Research Results Auditor

Quick start

Collect the concrete result artifact, the exact source claim, and the evaluation protocol.
Classify the requested assurance as exploratory, confirmatory, independently verified, or operational/high-stakes.
In tracked work, initialize results-audit.json and results-audit.md with scripts/init_results_audit.py.
Reconstruct the experiment binding, submitted run, verifier decision, gate result, lineage, artifact digests, and failure history before interpreting the headline metric.
Audit protocol integrity, metrics, baselines, uncertainty, selection, outcome accounting, provenance, confounds, and actual independence.
Record one bounded audit object per claim under review; do not hide conflicting runs inside a prose summary.
Validate with scripts/validate_results_audit.py; use the linked profile for orchestrated runs.
Hand the canonical JSON audit to research-paper-plan instead of asking it to infer support from figures or filenames.