{"name":"form-review-variations","title":"Form review variations","tables":[{"heading":"Page 06 · USA Patriot Act borrower ID, 47 facts (Mistral alone 40)","columns":["Model","Picture of the page","Reviews","Runs","Score of 47, each run","Table errors fixed of 5","Broke","Time / run","Cost / run"],"rows":[["Claude Sonnet 5.5","whole page, 150 DPI","ID table","4","44 · 44 · 44 · 45","4 · 4 · 4 · 5","0","9–13 s","$0.019–0.023"],["Claude Sonnet 5.5","whole page, 150 DPI","every form block","4","42 · 44 · 42 · 45","2 · 4 · 2 · 5","0","12–16 s","$0.038–0.042"],["Claude Sonnet 5.5","PDF (the 300 DPI scan)","ID table","3","40 graded","4 · 5 by hand","0","8–12 s","$0.016–0.021"],["Claude Sonnet 5.5","crop of the table, 300 DPI","ID table","1","43","3","0","7 s","$0.013"],["GPT-6 Luna","whole page, 150 DPI","every form block","1","42","1","0","11 s","$0.0014"],["GPT-6 Luna","crop of the table, 300 DPI","ID table","1","41","1","0","9 s","$0.0006"],["GPT-5.6 Terra","crop of the table, 300 DPI","ID table","1","39","0","1","5 s","$0.0079"],["Opus 5.5 check for reference","whole page","whole page","1","45","3","0","61 s","$0.075"],["Mistral OCR 4.1 alone","","","","40","","","","$0.005"]],"note":"Table errors: the 5 Mistral errors in the handwritten Primary ID table. The PDF run graded 40 because the grader quoted the annotation's old fields; printing the table from two more PDF runs showed 4 and then all 5 fixed."},{"heading":"Pages 13–15 · hand-printed tax forms, held out","columns":["Page","Facts","Mistral alone","Luna, every form block","Qwen3.8 Flash, every form block","Sonnet 5.5, table, image","Sonnet 5.5, table, PDF","Opus 5.5 check"],"rows":[["13 · Form 1040, page 1","91","71","72","failed","76","79","82"],["14 · Form 1040, page 2","69","60","failed","failed","64","","66"],["15 · Schedule E, rental income","107","86","91","93","89 broke 3","","96"]]},{"heading":"Time and cost per page","columns":["Time and cost per page","Luna","Qwen3.8 Flash","Sonnet 5.5, image","Sonnet 5.5, PDF","Opus check incl. Mistral"],"rows":[["13 · Form 1040, page 1","50 s · $0.004","","21 s · $0.046","23 s · $0.047","57 s · $0.106"],["14 · Form 1040, page 2","","","24 s · $0.062","","45 s · $0.105"],["15 · Schedule E","43 s · $0.003","158 s · $0.008","27 s · $0.052","","31 s · $0.081"]]}],"log":[{"date":"261004","line":"New board. Sonnet 5.5 fixes 4–5 of the 5 handwritten table errors on page 06; Luna 1, Terra 0."}],"updatedAt":"2026-10-04T01:04:16.263Z"}