first (/v1/parse/quote) - text-embedded flat $0.01/document, scanned OCR $0.015/page. Measured latency & success rate: /stats, also probed by Market402 (same operator, public rules).

Agent docs: /agents.md · /llms.txt · /catalog