Drex turns document parsing into per-page judgment: top Parse Index of 0.834, with correction rewriting only doubtful pages
Synopsis
Nace released Drex, a decision model for document processing that, after the parsing engine returns layout-aware Markdown, judges page by page whether text matches the page image, whether a table kept its shape and whether a ticked box survived, and triggers correction only where something is in doubt. On Nace's own parsing leaderboard, NDI plus Drex scores a Parse Index of 0.834, ahead of all 12 parsers tested, including LlamaParse Agentic at 0.820; it improves on the engine alone on OmniDocBench, MDPBench and ExtractBench, and is within noise of it on ParseBench and RealDocBench.
Interpretation
Document parsing gains a per-page judgment layer: the engine first reads a file into Markdown that keeps each block's place on the page, and Drex then checks whether the read text agrees with the page image and decides where a second look is needed. Earlier paths either used OCR, which recovers words but turns a table into a run of numbers and loses ticked boxes, or sent every page to a frontier model, which is slow and expensive at volume. On Nace's parsing leaderboard, NDI plus Drex scores a Parse Index of 0.834, first among the 12 parsers tested; NDI alone scores 0.820, LlamaParse Agentic 0.820 and GPT-6 Astra (low) 0.813.
The correction step rewrites only the parts of a page judged doubtful rather than the whole page, so a dense page no longer means a long rewrite. The speed comparison gives both approaches the same first reading of every page, so the only thing that changes is the correction step, which attributes the speed difference to correction scope. In median correction time per page, ParseBench falls from 15.6 s to 9.8 s, RealDocBench from 20.2 s to 12.5 s and MDPBench (English) from 26.7 s to 19.1 s; on the slowest 10% of pages the gains are 55%, 57% and 53%.
Drex answers narrow questions, such as whether a line of OCR text matches the image, with a calibrated probability, which lets the engine check every page and spend heavy reading only where something is in doubt. At about 50 milliseconds and about $2 per million answers, per-page checking becomes affordable across a whole pipeline. On phone photos of documents it scores 91.9 on MDPBench, 8.3 points ahead of the next best parser, and 95.7 on OmniDocBench, 5.4 points ahead of the next best parser.
Perspective
The demonstrated scope is a comparison of 13 parsers on Nace's own parsing leaderboard, per-benchmark results on six benchmarks, and correction-time comparisons on ParseBench, RealDocBench and MDPBench (English); the timing figures are median correction time per page averaged over two runs.
On ParseBench and RealDocBench the gap between NDI plus Drex and NDI alone is within noise; the Parse Index and leaderboard are maintained by the releasing party, and costs use each provider's list price.