From inner diversity to outer crowds: How instructions reshape error structure before aggregation

Diversity is widely regarded as a core condition for collective intelligence, yet its effects in estimation tasks are often conditional and mechanistically unclear. Here we offer a descriptive reanalysis and conceptual synthesis of inner-crowd datasets, most drawn from our earlier studies, to clarify how instruction-induced second estimates change crowd error before aggregation. Rather than presenting a new experimental discovery, the paper reinterprets existing data through a common lens. Across datasets, we compute item-wise crowd mean squared error (MSE) for Round 1 (Own) and Round 2 (instruction-induced alternative) estimates, and quantify item-wise change in error (ΔMSE) to describe error redistribution across questions. We focus directly on crowd error, bias summaries, dispersion-related movement, and item-level redistribution; possible changes in dependence are discussed more cautiously because these datasets do not identify them cleanly. Estimating from the perspective of “people in general” improves collective accuracy in one 20-item set but not in a related 8-item set, showing strong item-pool dependence. A repeated-estimation baseline yields near-zero net improvement, indicating that second-try effects alone are insufficient. Dialectical bootstrapping worsens mean crowd accuracy, consistent with bias deterioration. A disagreeing-perspective instruction increases item-wise movement and dispersion but does not improve crowd accuracy on average, illustrating redistribution without net gain. Finally, in a limited pseudo-replication of cognitive-process diversity claims, mixed-instruction crowds in these datasets do not outperform the best single-instruction baseline across group sizes. The contribution is therefore a mechanistic re-interpretation of inner-crowd data: instruction-induced transformations often redistribute error across items, and collective gains depend on when they add useful signal without worsening bias.