← ClaudeAtlas

cite-checklisted

Paper citation and claims verifier. Use before submitting a paper, after major eval runs that change numbers, or when reviewing a paper draft. Traces every numeric claim in the paper to a result JSON on disk; flags UNTRACED, MISMATCH, STALE_REF, MISSING_CITE, PHANTOM categories.
SamyakJhaveri/loam · ★ 0 · Data & Documents · score 71
Install: claude install-skill SamyakJhaveri/loam
# Paper Citation and Claims Verifier Use when auditing a paper draft for numerical accuracy, stale references, and unsubstantiated claims. Traces every number and assertion back to actual result files on disk. **Trigger:** When user types `/cite-check` with an optional file path. ## Iron Law ``` NO NUMBER IN THE PAPER WITHOUT A TRACEABLE DATA SOURCE ``` ## Arguments - `$ARGUMENTS` — optional: path to paper draft file. Defaults to `docs/paper_draft.md`. ## Anti-Rationalization Table | Excuse | Reality | |--------|---------| | "The numbers are approximately right" | Reviewers check exact numbers. "Approximately" gets your paper rejected | | "I'll update the numbers before submission" | Stale numbers propagate — every draft review builds intuition on wrong data | | "Only the final version matters" | SC papers are reviewed once. There is no "fix it later" after submission | | "The trend is correct even if the number is off" | A wrong number undermines trust in every other number in the paper | ## Red Flags — STOP and Escalate - A percentage or count in the paper that does not match any result file on disk - A reference to a model that has been dropped from the campaign (e.g., azure-gpt-4.1) - A table or figure referenced in text that does not exist in the draft - A claim about "all models" or "all kernels" that actually excludes some - A speedup ratio cited from wall-clock timing (known to be unreliable — see known-issues.md) If any red flag triggers: mark the claim as