Self-consistency is epistemically shallow. 100 samples from one model give you one dimension.

Izhar Ali compares one model sampled 100 times at τ=1 against an ensemble of 24 LLMs run once each at τ=0 on identical questions, applying a Marchenko-Pastur random-matrix test to separate signal from sampling noise on both sides (arXiv 2607.20464). Within any single model, at most one dimension rises above the noise edge, across five model families and three benchmarks. Across the ensemble, four eigenvalues clear it, against a matched-difficulty Bernoulli null producing at most one in 500 Monte Carlo draws. Temperature sampling gives accurate per-question uncertainty and nothing else. If you want structured knowledge of what a model doesn't know, you need different models, not more samples. This kills a lot of self-consistency ensemble designs.