Before you start
Settle one question before you install anything. May you send a manuscript you were asked to referee to a language model at all?
Your venue’s rules come first
Section titled “Your venue’s rules come first”refereekit sends manuscript text to a model backend. Whether you may do that is your venue’s decision, not the tool’s. The venues that have written a rule down do not agree. Two examples follow, quoted from each venue’s own page as of 2026-08-15.
NeurIPS 2025 prohibits it. Its policy for reviewers is at
https://neurips.cc/Conferences/2025/LLM. It says: “You must
keep everything relating to the review process confidential. … Do not
talk about or share submissions with anyone or any LLMs.” Zero-retention
terms make no exception: a model that keeps nothing has still been shared
with. Use the venue’s own review
interface.
ICLR 2026 permits it with disclosure. Its reviewer guide, at
https://iclr.cc/Conferences/2026/ReviewerGuide, says that “The use of
LLMs is allowed as a general-purpose writing assistance tool”. It goes on:
“we mandate that reviewers disclose the use of LLMs in their reviews. The
review form will include a field to specify how you used LLMs, if at all.”
Write that field yourself, in your own words.
Rules change each year, and most venues have written nothing. Where a venue is silent, the judgment is yours. Read its current reviewer page before each review.
refereekit knows one prohibition and cannot discover others
Section titled “refereekit knows one prohibition and cannot discover others”Before any step sends the manuscript to a model, refereekit checks the
venue. This is the venue gate. It reads refereekit’s list of venues that
forbid this, which has exactly one built-in entry, NeurIPS. Every venue
the list does not name is permitted. The gate acts only on a venue you
named, so a NeurIPS paper started without its venue is sent. It keeps
you from forgetting a rule you know, and never finds one for you. Extend
the list in the file venue-policy.toml in your reviews folder. The venue gate
has the details.
Confidentiality is your obligation
Section titled “Confidentiality is your obligation”The manuscript goes to one place: the model backend you chose. You also attest that this backend keeps nothing: zero retention. refereekit cannot check it. The attestation says what it means for each kind of backend.
Each review lives under work/ in your reviews folder. The folder’s
.gitignore keeps work/ out of a Git repository. That is a rule about
paths, not a check of what a file holds; see
Repository rules.
Author responses count as manuscript text, because a rebuttal quotes the paper. A venue that forbids sending the paper forbids sending the response too.
What refereekit does not do
Section titled “What refereekit does not do”- It does not judge the paper. You type your verdict in your private view, and the drafts are written from it. Nothing in refereekit recommends accepting or rejecting a paper.
- It does not fill in scores. On an OpenReview form, every field with fixed choices, such as a rating or a confidence, stays blank for you. A check of quoted words cannot justify a 3 over a 4.
- It does not check that a statement is true. A PASS means the quoted words occur on the named page, and nothing more. What verification means says how narrow that is.
- It does not refuse a draft with nothing behind it. The report is
drafted even when no checked evidence backs it. Every citation the draft
makes outside the checked evidence is flagged. So a draft with no flags
may simply cite nothing. The private view then shows a line that begins
EMPTY POOL. - It does not submit. You paste the drafts into the venue’s system yourself.