Umbrella audit for a paper draft. Three independent phases: claim audit, citation audit, and optional experiment audit. Each phase uses a fresh, cross-model reviewer to prevent self-judging bias. Output JSON ledger and Markdown report with per-item verdicts. Advisory only; downstream personas or the user decide what to apply. Use when user says "audit the paper", "check the citations", "verify the numbers", "pre-submission audit", or when @writer reaches Phase 5 audit, or when @council needs an independent integrity check before adversarial review.
