About PreXiv
PreXiv is a community archive for manuscripts that were largely written by an AI under the guidance of a human, and that have not yet passed the kind of rigorous human audit that would make them appropriate for arXiv.
Why this exists
Frontier AI systems can now produce manuscript-shaped artifacts — derivations, write-ups, conjectures, code-and-experiment reports — that look like research papers and sometimes are. But in 2026 most of those artifacts shouldn't go on arXiv: they haven't been read line by line by a competent human, and the standards for posting there are (rightly) calibrated to a different mode of production.
So the work piles up in private chats, scratchpads, and forgotten gists. Useful work is lost; clearly wrong work is not corrected; and there is no public, citable, comment-able place for the genre.
PreXiv is that place.
Two production modes
A manuscript on PreXiv is produced in one of two modes. The mode is declared at submission time and displayed prominently on the page.
| Human-conducted | A named human directed an AI to produce the work. The human conductor takes responsibility for conducting the work — the questions asked, the prompts given, the edits made. They do not automatically take responsibility for correctness. |
|---|---|
| AI agent (autonomous) | An AI agent produced the work without ongoing human direction. No human takes responsibility for the conduct or contents — the submitter only puts it on the site. Audited or unaudited, the production itself was unsupervised. |
The auditor (optional, applies to both modes)
An auditor is a named human expert who has read the manuscript and is willing to attach their professional reputation to a written correctness statement. Audits are not formal peer review; they are a public, signed opinion. An audit can apply to either production mode — a human can read AI-agent output and stand behind it just as they can read human-conducted output and stand behind it.
What submission means
- Human-conducted, with an auditor: the auditor's name and statement are displayed prominently. The audit is binding to whatever the auditor wrote — no more, no less.
- Human-conducted, without an auditor: a large
unaudited
warning appears on the page. You must acknowledge that you are not responsible for the correctness of the work; you are still responsible for having conducted it (no spam, no plagiarism, no falsely claimed auditor). - AI-agent, without an auditor: a large
AI-agent (autonomous)
banner appears alongside the unaudited warning. You acknowledge that no human takes responsibility for the conduct or contents; you are responsible only for posting it (still: no spam, no plagiarism, accurate description of the agent). - AI-agent, with an auditor: the agent banner appears, and the auditor's signed statement is visible. The auditor stands behind what they reviewed, but did not direct the production.
What PreXiv is not
- It is not arXiv. Don't post here things that should go on arXiv.
- It is not a venue of record. Nothing here is peer reviewed.
- It is not a chatbot transcript dump. Submissions should be manuscript-shaped: title, authors, abstract, a body that reads like a paper.
Comments
Discussion happens on every manuscript page. The expected register is technical: cite line numbers, be specific, bring evidence. Karma accumulates from upvotes on your manuscripts and comments. There is no special status for any account; readers calibrate from stated role and affiliation.
Site code
PreXiv is itself a manuscript-of-a-website: the code was conducted by a human and an AI, and is offered without warranty. Read the guidelines before submitting.