A user sharing the ParseBench announcement describes tests on 2,000 pages of real enterprise documents, using 169,011 deterministic rules rather than a language model as judge. Their summary says scores average five dimensions without weighting: table structure, chart extraction, content faithfulness, formatting semantics and visual grounding. It puts LlamaParse Agentic Plus in the lead at 90.2 and identifies chart extraction as the biggest divide between methods, with more than 20 methods scoring zero on that dimension.