@gerardsans
Absolutely. That shows the underlying distribution topology for that particular set of rollouts. The problem with the current paradigm is that most papers still treat stochastic evaluations as if they were perfect binaries not high dimensional gradients that require a few more dimensions to remove the deterministic assumptions and show coverage, training support, geometric envelope, combinatorial expansion (diverse, fixed) and resistance to interference (reordering, token replacement/removal, distractors).