Проверка фактов по протоколу CoVe

research

verify-claims is a Claude Code skill that runs the Chain-of-Verification (CoVe) protocol from Dhuliawala et al. 2023 (arXiv:2309.11495) on a draft file or text block containing factual assertions. It extracts five claim types — citation pairs (what the draft attributes to which work), numerical facts, negative literature claims, named entities, and dataset assertions — then spawns the claim-verifier agent in a forked, fresh context that never sees the original draft, eliminating confirmation bias. Allowed tools are Read, Grep, Glob, Task, and Write; arguments include a file path (.md, .qmd, .tex), --source flags, and --no-fail-closed. Each claim receives a verdict of HIGH-WARN, MED-WARN, LOW-WARN, or EXPLAINED — HIGH-WARN claims block /commit unless explicitly overridden. Ideal for researchers who need to fact-check citations before submission, catch hallucinations after AI-assisted writing, or validate that cited papers actually support the attributed claims.