Definition
Plain language
Re-running an AI's process from the exact point where one change would matter, holding everything else fixed, to see if the change actually helped.
As stated in the literature
An optimization technique that rewinds an execution to the first state affected by an edit, freezes the upstream prefix, and replays only the downstream suffix, isolating the edit's effect from stochastic noise while reusing cached computation.
Also called: counterfactual replay optimization
Why it matters: It isolates the true effect of a single change from random run-to-run variation, so you can tell whether an edit actually mattered.
For example, to test whether changing one instruction helped, the system rewinds to the exact moment that instruction took effect, keeps everything before it identical, and re-runs only what came after.
Heard on the show
“They call it counterfactual replay optimization.”Episode 096 — How Treating an AI Agent's Execution Like Git Recovers a Coordination Penalty