Reading order is the part of PDF parsing that quietly breaks downstream NLP

If you have ever run NER or a summarizer over text extracted from a two-column paper, you have seen it: the extractor reads straight across both columns, and every sentence ends up as halves of two different sentences. Tables flatten into word soup, formulas turn into stray symbols, and running headers land in the middle of paragraphs. Nothing downstream can repair that. We have spent a good while on this and just released the code and weights for a 1.3B document parsing model, OXR from SoMarkAI. It takes PDFs or page images and returns Markdown or JSON where every element is typed (13 types,