StreamOPD introduces a post‑training recipe that blends verifiable streaming‑video data, thinking‑mode on‑policy distillation, and instruct‑mode deployment to enable memory‑free, recent‑window video understanding. The approach lifts StreamingBench accuracy from 77.9% to 83.9% (within 0.3 pts of a 9B teacher) and improves OVO‑Bench (excluding the hallucination‑detection subtask) by 9.1 points, while the added Spatio‑Temporal CueGate module further boosts performance on OVO‑Bench and Video‑MME. Experiments show that replacing the teacher with a frozen copy of the student’s initial policy retains most gains and even raises hallucination‑detection to 57.0%, demonstrating that the improvements are limited by the recipe rather than abstention loss.