Reviewing for a journal

A journal review runs the same steps as Your first review, start to finish. This page covers only what differs when the paper is real and the venue is a journal.

Before you start

Settle three things before you start, so none of them stops you halfway.

• Your journal’s rule. Read what it says about sending a submission to a language model. refereekit knows the rules of very few venues, so it cannot tell you your journal’s. If the journal forbids it, add the journal to venue-policy.toml in your reviews folder. Before you start says why.
• Model access. Choose your deployment and enter its key in model access, never in the chat.
• The attestation. Declare, in the same place, that your account keeps nothing. Without it, every step that would send the paper refuses. The attestation says what you assert on each deployment.

Start the session

Name the journal when you start:

You: Start a review of paper.pdf for PRX.

The agent starts a session and gets back its reference and the page count. The session keeps a copy of the PDF and records the journal, so you never name it again.

The journal does two jobs. It runs the venue gate before anything is sent. And it picks which of your venue notes the drafts recall. A note you stored while reviewing an earlier PRX paper shapes this draft too.

Use one session per paper. A session is bound to the PDF it started from. If the authors send a revision, ask the agent to compare the session with it. The revision is stored beside the original (see Comparing a revision). A second paper gets a session of its own.

The run

Open the review, and ask your questions as in the practice run. The difference is that a real model answers now, so think over each question before you send it. There is no automatic summary. Ask for one if you want it.

Before your first real run, try one refusal that needs no key. On the tutorial session, before you make the attestation, ask the agent to draft the report again. The private view shows:

refusing to send: backend is not marked zero_retention

That is the gate working: nothing was sent.

Read the flags

The review ends with the triage list: every citation that did not check out, grouped by reason. A draft can also carry flags. Each flag gives one of two reasons.

• not in verified pool: the draft cites a page, equation or figure your questions never established. Ask about it, or cut the sentence.
• failed re-verification: the draft’s words for a page do not match that page. Check the sentence against the PDF yourself.

A draft with no flags and an empty claim pool cited nothing at all. The private view marks that case with a line that begins EMPTY POOL. The two draft flags has both in full.

Edit the drafts

The report and the letter are starting points. Edit them in /refx-drafts in pi, or with the window’s “Open it in your editor”. Once you have edited one, a new draft asks you first whether to replace it. Your edited copy is kept either way.

Rather than edit by hand, you can ask for one part again, such as a shorter summary. If the drafted voice is wrong, too hedged or too terse, fix the style guide instead (see Your voice).

Submit, then close

Ask the agent to check and finish, as in the practice run. Paste the drafts into the journal’s own system yourself. Ask the agent to delete the session when the journal’s terms say so.

Afterward

Store a short venue note in your own words, about how you weighed this paper for this journal:

PRX: lean accept-after-major on approximate-but-validated theory

Every later draft for PRX recalls it. A note that repeats the paper’s words is refused (see Your voice).