Definition
Plain language
Re-running an AI's task without one of its helper steps to see whether that step actually mattered.
As stated in the literature
A counterfactual rollout used in contrastive RL rewards in which a designated stage (e.g., the verification stage) is skipped so the system's pre-verification output can be compared to its full output for credit assignment.
Why it matters: Counterfactual comparisons like this let you attribute credit cleanly to individual pipeline stages, so you can train each one against signal that's actually about it.
For example, the system runs the agent twice on the same task — once with its verification step and once without — and rewards the verifier proportionally to how much the verified output improves over the unverified one.
Heard on the show
“… Once in a "shadow pass" where the Navigator still synthesizes an answer, but only from the pre-verification graph — the …”Episode 051 — Why Parallel Sampling Plateaus, And What Evidence Graphs Do Instead