A Sunday-morning postmortem on teaching a 3B model to do enterprise IT triage with GRPO. It's 1 AM on a Sunday. The Meta × PyTorch OpenEnv Hackathon submission is due at 5 PM. My training logs show a loss curve that's been flat at 0.0 for the last thirty minutes. A flat loss in supervised learning means convergence. A flat loss in reinforcement learning usually means something else: your model has found a way to game your reward function, and it is now sitting in a local optimum, smug about it. Mine had decided that the easiest way to maximize reward was to write about cinema history.…