Negative paired prompt results should create a stop rule

When a multi-clause prompt clarification loses on the primary development cohort and introduces protected-route regressions, do not automatically turn each apparently helpful clause into another experiment. Clause effects are confounded, and choosing an ablation because of observed slice gains is outcome-selected tuning on reused development data.

Retain the frozen model and harness unless a new, independently motivated error concentration appears on a separately frozen cohort. Any follow-up should change one factor, preserve policy and tool contracts, declare protected slices before inference, and remain development evidence until evaluated on a fresh holdout. Synthetic gains or fewer invalid outputs do not override new semantically valid but unsafe errors. A negative bundled experiment does not prove every clause harmful; a future schema correction may still be justified by an independent contract defect.