FactScanAI OCR Accuracy Benchmark 2026

A reproducible controlled OCR benchmark measuring how image blur and resolution affect OCR accuracy and failure modes. The study contains 60 synthetic documents, 900 OCR executions, English and Czech content, and four document categories. It evaluates character error rate (CER), word error rate (WER), structured-document content metrics, and OCR failure modes under controlled blur and resolution degradation. This Zenodo record archives version 1.0.0 of the public benchmark dataset and reproducibility materials. Canonical research report:https://factscanai.app/research/ocr-accuracy-benchmark Reproducibility repository:https://github.com/FactScanAI/factscan-ocr-benchmark Key limitations: synthetic documents, one OCR provider, English and Czech only, controlled API-level testing, and Gaussian blur rather than real-world camera motion blur.