Phantom Clipping: Token-Level Strip Plot

Each dot is one token. Phantom-clipped tokens (marked with X) land in the red clip zone only because of the precision gap β, not because the policy actually changed.

0
1000
Actual ratio exp(α + β)
adv > 0 (reinforce)
adv ≤ 0 (penalize)
phantom-clipped
shaded = clip zone