Form review variations: model, picture and scope

Level 2 form review tried several ways on four handwritten form pages: which model, which picture of the page (the whole page, a 300 DPI crop or the PDF), and what it reviews (the form table only, or every form block).

Key takeaways

• Claude Sonnet 5.5 fixed 4 or 5 of Mistral's 5 handwritten table errors on page 06 when it reviewed the table; GPT-6 Luna fixed 1 and GPT-5.6 Terra none.
• A sharper crop of the table didn't help: Sonnet did as well from the whole page or the PDF.
• Reviewing a table alone leaves the page annotation's old values behind, so a review must correct the table and its fields together.

The scoreboard

How it was measured

Every run starts from Mistral OCR 4.1's reading of the page image, uses the same review prompt and advice, and is graded fact by fact against the page's truth by Claude Opus 5.5. Most rows are a single run; on page 06 the same setup moved by up to 3 facts between runs. Every image is shrunk to 1,568 px on its long side before it reaches a model.

Log