GPT-5.1 Checkbox Hallucinations: A Document Extraction War Story Stuck at 80% Accuracy

ComfortableThought55 · reddit · 2026-11-20

A developer details building multi-form document extraction on GPT-5.1: a generic prompt with per-form runtime specs, plus a fix for checkbox hallucinations caused by image downscaling — using Tesseract OCR to locate the checkbox page, cropping and sharpening that region into an extra image. In 10 days of shadow production, accuracy sits at 80%; OCR misses the page in 30% of runs, and prompt tuning remains painfully non-deterministic. They ask for advice on reliable checkbox reading and OCR-free page detection.

Original post →

More from coding & agent

coding & agent channel →