Abstract
This paper asks a practical question that most drift work ignores: once drift is detected, can the behavior actually be corrected through ordinary interventions such as prompt changes and resets?
What it establishes
- Different model families resist correction to different degrees.
- Correctability is measurable, not just anecdotal.
- Drift alerts are more useful when paired with a sense of intervention realism.
Why it matters
Rigidity turns correction into a first-class research topic. It shifts the question from "did behavior change?" to "what kind of change is this, and what can a team realistically do about it?"