Definition
Plain language
A score for how well data fits a curve — near one means the points sit right on it, near zero means a shapeless cloud.
As stated in the literature
The coefficient of determination, the fraction of variance in an outcome explained by a predictor; can go negative when the predictor is worse than the mean baseline, as raw token-count measures do on real agent traces in the Effective Feedback Compute analysis.
Also called: R²
Why it matters: It tells you how much a model actually explains, and a negative value is a red flag that the model is worse than just guessing the average.
For example, an R-squared near one means a fitted line passes right through the data points, while near zero means the points scatter like a shapeless cloud.
Heard on the show
“And the correlation between what it says and what it does climbs from basically uncorrelated — an R-squared of about a quarter — to around two-thirds.”Episode 152 — Training a Model to Mean What It Says, And Why That Isn't the Same as Being Good