--- rl_score_centering.v1.py	2026-11-04 02:50:30.462151190 +0000
+++ rl_score_centering.py	2026-11-04 02:50:30.471151150 +0000
@@ -22,7 +22,7 @@
     p_tail = torch.clamp(1 - p_head.sum(-1), min=EPS); q_tail = torch.clamp(1 - q_head.sum(-1), min=EPS)
     rho = q_tail / p_tail
     alpha = rho * w(1 / rho)
-    head_weight = w(torch.exp(head_logp - samp_head))
+    head_weight = w(torch.exp(torch.where(hm > 0, head_logp - samp_head, torch.zeros_like(head_logp))))  # empty slots hold -1e9
     residual = (q_head * head_weight - alpha[:, None] * p_head) * hm
     correction = (residual.detach() * head_logp).sum(-1)
     token_weight = w(torch.exp(token_logp - samp_tok))