{"name":"forms-level-2","title":"Forms · Level 2 Parse","tables":[{"heading":"Fifteen form pages, 707 facts (seven held out)","columns":["Pages","Facts","Mistral alone","Opus 5.5 check","GPT-6 Luna","Qwen3.8 Flash"],"rows":[["01–08, the pages the advice and flags were built from","386","369","384","372 (378 with figures)","244 of 250 on 6 pages (Mistral 241, Luna 244); 2 failed"],["09–12, held out","54","48","53","51","50"],["13 · Form 1040, page 1, hand print","91","71","82","72","failed: 3 failed calls"],["14 · Form 1040, page 2, hand print","69","60","66","failed: stopped at 32,000 output tokens","failed: an unreadable reply"],["15 · Schedule E, hand print","107","86","96","91","93"],["All 15","707","634","681","+12 on 14 pages","+12 on 11 pages"]],"note":"Luna and Qwen review every form block; figures are left as Mistral described them. The flag rules alone mark 19% of the parts (42% of the text), and those parts hold 56 of Mistral's 73 errors."},{"heading":"Eight form pages, 386 facts: model, scope and advice","columns":["Second pass","What it reviews","Advice file","All 386","Gain","Values 230","Choices 29","Boxes 116","Signatures 8","Drawings 3","Wrong","Missing","Time / page","Cost / page"],"rows":[["Opus 5.5 check","whole page","no","384","+15","228","29","116","8","3","2","0","+16 s","+$0.058"],["GPT-6 Luna","form blocks","yes","378","+9","222","29","116","8","3","6","2","+17 s","+$0.0015"],["GPT-6 Luna","form blocks","no","378","+9","222","29","116","8","3","6","2","+21 s","+$0.0017"],["GPT-6 Luna","whole page","yes","378","+9","222","29","116","8","3","6","2","+23 s","+$0.0020"],["GPT-6 Luna","whole page","no","378","+9","222","29","116","8","3","6","2","+16 s","+$0.0013"],["GPT-6 Luna","flagged form blocks","yes","377","+8","222","29","115","8","3","5","4","+17 s","+$0.0013"],["GPT-6 Luna","flagged form blocks","no","376","+7","222","29","116","8","1","4","6","+21 s","+$0.0014"],["Mistral OCR 4.1 alone","none","","369","","219","28","116","6","0","11","6","",""]],"note":"Measured before figures left the forms scope. Time and cost are the second pass only; Mistral adds about 22 s and $0.005 a page."},{"heading":"Qwen3.8 Flash against GPT-6 Luna: the 6 pages Qwen finished","columns":["Second pass","6 pages 283 facts","Wrong","Missing","Time / page","Cost / page","Pages it failed"],"rows":[["Opus 5.5 check","283","0","0","+15 s","+$0.052","none"],["GPT-6 Luna","280","1","2","+17 s","+$0.0014","none"],["Qwen3.8 Flash","279","2","2","+119 s","+$0.0045","IRS 4506-C and Patriot Act ID: timed out, 3 attempts each"],["Mistral OCR 4.1 alone","273","","","","",""]]},{"heading":"Four held-out pages, 54 facts","columns":["Page","Mistral alone","Opus check","GPT-6 Luna","Qwen3.8 Flash"],"rows":[["Contact information: handwritten address, phone, email","7/8","8/8","8/8","7/8"],["Notary instructions: a ticked option, two sets of initials","4/6","6/6","6/6","6/6"],["Indemnity agreement, typed \"contract\" by Mistral","18/20","19/20","18/20 · skipped","18/20 · skipped"],["Signature affidavit, typed \"id_document\" by Mistral","19/20","20/20","19/20 · skipped","19/20 · skipped"],["All 4","48","53","51","50"]]}],"log":[{"date":"261004","line":"New board. On 15 form pages, Luna adds 12 facts to Mistral's 634 of 707; the Opus 5.5 check reaches 681."}],"updatedAt":"2026-10-04T01:04:15.861Z"}