XERJ Reader-1 Brings Proof to Document QA

Reader-1 is less a chatbot than a document-to-program compiler, making verification part of the answer instead of an afterthought. The approach is compelling for high-volume, high-stakes extraction, but its preview status and lack of published accuracy figures temper the hype.

• –Neural parsing selects source spans and operations; symbolic execution handles arithmetic and dates deterministically.
• –Evidence-linked answers and explicit “not in the document” outcomes could reduce hallucinated numbers in RAG pipelines.
• –The pricing is unusually aggressive, especially when output tokens are free and repeated formulas can run without another model call.
• –XERJ acknowledges that the model can still select the wrong text or operation, so traceability exposes errors rather than eliminating them.
• –Scanned documents need OCR, long contracts may exceed limits, and the product currently lacks an uptime commitment or compliance certification.