Paper review

Paper Review

Overview

A pre-submission novelty + reviewer pass that grounds every judgment in a retrieved paper. Naive LLM review asserts novelty from memory and invents citations; this skill retrieves real prior work, quotes it, and cross-checks the draft's bibliography — then surfaces (without adjudicating) any overlap with a lab or co-author line that has to be disclosed.

Core principle: a novelty claim is only as good as the paper it is checked against. Every verdict cites a record retrieved this run with a verbatim snippet, or it is marked unverified — the analogue of techne:paper's "never hand-type a result number".

When to use

• A first-author draft is approaching submission and needs a distinctness + related-work check.
• "Review <name> for novelty", "what am I missing", "is C2 actually new".
• NOT for copy-editing prose; NOT for a paper with no stated contributions yet (scaffold first with techne:paper).

Inputs

• <name> — the paper directory under papers/ (e.g. velocity-fl-systems).
• --from <repo> (optional) — the source code repo; grounds quantitative claims against its harvest output.

Config

Read ## paper-review from <repo>/.claude/skill-context.md: lab_line (overlap source of truth, default LINEAGE.md), retrieval (default OpenAlex + web; Semantic Scholar only if S2_API_KEY is set), mailto (OpenAlex polite pool, default git config user.email), report (default papers/<name>/novelty-review.md). Sensible defaults if absent.

Procedure

Refuse if papers/<name>/ does not exist. Locate main.tex and its \input files. Then run a mandatory citation-integrity gate, then OpenNovelty's four-phase pipeline (arXiv 2601.01576):

0. Citation integrity (MANDATORY — do this first). Before assessing novelty, verify the draft's own citations are real. Extract every \cite key and its references.bib entry; verify each against an authoritative record — arXiv ids via the arXiv API, DOIs / venues via OpenAlex or DOI content-negotiation. See citation-verify. Confirm the id/DOI resolves to the same paper (title match), with correct authors and year, and read the abstract to check any claim the draft makes about that work ("parameter-free", "first to"). A fabricated id, a mismatched title, or an unsupported claim-about-a-paper is a stop-ship finding — fix or flag it before anything else. A single wrong citation sinks the paper.

1. Extract claims. Parse main.tex for the title, abstract, and stated contributions (the % (C1) … or \item list after "Our contributions" / "In this paper"). Collect flag-claims: first, novel, outperforms, state-of-the-art, and any N× / N% number. Show the extracted contribution list to the user to confirm or edit before retrieving — the claims drive everything downstream.

2. Retrieve. Per contribution, build a focused query from its noun phrases and search OpenAlex; collect {title, year, DOI, citations, abstract}. Rank by relevance, then citation count; web-search fills gaps. Log every query. See retrieval recipes. Only derived keyword queries leave the machine — never the paper text (drafts are private, often double-blind).

3. Compare — each contribution against its top candidates. Ground every judgment in a verbatim snippet from the retrieved abstract. Classify: novel / incremental / overlaps-with [cite]. A verdict with no retrieved record + snippet is unverified, never a confident assertion.

4. Synthesize into the configured report from the template:

   • §0 Citation integrity — each cited work → verified (source) / mismatch-fixed / stop-ship; stop-ship findings first. No fabricated or mischaracterized citations.
   • §1 Novelty — per contribution: verdict + closest prior work (cited + snippet) + what is distinct.
   • §2 Related-work gaps — retrieved papers absent from references.bib (cross-check DOIs and titles per the recipe). The closely-related work the author overlooked.
   • §3 Claim-support — each flag-claim → supported / unsupported / overclaim. First-ness from retrieval; quantitative claims against the --from repo's harvest output (else flag "verify against source").
   • §4 Lab-overlap — surface, do not adjudicate. Parse lab_line; list adjacent lab papers and which draft claims overlap; emit the COPE disclosure checklist. Never render a too-close / go-no-go verdict — that boundary is the author's and advisor's to agree.
   • § Provenance — every query and the records it returned, re-runnable.

Report the path + a one-line summary (N contributions, M novel / K overlaps, P uncited gaps, lab-overlap: needs-discussion). Never edit the paper — this pass is advisory.

Common mistakes

• Trusting the draft's existing citations. Verify them too (§0) — a seed bib or a recalled citation can be fabricated or mischaracterized. arXiv id → arXiv API; venue → OpenAlex / DOI; a claim about a paper → its abstract. A single wrong citation sinks the paper.
• Asserting novelty from memory. The failure this skill exists to stop. No verdict without a retrieved record + a quoted snippet logged in § Provenance.
• Inventing or half-remembering a citation. A fabricated reference is worse than none. If a search did not return it, do not cite it.
• Eyeballing the bibliography. §2 gaps come from parsing references.bib (DOIs + titles) and comparing to the retrieved set — not from guessing what is already cited.
• Adjudicating the lab boundary. §4 surfaces adjacency + the disclosure checklist and stops. Scoring it go/no-go usurps a human agreement.
• Relying on keyless Semantic Scholar. It returns 429. OpenAlex is primary; S2 only with a key.

Why this skill is careful

The output is one advisory artifact — novelty-review.md — and it never touches the draft. Its value is that every line can be checked: a cited paper, a quoted snippet, a logged query. Sibling of techne:paper, which scaffolds the draft this one reviews.