RED REVIEW
RED REVIEW is a forensic audit tool for clinical-AI preprints that produces executable evidence instead of prose critique.
Core trick
Resolves each citation to its DOI, fetches the cited paper, and uses extended thinking to check whether it supports the stated claim.
Description
RED REVIEW is a forensic audit tool for clinical-AI preprints that produces executable evidence instead of prose critique. Paste an arXiv URL. In under two minutes, it returns a structured red-team report backed by runnable artifacts: a citation whose content does not actually support the paper's claim, a reference domain mismatched between paper and source, a statistical test misused since 2012.
Three Claude Code agents run in parallel (Citation Forensics, Guideline Auditor, Statistical Critic) unified through a skills-first architecture with CONSORT-AI and TRIPOD+AI authored as progressive-disclosure item files.
The hero capability is Citation Semantic Verification: Opus 4.7 with extended thinking fetches each cited paper, reads it in full, then checks whether the citation actually supports the claim being made. This inverts the "LLMs hallucinate citations" problem, applied to human authors. On a real published paper, it caught a methodology citation that pointed to the test's implementation paper rather than its origin, and mischaracterized the test as a t-test when it is U-statistics-based.
AI peer review that produces prose critique manufactures confident wrong answers. RED REVIEW produces evidence.
README excerpt
# RED REVIEW A forensic auditor for clinical-AI preprints, backed by executable evidence. Paste an arXiv URL or upload a PDF and the system returns a structured red-review report. Each finding links to a concrete artifact: a resolver failure, a verbatim quote from a cited paper, a recomputed statistic, or a checklist item with rubric reasoning. Built for reviewers, journal editors, and clinical AI researchers who need traceable audit trails rather than vibes.  ## What it does Input is an arXiv URL or a local PDF. Output is a single-page audit report with three classes of evidence. Citation forensics resolves every reference (DOI, arXiv id, PMID) against authoritative sources and flags unresolvables. For high-impact citations, it fetches the cited paper's full text and judges whether the citing claim is actually supported. Guideline compliance audits the paper against CONSORT-AI and TRIPOD+AI item-by-item. Each item produces a verdict (pass, fail, unclear, not_applicable) with supporting quotes or absence evidence. Statistical validity recomputes published statistics where the data permits. Primary targets are DeLong AUC comparisons, the misuse of DeLong on nested models, and absent Firth penalisation in rare-outcome logistic regression. ## How it works Three Claude Code agents run in parallel through a thin orchestrator. Each agent writes its own JSON output file. A deterministic merge step then produces the final repor