| name | citation_accuracy |
| description | Verifies every Markdown link citation in `_papers/` and `_ideas/` notes resolves to the correct paper and that the attributed claim actually matches what the paper says. Gate skill — run before finalizing literature_review/go/check outputs. |
citation_accuracy — claim-to-citation gate
Process
- Resolve every link. For each Markdown link citation
[Authors et al., Year — Title](link) in the output being finalized, confirm the link is live and actually points to the claimed paper (not a 404, not a different paper, not a search results page).
- Match claim to source. For each specific claim attributed to a paper (a number, a method name, a conclusion), confirm the paper itself actually states that — not a paraphrase that drifted into overclaiming, and not a claim from a different paper that got misattributed.
- Check for exaggeration drift. Compare the summary/comparison language against the paper's actual hedging — if the paper says "may improve" and the note says "proves," that's a violation.
- Block finalization on failure. If any citation fails resolution or claim-matching, the output is not done — fix the citation or the claim before it's presented as final.
Anti-Rationalization
- "The link looks right" — looking right isn't resolving; actually check the link target when possible.
- "It's basically what the paper says" — "basically" is where overclaiming hides; check the literal wording of the claim.
- "This is the last step, no time to double check" — this skill IS the double-check; skipping it defeats its purpose.
Evidence
Citation accuracy passes when: every link in the output resolves to its claimed paper, and every specific claim attributed to a paper is directly supported by that paper's actual text (not a paraphrase drift).
Red Flags
- A citation link that 404s or points to the wrong paper
- A claim attributed to paper A that's actually from paper B
- Hedged language in the source ("may", "suggests") restated as certainty ("proves", "shows definitively")
