Held-out monitoring (NO gradient): analytic per-position acceptance
a_k = 1 − ½‖p^d_k − p^t_k‖₁ and the expected accepted length
τ = 1 + Σ_{j=1}^{γ} Π_{i≤j} a_i (the +1 is the target bonus token). This is
the cheap proxy to early-stop on — τ is the number that matters, not loss.
Returns means over the batch.
Held-out monitoring (NO gradient): analytic per-position acceptance a_k = 1 − ½‖p^d_k − p^t_k‖₁ and the expected accepted length τ = 1 + Σ_{j=1}^{γ} Π_{i≤j} a_i (the +1 is the target bonus token). This is the cheap proxy to early-stop on — τ is the number that matters, not loss. Returns means over the batch.