Definition
Plain language
The property that your measuring stick stays fixed — a score early in a process means the same thing as a score later, so you can compare them.
As stated in the literature
The assumption that the evaluation function is constant over the course of a search, which underwrites comparison and the convergence guarantees of self-improvement loops; broken the moment the judge itself is allowed to evolve, which the Red Queen framework restores epoch-locally.
Why it matters: Without a fixed measuring stick, you can't tell whether a self-improvement loop is genuinely getting better or whether the grader just changed underneath it.
For example, it holds when a score of 80 measured at the start of a search means the same thing as a score of 80 measured much later, so the two can be compared directly.
Heard on the show
“The technical word is stationarity, and it's not a detail, it's the whole foundation.”Episode 178 — How an AI Reviewer Learned to Stop Going Easy on AI Writing