agentleFS
gh-xjgh-xj/agent-repo-kit/skills/paper-vetting/SKILL.md

paper-vetting

Skill0 starsChanged 7 months agoMITView on GitHub

Use BEFORE reading or summarizing any research paper. Triangulate trust across three independent lenses (team / citation context / claim-level evidence), then decide how skeptically to read it. Triggers: 'research this paper', 'read this paper', 'summarize this arxiv', 'what's the reputation of', 'is this paper credible', 'vet this paper', '评估这篇论文', '这篇论文靠谱吗', '这个团队厉害吗', any arxiv.org / openreview.net / aclanthology.org / proceedings URL handed in for review.

The file

We split this skill into the rules it gives an agent

5 rules. About 0 are common, 4 are distinctive, and 1 is very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈4

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 1

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Default: run Phases 1–4, present verdict, then ask before Phase 5.

    line 49
  2. Very specific:

    Use the cheapest tool that works (see references/sources.md tool ladder):

    line 53
  3. Distinctive:

    Read the paper with the calibrated posture.

    line 130
  4. Distinctive:

    Three lenses run independently: before triangulation. Do not let the team lens leak into the citation-context lens — that's anchoring.

    line 134
  5. Distinctive:

    The falsifier is mandatory: every verdict ends with what would change it.

    line 142

≈ “Common” is an estimate: at least 1 in 5 of the 350,495 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Paper Vetting

Vet a paper through three independent lenses, then triangulate. Output is a calibrated reading recommendation, not a single trust score. Reputation alone tells you how seriously to read; only the convergence of lenses tells you whether the claims survive.

Methodology-first. Same flow regardless of field, but the ML/Agent lens has its own pitfall list — see references/ml-agent-pitfalls.md.

Why three independent lenses

Single-lens vetting fails predictably:

  • Team-only → famous-name halo: senior teams write wrong papers and you anchor on credentials.
  • Citation-count-only → hype amplification: viral wrong papers and ceremonial citations both inflate the number.
  • Self-reported-rigor-only → checklist theater: authors check boxes without substance.

Independence prevents each failure mode. Mismatch between lenses is itself the signal — strong team + critical citation context + no released code = a different problem than weak team + replicated result + open code.

When to use

  • User hands over an arxiv / OpenReview / venue URL and asks to "research", "read", "summarize", or "vet" it.
  • User asks about a paper's team or credibility.
  • User is deciding whether to invest time in a paper.

When NOT to use

  • User wants you to implement a method from a paper they already trust → just read.
  • User wants a literature review across many papers → use a survey workflow.
  • The "paper" is a blog post — most lenses still apply, but venue/checklist signals will be empty.

Process

Phase 1: Identify        →  metadata via arXiv/OpenAlex (cheap), see sources.md tool ladder
Phase 2: Three lenses    →  run in parallel; each lens votes independently
   2a Team               →  position, m-quotient, field-fit, recency, Connected-Papers neighborhood
   2b Citation context   →  scite supporting/contrasting, Sem.Sch. highly-influential cites,
                            PubPeer flags, Retraction Watch
   2c Claim-level        →  reproducibility checklist (in-paper), code/data/weights, leaderboard,
                            for ML/Agent: contamination check (see ml-agent-pitfalls.md)
Phase 3: Triangulate     →  convergence → confident verdict; divergence → flag the mismatch
Phase 4: Verdict         →  trust band + reading posture + skepticism flags + falsifier
Phase 5: Read (opt-in)   →  read paper through the calibrated lens

Default: run Phases 1–4, present verdict, then ask before Phase 5. Skip vetting only on explicit "skip vetting".

Phase 1 — Identify

Use the cheapest tool that works (see references/sources.md tool ladder):

  • arXiv id → https://arxiv.org/abs/<id> (canonical) or arXiv API.
  • DOI → OpenAlex API (/works/doi:<doi>).
  • OpenReview → use the forum page; reviews + scores are gold.

Extract: title, abstract, version, year, all authors with affiliations, corresponding-author flags, venue (or "preprint"), and — for ML papers — whether a NeurIPS/ICML/ICLR/JMLR-style author-completed reproducibility checklist is included. The checklist is its own signal regardless of content.

Phase 2 — Three lenses (run in parallel)

Dispatch lens lookups concurrently. Cap each lens at 3–4 web calls; stop early when the picture is clear.

Lens A — Team (who)

For corresponding + visibly senior co-authors only (not all 20):

  • Position + institution + tenure status.
  • m-quotient = h-index / years-since-first-publication. Better than raw h for career-stage normalization. Round, don't fabricate.
  • Field-fit: is this paper in their published lane? (Connected Papers neighborhood reality-check.)
  • Recency: first/corresponding output in last 18 months.
  • Subfield reputation: invited talks, keynotes, named in conferences.
  • Practitioner credibility: code shipped, frameworks in use, production deployment.

Output: 2–3 sentences per author + a one-word characterization (heavyweight / solid / rising / unknown / tourist). Mark ? when signal is thin — do not invent.

