Audit supervised fine-tuning datasets against the behavior and task they are meant to teach. Use when inspecting SFT JSONL, chat messages, instruction-response pairs, tool or agent trajectories, code corpora, synthetic examples, revised datasets, base-model evals, pass@k skill maps, train-validation-test splits, benchmark contamination, duplicate lineage, answer correctness, token limits, data mixtures, or train-readiness claims. Produce an evidence-backed row catalog, quality gates, duplicate and contamination report, corpus composition analysis, and a train, review, replace, reject, or eval-only decision.
