Nov 29, 2025
Accept/Reject Preference Learning
How TypeOS captures micro-level accept/reject decisions inside real documents to build the world's first fine-grained writing preference dataset.
Research
Advancing human taste, formatting, and agentic writing intelligence.
50,000+Documents in TypeBench corpus
99.98%AI detection accuracy (TypeBench 2025)
94%Essay grading correlation with human raters
3Published research benchmarks
TypeOS research is conducted by the TypeOS Research team. Benchmarks are developed from real document editing behavior and evaluated against human expert panels. All methodology details are published in full in each paper.
Nov 15, 2025
TypeBench: Measuring Human Taste
TypeBench is the primary evaluation corpus for TypeOS AI tools. Built from 50,000+ real documents, it validates AI detection accuracy (99.98%), essay grading correlation with human raters (r = 0.94), and writing taste alignment. The benchmark spans academic essays, blog posts, reports, and creative writing generated by 7 major AI models.
Oct 23, 2025
FormatBench: Structural Fidelity
Evaluating formatting accuracy in AI writing: APA, MLA, Chicago, legal contracts, business documents, and structural fidelity.