Phase 2 Calibration: Per‑Category OOD Thresholds + Group‑Relative Reward Normalization in My Scene Compiler
4d ago · 12 min read · I didn’t add per‑category OOD thresholds because it was academically elegant. I added them because my baseline runs were telling me the same story over and over: some prompt categories were systematically getting mis-gated by a single global uncertai...
Join discussion