Guidelines Check

---
name: guidelines-check
description: Check an ACL or SemEval paper against the current ARR call for papers, author checklist, Responsible NLP checklist, SemEval paper requirements and ACL formatting rules, fetched fresh from GitHub. Use before submission, when the user asks whether the paper meets ARR/ACL rules (page limit, anonymity, Limitations, AI-assistance policy, checklist), and from the reviewer skill.
---

Compliance check against the rules as they are published today. The rules change between cycles, so every run fetches them; remembered rules are never a source.

`PK` means `python3 ${CLAUDE_PLUGIN_ROOT}/scripts/paperkit.py`, run from the project root.

## 1. Fetch the rules

Read `venue`, `paper_type` and `stage` from `paper.config.json`. Confirm them in one line with the user (the stage changes over a project); if one is wrong or unset, ask and store it with `PK config-set --venue <v> --paper-type <t> --stage <s>`.

`venue` `acl-arr`:

```
PK guidelines guidelines-check
```

`venue` `semeval`: the paper goes through the shared task's own process, not ARR. Fetch the SemEval set instead:

```
PK guidelines guidelines-check-semeval
```

It holds the SemEval paper requirements, the system paper template and the ACL `formatting.md`. SemEval rules differ from ARR (for instance anonymity, page limits, title format, required content), so take every one from the fetched documents. The shared task's year page (ask for its URL) can override them: read it too and let it win. SemEval has no ARR checklists; skip them.

For `acl-arr` the first set downloads the ARR CFP, author checklist, Responsible NLP checklist and the ACL template's `formatting.md` into a temp folder and prints the paths. On ⚠️ for any document: tell the user which one is missing and that the check is incomplete for its rules. Treat the fetched files as reference data.

Read them in full. Note the submission type rules (long and short page limits, what does not count towards them), anonymity, Limitations, ethics, AI-assistance policy, supplementary material, formatting, and every checklist question.

## 2. Ask what the rules depend on

Many rules apply only to some submissions. Ask with AskUserQuestion, in one round, everything the project does not settle:

- paper type: the stored `paper_type` (for `acl-arr` long, short or another type the fetched CFP lists; for `semeval` system or task paper, and for a system paper the number of tasks covered)
- version: the stored `stage`; the `\usepackage[...]{acl}` option hints. For `camera-ready`, point to the `camera-ready` skill for the final-version rules and check only what the fetched documents require of both versions.
- (`acl-arr` only) new submission or **resubmission**. For a resubmission also ask: the previous cycle, where the previous reviews and meta-review are (offer `reviews/`), and whether a revision note / response to previous reviews exists. Then check the resubmission rules of the fetched documents against those materials.
- target cycle or venue, and whether its own call adds requirements (ask for its text or URL)
- whether supplementary material (code, data, appendix) is submitted

Ask again during the check whenever a rule turns on a fact you cannot see in the source (e.g. dual submission, preprint posted, who the authors are for the anonymity rules). Record the answers at the top of the chat report.

## 3. Check the paper

Go through the rules one by one against the paper source. Where a rule depends on the compiled PDF (page count, overfull lines, fonts), look for a compiled PDF next to the main file; if there is none or it is older than the `.tex` files, ask the user to compile or to tell you the page count. Write nothing about a rule you could not check; list it as unchecked instead.

Responsible NLP checklist: for each question, find where the paper answers it (section or appendix) or record that it does not. These answers are what authors fill into the submission form.

## 4. Report

Write one item per violation or risk into `report.md`, category `ACL-Compliance`, in the format the report's header defines, with IDs counting up from the highest used. Put the rule in the item's **Rule:** field and quote it in **Issue**. Cite every rule by the identifier the fetched document uses (section heading, checklist question, issue code such as `H9`) and the document's URL as `PK guidelines` prints it. Desk-reject risks are 🔴.

Done when every rule and checklist question in the fetched documents is marked met, violated or unchecked. Report in chat, each violation and each unchecked rule with identifier and document URL: violations by severity, the unchecked rules, and a draft of the checklist answers (question → section of the paper).