Lens B — Citation context (how peers actually use it)

Citation count is volume; citation context is direction.

  • scite.ai Smart Citations — supporting / contrasting / mentioning ratio. <5% contrasting on a methods paper with >100 cites is a healthy sign; >20% contrasting is a red flag.
  • Semantic Scholar Highly-Influential Citations — citations where the citing paper built on this one, not just referenced it.
  • Connected Papers / ResearchRabbit — is this paper a node in the live conversation, or isolated?
  • PubPeer — search the title or DOI. Any post-publication concerns?
  • Retraction Watch — retracted? expression of concern?
  • Replication record — has anyone independently reproduced the headline claim?

Output: one short paragraph + a label (well-received / contested / ignored / flagged / too-new-to-tell).

Lens C — Claim-level evidence (does the result survive scrutiny)

Reading what the paper self-discloses + what's externally checkable.

  • Reproducibility checklist (NeurIPS / JMLR / MLRC) — is one included? what's checked yes/no? (Authors are rewarded for honesty about limitations — "no" answers aren't disqualifying, but missing fields are.)
  • Code / data / weights / configs — released? last commit recent? stars? independent forks?
  • Leaderboard / benchmark — entry on Papers With Code? still SOTA, or has the field moved on?
  • For ML/Agent papers — run the contamination + overclaim checks in references/ml-agent-pitfalls.md.
  • Methodology section — sample size, ablations, statistical tests, preregistration.

Output: short paragraph + label (fully-evidenced / partially-evidenced / claimed-only / unverifiable).

Phase 3 — Triangulate

Plot the three labels side-by-side. Three patterns:

PatternLensesRead
ConvergentAll three positiveHigh trust; mine for techniques.
ConvergentAll three negativeLow trust; skim for ideas, do not cite numbers.
DivergentTeam strong, others weak"Hype paper from a real team" — read the method, distrust the claims.
DivergentTeam weak, others strong"Underdog actually right" — rare but real; trust the result.
MixedOne ambiguousUse the ambiguous lens to define skepticism flags.

The mismatch is the story. Don't average lenses into a single number.

Phase 4 — Verdict

Output four things:

  1. Trust band: A (read closely, mostly trust) / B (read with healthy skepticism) / C (skim, verify externally) / D (don't rely on without independent confirmation).
  2. Reading posture: which sections to read carefully, skim, or cross-check. Tie posture to specific lenses (e.g., "Lens C flagged no code → don't trust the implementation details").
  3. Skepticism flags: 2–4 specific things to watch for, each tied to which lens surfaced it.
  4. Falsifier — what would change the verdict: a concrete observable thing (replication on a leaderboard, retraction posted, code released, etc.). Without this, the verdict is static instead of a live hypothesis.

See references/vetting-rubric.md for the report template.

Phase 5 — Read (only if user confirms)

Read the paper with the calibrated posture. Map skepticism flags onto specific sections. For ML/Agent papers, do the contamination check before believing any benchmark number.

Non-negotiables

  • Three lenses run independently before triangulation. Do not let the team lens leak into the citation-context lens — that's anchoring.
  • m-quotient over raw h-index for career-stage fairness. Mention raw h only when career length is unknown.
  • Round, don't fabricate. h ~30k+ not h = 29,847. Scite ratios as percentages, not three-decimal floats.
  • Author-count > 10 is a yellow flag. Score only corresponding + visibly senior co-authors; name the rest as "team labor".
  • Self-citation pattern matters. ≥30% of citations being the team's own work is a flag, surface it explicitly.
  • Field decay: AI/LLM/Agent papers older than 12 months → discount Lens B's volume signal; the field moves faster than citations accrue.
  • Don't conflate seniority with correctness. Famous teams write wrong papers; surface the specific reasoning, not the credentials.
  • Honest uncertainty: "no public signal on <name>" beats an invented score.
  • The falsifier is mandatory: every verdict ends with what would change it.

References

FileUse for
references/sources.mdTool ladder — cheap → better → best path for each lens. Includes APIs (arXiv, OpenAlex, Semantic Scholar).
references/influence-metrics.mdh-index, m-quotient, g-index, FWCI, scite Smart Citations, Altmetric — what each means and where it fails.
references/vetting-rubric.mdThe three-lens scoring rubric + report template + Keshav's 5 Cs as a complementary frame.
references/ml-agent-pitfalls.mdML/LLM/Agent-specific failure modes: benchmark contamination, leakage, overclaiming, NeurIPS-checklist reads.

Common mistakes

  • Reading the paper before vetting it (cognitive anchoring).
  • Collapsing the three lenses into a weighted sum — kills the mismatch signal that's the whole point.
  • Treating raw h-index as career-stage-normalized. Use m-quotient.
  • Skipping Lens B because "the team is famous" — that's exactly when Lens B matters most.
  • Skipping the NeurIPS reproducibility checklist when one is included — it's a free signal sitting in the paper.
  • Forgetting the ML-specific contamination check; it is the dominant failure mode for benchmark claims.
  • Skipping the falsifier line — without it the verdict is a static judgment, not a live hypothesis.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.