A user practice of treating AI errors as informative, deliberately examining what a malfunction reveals about a system rather than discarding it. It denotes a discrete investigative use of an observed error.
EVENT. Counted once per episode in which a user, upon encountering a model error, explicitly probes or analyses the error to infer system behaviour (follow-up questions about the failure, deliberate re-elicitation). Simply re-prompting for a correct answer is excluded.
Proposed measurement protocol (not yet empirically validated): Rater-coded count of error-leveraging episodes per session, conditioned on detected error turns; could be reported as episodes per 100 error events. Coder reliability on the leverage-vs-retry distinction via Cohen's kappa.
https://andreasehstandlicenseofclarityloc.github.io/augmanitai-periodic/#topic-system-behavior