Determine the effect of disclosure mode on rarity-dependent engagement

Determine whether the reported elicitation-condition effects on language-model explanatory engagement hold equally across disclosure modes, or whether they are stronger or present only when the model is directly informed of the tool failure versus when it must infer the failure unprompted.

Background

The study pools results across two disclosure modes: immediate disclosure, in which the model is told that a failure occurred, and silent-batch-reveal disclosure, in which the model must detect the failure without being explicitly informed. Because the primary results are largely pooled across these modes, the reported effects of failure rarity and elicitation condition may be confounded with how the anomalous outcome is revealed.

The paper identifies disclosure mode as an unresolved methodological factor and reports a disclosure-mode-split analysis for selected findings. However, it does not establish whether all condition-level effects generalize across disclosure modes or whether some effects depend specifically on one mode of revealing the failure.