False Cascades and the Cost of Truth

We study sequential social learning with continuous actions and conformity when agents can endogenously generate hard, publicly verifiable evidence. Actions transmit soft information whose visibility depends on responsiveness to private signals and on observational granularity; investigations produce hard evidence only in the true state and, once found, are disclosed and reset public beliefs. We deliver two primitives. First, a soft-channel informativeness threshold characterizes when actions remain publicly revealing under coarse observation, clarifying why learning can be locally mute at classical cascade boundaries even with continuous actions. Second, a transparent knife-edge for verification guarantees boundary breakability: whenever the expected private return to a one-shot investigation exceeds its cost, occasional disclosures overturn any false cascade with positive probability in finite time. Combining both yields a compact resilience frontier with clean comparative statics in signal quality, responsiveness/conformity, observability, verification quality, and investigation costs. The results provide policy levers -- subsidies to investigation, investments in verification quality, disclosure incentives, and platform design -- that ensure correction of wrong cascades.