We study sequential social learning when agents can sometimes pay to verify a claim and obtain hard, publicly checkable evidence. Each agent observes the public history, receives a private signal, may investigate at a cost (succeeding only when the claim is true), and can disclose or conceal any proof. Actions are binary or continuous, with a conformity pull toward the prevailing consensus. We characterize when false cascades persist and when societies self-correct. In the binary benchmark, we derive an investigation cutoff and show how its location relative to classic cascade bands governs breakability; a simple knife-edge condition guarantees that any wrong cascade at the boundary is overturned with positive probability. With continuous actions, coarse observation and conformity can recreate cascades, yet occasional disclosures collapse them. These forces yield a tractable "resilience frontier" with transparent comparative statics and policy levers.
Lukyanov et al. (2025) studied this question.