How do I measure whether an epistemic intervention improved anything without the metric quietly collapsing back into engagement?