How XERJ launched XERJ Reader-1 on X
Answers document questions with verifiable programs that cite exact source words
What XERJ launched

I've built a new AI model that is not allowed to generate text. Instead, it produces a verifiable program that answers your prompt using the original reference. XERJ Reader-1 is a neurosymbolic model for working with data. It points to the exact words in your document, writes a https://t.co/JXHkFm0Om8
Transcript
How LaunchVideo.com rated the launch
LaunchVideo.com rating
About the launch
XERJ Reader-1 is a hosted model that answers questions over a document by writing a small program over the exact words in the source, then letting a separate engine compute the value. It opened in preview, by waitlist, on 12 November 2026 at platform.xerj.ai, for teams that have to stand behind a figure, including invoice totals and due dates, contract notice periods, email triage, operations logs, and payments checks. The launch is aimed at agent and back-office work where someone has to trace a figure back to the page. Ivan Novikov, founder and lead maintainer of XERJ, posted the announcement. Reader-1 is separate from Query-1, the model that ships inside XERJ, the Rust search engine he maintains for coding agents. On Reader-1 the neural part writes only the formula, pointing at character spans in the document. The symbolic part runs that formula with decimal numbers and real calendars, and the API returns the program, the quoted words, and the offsets. A stored formula can be executed again later with no model involved. If the text does not contain the answer, the call declines. Preview pricing is $0.01 per million input tokens, counting the document plus the question, with output free. Approved accounts start with a $1.00 balance and no card. XERJ's own measurements put a question on a short document at about 1.2 to 1.6 seconds, and the project promises no particular speed during the preview. They have published no accuracy figures. A formula can still quote the wrong span or pick the wrong operation, a scanned page has to be turned into text first, and the model does not summarise, estimate, or bring in outside knowledge. Documents and questions sent to the API are not stored.
Tags
Questions
Frequently asked questions
The XERJ launch post reached 1.7M views on X, with 785 likes, 205 reposts, 236 comments, 813 bookmarks.
XERJ launched on X on November 12, 2026 with the message "Answers document questions with verifiable programs that cite exact source words".
XERJ Reader-1 is a hosted model that answers questions over a document by writing a small program over the exact words in the source, then letting a separate engine compute the value.
LaunchVideo.com rated the XERJ launch 3.20 out of 5: hook 2/5, editing / creativity 2/5, copy 4/5, sentiment of launch 4/5, distribution strategy 4/5.





