What problem does it solve? Research papers often contain hidden experimental flaws: metrics compared across different dataset versions (e.g., WDBC 699 vs 569 samples), ablation results that cannot be reproduced, bibliography entries out of sync between thebibliography and references.bib, and figures cited redundantly. This Skill systematically audits a paper's experiment directory to detect and flag these integrity issues before publication or peer review. ## Core Features & Use Cases - Dataset Version Locking: Records and fixes dataset versions before auditing, blocking cross-version metric comparisons that exceed the 0.5% difference threshold with a MISMATCH flag. - Ablation Reproducibility Checks: Re-runs each ablation configuration's independent script and compares output JSON/CSV against claimed values, marking differences above 0.5% as inconsistent. - Citation & Figure Validation: Verifies one-to-one correspondence between thebibliography and references.bib entries, and detects figures (like fig6) cited more than twice in the text for deduplication. - Multi-Source Cross-Validation: Requires at least two independently generated JSON/CSV outputs per metric; single-source metrics are marked UNVERIFIED. - Use Case: Given a LaTeX paper directory with WDBC experiments, ablation scripts, and output JSON files, the Skill produces a structured audit log where each of six verification checks is marked pass/fail with traceable evidence, and any MISMATCH blocks downstream conclusions. ## Quick Start Audit the experiments in my LaTeX paper directory, checking dataset version consistency, ablation reproducibility, bibliography synchronization, and figure citation redundancy.
