-
artifact v4: clarify whitening mean/var are per masked TOKEN (token-weighted), not per sample
-
artifact v4: clarify whitening mean/var are per masked TOKEN (token-weighted), not per sample
-
artifact v3: add "what normalize_advantages does" (global batch z-score whitening)
-
artifact v3: add "what normalize_advantages does" (global batch z-score whitening)
-
artifact v2: sharper root cause + exact-context code + worked one-dissenter example; fix 16<->0 (fully-unanimous is SAFE)
-
artifact(html): GRPO 1e6 advantage blow-up incident report
-
artifact: GRPO ±1e6 advantage blow-up incident (reward LOO std-norm) 2026-07-07