/evaluate

Scores a finished draft against what the target venue says it looks for. Each section gets a 1 to 10 score with the reason, the evidence, and the one change that would raise it most. The paper gets a weighted overall score. The report lands in paper-out/evaluation.md.

It scores. It does not fix: fixes go to /review <section> and /write <section>. It does not predict accept or reject, because that would be a guess presented as a measurement.

It is manual-only: a whole-paper read that writes a file is something you time yourself.

Usage

/evaluate                         # finds the manuscript, asks for the venue text if there is none saved
/evaluate main.tex

Before anything else

1. Check for paper-out/STATE.md and read it, with decisions.md and open-questions.md. Read references/session-state.md in paper-writing-style before writing any state.
2. Load the style rules. Invoke paper-writing-style. The project's CLAUDE.md outranks it.
3. Lock the language. Language of the user's first message, whole run.
4. Check the draft is complete. Follow the \input list in main.tex and read every section file. If any is empty or a placeholder, stop, name them, and point at /draft <section> or /write <section>. Do not score an outline.

Workflow

1. Get the venue criteria

Check for paper-out/venue-criteria.md.

• It exists: read it and say in one line what you loaded, for example Using venue-criteria.md (pasted 2026-11-02, 7 criteria, weighted), so a stale file is visible. The researcher does not paste again unless they say the text changed.
• It does not exist: ask the researcher to paste the venue text: the review criteria, call for papers, or reviewer guidelines. Without it, refuse and say why. Never score against criteria you remember for the venue.
• The researcher has new text: replace the pasted section, and append the replacement to decisions.md.

Save, extract and confirm the criteria as references/venue-criteria.md sets out. Show the criteria table with the sentence each came from, mark every weight the text did not give as unstated (equal weight, said so), and get confirmation before scoring. If the text names only paper-level criteria, ask how each maps to sections rather than inventing the mapping.

2. Read the exemplars

If exemplars/ holds papers, label each one's structure with the categorization method, keep what recurs, and record how many papers it recurred in (6 of 8). Structure only, never sentences or results. Save the recurring pattern in venue-criteria.md so a later /review can reuse it.

If exemplars/ is empty, continue, say plainly that the exemplar comparison is missing, and mark every cell that depends on it unscored. Never fill it from memory of what the venue's papers look like.

3. Score

Read the manuscript section by section and score as references/rubric.md sets out: 1 to 10 per criterion with the anchors, a quote from the draft as evidence, the exemplar count where one applies, and the single change that would raise the section most. Methods, related work and other sections with no /write model are scored on the venue's criteria and clarity only, never against an invented structure.

4. Compute and write

Compute the section scores and the weighted overall score exactly as the rubric defines them, so the reader can recompute them from the table. Write paper-out/evaluation.md from references/report-template.md, overwriting the previous one.

5. Update paper-out/

• evaluation.md: written.
• venue-criteria.md: created or updated.
• STATE.md: the overall score and the evaluation date, and the weakest section as the next step.
• decisions.md: one dated line with the overall score, and the previous one if there was one, since the report is overwritten.
• open-questions.md: each criterion that could not be scored, and what would unblock it.

Name every file you wrote. A run is not finished until this is done.

6. Hand off

Name the three lowest sections and the skill for each: /review <section> for a diagnosis, then /write <section> to apply it. Say once that /review updates that section's score in evaluation.md when it runs, so the report stays current without running /evaluate again.

Hard rules

• Never run without pasted venue criteria. Refuse, and ask for the paste.
• Never ask the researcher to paste again while venue-criteria.md exists and is readable.
• Never fill a criterion, weight or exemplar pattern from memory. Unknown or unscored stays marked.
• Never predict accept, reject or a decision. Report scores, reasons and the change that raises each.
• Never score a section that is empty or a placeholder. Stop and say which.
• Never copy an exemplar's sentences or results. Structure only.
• Never rewrite the researcher's text. This skill scores.
• Never give a score without its reason, evidence and criterion.
• Never finish a run without updating paper-out/. Say which files you changed.
• Never emit an em dash (---).