Research paper integrity

This skill is the index. Each area below lives in its own reference file so you only load what the current task needs. Read the file before acting on that area, not after, because these rules change what you write rather than how you check it afterwards.

What lives where

scripts/check_prose.py is a fast mechanical pass for em dashes, banned filler, hedge stacking, and suspicious bib entries. It catches the boring violations so your attention goes to the ones that need judgment. Per the user's global rule, do not run it yourself: write it into the relevant workspace if needed and hand over the exact command.

The three commitments

Everything in this skill reduces to three things the user cares about. Keep them in mind even when you have only loaded this index.

1. The text contains no em dashes and no filler. Not because em dashes are wrong in general, but because in this user's papers they signal a sentence that was never restructured, and filler signals a claim that was never sharpened. Replace an em dash with a period, a colon, a comma, or parentheses, whichever the actual logic calls for. Details and the filler list are in references/writing-style.md.

2. Every sentence has been asked twice whether it needs to exist. First pass: does this sentence carry information the reader does not already have? Second pass: if it does, is this the shortest honest form of it? A paragraph that survives both passes is usually a third shorter than the draft. Cutting is the main editing operation here, not rewording.

3. Nothing is asserted that has not been checked, and no reference is cited that has not been confirmed to exist. Fabricated citations and confidently wrong numbers are the two failure modes that end a paper's credibility, and both look completely normal in a draft. Treat every number, every "prior work shows", and every bib entry as unverified until you have traced it. Procedure in references/hallucination-audit.md.

Default workflow when handed a draft

Adapt this, do not perform it mechanically. If the user asked only for a citation check, do only that.

1. Read the draft fully before changing anything. Local edits made without the whole argument in mind are how papers lose their thread.
2. Run the hallucination audit first. There is no point polishing a sentence built on a claim that turns out to be unsupported, and finding a broken citation early changes what the surrounding paragraph should say.
3. Then the prose pass: cut, de-fluff, remove em dashes, tighten.
4. Then structure: does each section do the job references/conference-writing.md describes, in the order a reviewer reads.
5. Report what you changed and, separately, what you could not verify. The second list matters more than the first. Never quietly leave an unverifiable claim in place without flagging it.

How to report

Give the user the edited text plus a short audit section. Something like:

## Changes
- Cut 4 sentences from the intro that restated the abstract
- Removed 11 em dashes, restructured the 3 sentences that needed it
- Section 4.2: "significantly improves" -> stated the actual delta and the test

## Could not verify (needs your call)
- [Smith 2021] claimed for the pretraining result: no paper by that name found in
  ACL Anthology or arXiv. Either the citation is wrong or the claim needs a source.
- Table 2 baseline number differs from the number quoted in Section 5.

Flagging honestly is the whole value here. A silent pass that leaves a fabricated citation in place is worse than no pass at all, because it launders the error.