Definition
Plain language
When an AI system mistakenly promotes a result that wasn't really better, corrupting every later comparison built on it.
As stated in the literature
A failure mode in shared-state experiment search where a falsely-promoted candidate (an improvement inside the noise floor) becomes the reference baseline, corrupting downstream comparisons; mitigated by replication gating.
Why it matters: It quietly corrupts an entire chain of automated comparisons, which is why systems gate promotions behind repeated checks before declaring a winner.
For example, if a tweak that wasn't really better gets crowned the new best version, every later experiment is unfairly measured against that flawed baseline.
Heard on the show
“They call this champion pollution.”Episode 095 — Seven Wins to Zero: How Organizing AI Agents Like a Lab Changes the Search