What problem does it solve? Pre-submission papers often contain experimental numbers (F1, AUC, Recall) that cannot be traced back to actual code runs, risking fabricated or mismatched claims slipping into submissions. This Skill performs a deep semantic review of a paper and audits every numeric claim against independently executed code outputs before submission. ## Core Features & Use Cases - Experimental Data Integrity Audit: Extracts all numeric claims from the paper, checks for corresponding scripts in the code directory, and labels each claim CLOSE, MISMATCH, or FABRICATED based on independent reproduction. - Multi-Round Review Pipeline: Runs a structured workflow of L1-L2 structural checks, semantic review, fixes, deep multi-reviewer deliberation, and L5-L7 compliance checks, with independent reviewers and strict token budgets (10k per round, 50k cumulative). - Use Case: Before submitting a machine learning paper claiming F1=0.92, run this Skill to re-execute the training script, archive JSON/CSV evidence, and confirm the claimed metric matches the reproduced result—or flag it as fabricated if no supporting code exists. ## Quick Start Review my paper in the papers/ directory, audit every experimental metric against the code in 03-code/, and report which claims are CLOSE, MISMATCH, or FABRICATED.
