Two Motives for Verification in Information Cascades

Technical Summary: Two Motives for Verification in Information Cascades

Problem Statement
This paper investigates sequential social learning in an environment where agents can endogenously produce public evidence through costly verification. Unlike canonical information cascade models (Bikhchandani et al., 1992; 2024) where public news arrives exogenously, this model posits that producing evidence requires a private investment. The central question is how the decision to verify interacts with social learning when the decision itself is public and when verification yields two distinct private returns: improving the agent's own action accuracy and securing a "first-discovery" reward. The authors ask whether the opportunity to produce such evidence persists indefinitely or if it can vanish, leading to "false cascade traps" where society stops learning despite the possibility of future proof.

Methodology
The authors construct a model of an infinite sequence of short-lived agents facing a fixed binary state $\theta \in \{0, 1\}$.

• Information Structure: Agents observe a public history $H_t$ and a private binary signal $s_t$ with precision $q$.
• Verification Technology: Agents may pay a fixed cost $c$ to run a one-sided test. Success ($e_t=1$) conclusively proves $\theta=1$ and is automatically disclosed, granting a discovery reward $V$. Failure ($e_t=0$) is public but inconclusive.
• Timing: The sequence is: observe public belief $\mu_t$ and private signal $s_t$ $\to$ choose investigation $i_t$ $\to$ observe outcome $e_t$ (if $i_t=1$) $\to$ choose action $a_t$.
• Equilibrium Concept: Perfect Bayesian Equilibrium where strategies are sequentially optimal and public beliefs update via Bayes' rule, accounting for the selection effect of the investigation decision itself.

The analysis decomposes the net gain from investigation, $G(x)$, into a diagnostic component (expected improvement in action accuracy) and a reward component (expected discovery benefit). The authors derive the exact geometry of the investigation set (the set of private posteriors $x$ for which $G(x) > 0$) and map these private incentives to public belief transitions.

Key Contributions and Results

1. Non-Monotone Investigation Incentives and Disconnected Regions:
   The primary theoretical contribution is the characterization of the investigation set. The diagnostic value of verification is "tent-shaped," peaking near the action threshold ($x=1/2$) and vanishing at high posteriors where the action is already determined. Conversely, the expected discovery reward is weakly increasing in the posterior probability of state 1.

   • Theorem 4.3: The equilibrium investigation set has at most two connected components. Under specific parameter conditions (specifically when the diagnostic motive is strong near the threshold but the reward motive is strong at high beliefs, and costs are intermediate), the investigation region becomes disconnected. There exists a central region where agents investigate for diagnostic reasons, a gap where they do not investigate, and a high-posterior region where agents investigate solely for the discovery reward. This is not a multiplicity result but a unique best response driven by two distinct motives.

2. Selection Effects in Public Learning:
   Because investigation is chosen based on private signals, an observed attempt reveals information about the investigator's signal type.

   • Proposition 5.1: The authors derive an exact public transition map that combines the selection effect, the test outcome, and the subsequent action.
   • Lemma 5.2: A failed investigation does not necessarily lower public belief. If the attempt reveals that the investigator had a favorable signal (which was willing to test despite the risk of failure), the "favorable selection" can outweigh the adverse information of the failure, causing public belief to rise.

3. Finite Verification Window and Cascade Traps:
   Contrary to the intuition that a positive probability of success implies eventual discovery, the paper proves that verification opportunities are finite almost surely when costs are strictly positive.

   • Theorem 5.3: Under either state, the total number of investigations is finite almost surely. In state 1, this follows from the geometric tail of independent failure probabilities. In state 0, infinite attempts would drive public belief to zero, eventually entering a region where no agent finds it profitable to investigate.
   • Proposition 5.6 & Corollary 5.8: The model identifies "cascade traps"—absorbing sets of public beliefs where no agent investigates and actions remain fixed. A lower trap (false cascade in state 1) always exists when $c>0$. A single failed investigation can move the public belief directly into such a trap, permanently halting learning.

4. Robustness:

   • Noisy Verification (Section 6.1): Introducing a small false-positive rate ($\eta > 0$) preserves the two-component geometry of the investigation region, though it removes the conclusive nature of success required for the finite-stopping proof.
   • Heterogeneous Costs (Section 6.2): With privately heterogeneous costs, the deterministic investigation rule becomes a smooth probability. The rise-fall-rise shape of the investigation rate persists, and the finite-verification result holds provided costs are bounded away from zero.

Significance and Claims
The paper claims to shift the focus of cascade fragility from whether exogenous news could break a cascade to whether such news will be produced. It demonstrates that:

• Verification incentives are not monotone; they can be sustained at very different belief levels for different reasons (diagnostic vs. reward).
• The decision to investigate conveys information that can counteract the negative signal of a failure.
• The opportunity to verify is endogenously finite. Even with a positive chance of proof, the system can fall into an absorbing false-cascade trap where no further agents find it optimal to investigate.

The authors explicitly state they do not offer a sufficient condition for eventual truth. Instead, they characterize the conditions under which verification opportunities disappear and how the geometry of private incentives shapes the public learning process. The work isolates the decision to produce evidence from the decision to conceal it, showing that the "cost of truth" can lead to a permanent cessation of learning even when the horizon is infinite.