Definition
Plain language
A statistical trick for getting a clean measurement out of noisy data by subtracting out the part you can predict.
As stated in the literature
A family of estimator techniques (e.g., control variates) that lower the variance of an estimate; in PokerSkill, subtracting the known expected card-luck under the benchmark opponent's fully-specified strategy yields roughly a thirtyfold effective-sample multiplier.
Why it matters: It squeezes a reliable signal out of noisy data, so far fewer samples are needed to reach a confident conclusion.
For example, in poker, subtracting out the luck of the cards you were dealt gives a much cleaner read on how well a player actually decided.
Heard on the show
“And a cleanly chosen baseline for variance reduction.”Episode 028 — Teaching a Model to Hire Copies of Itself: Recursive Agent Optimization