This is Yardstick Lab's first live category. The current ranking covers 72 public documents. It scores five specialized extraction platforms alongside three frontier LLMs called directly with the same schema — the do-it-yourself baseline a team would build in-house. One row each, at its best-scoring config.
Macro-average field accuracy from the current ranked dataset.
The leaderboard shows one config per system. This is the full set, including each product's standard and premium settings side by side.
| System | Config | Score |
|---|---|---|
 DocuPipe | High | 97.02% |
| DocuPipe | Standard | 96.14% |
 Claude Sonnet 5 | Direct LLM | 91.73% |
 Reducto | Deep Extract | 89.38% |
| Reducto | Standard | 81.11% |
 Extend | Default | 80.28% |
 GPT-5.5 | Direct LLM | 76.48% |
 Gemini 3.5 Flash | Direct LLM | 72.98% |
 Pulse AI | pulse-ultra-2 + extended reasoning | 70.95% |
 Unstructured | Auto partitioning | 67.67% |
These slices come from the same current run and show evidence for specific buyer workflows.
Hebrew and Arabic layouts, where column order and text direction have to be resolved before a single character is read.
Dense CJK scans, including vertical text and microfilmed sources.
