Forms · Level 2 Parse: second passes over Mistral OCR

A second pass that corrects Mistral OCR 4.1's reading of a form page before it is stored. Each row is one model and scope over the same Mistral readings.

Key takeaways

• An Opus 5.5 check lifts Mistral from 634 to 681 of 707 facts across 15 form pages.
• GPT-6 Luna adds 12 facts at about 3% of the check's cost, mostly boxes, choices, initials and short handwriting.
• On dense hand-printed tax forms Luna gains little and fails one page; Qwen3.8 Flash matches its gain on fewer pages, fails four and takes about 164 s a page.

The scoreboard

How it was measured

Every second pass reads Mistral's reading of the page image and the page itself, and returns only the parts it changed. Each change is checked: a part outside the pass's scope, or a field that is no longer a label and a value, is rejected. Readings are scored as on Level 1, fact by fact against each page's truth, by the same grader. Pages 09 to 15 are held out: they were never used to shape the prompt or the flag rules.

Log