# Held-Out Validation **Domain:** Machine Learning / Statistical Evaluation **Doc Type:** Canonical Method Node **Maturity:** Foundational ## Definition **Held-out validation** evaluates a fitted model on data not used to estimate its parameters. The separation tests whether learned structure generalizes beyond the examples that produced it. ## Historical Context [[wiki/Alexey Ivakhnenko|Alexey Ivakhnenko's]] 1971 polynomial systems retained or discarded candidate elements according to performance on a separate testing set, using external evaluation as a defense against [[wiki/Overfitting|overfitting]]. ## Governance Context Held-out evaluation reduces one form of self-confirmation, but its result still depends on how representative the held-out data and chosen metric are. ## See Also [[wiki/AI Benchmarking|AI Benchmarking]], [[wiki/Training Data|Training Data]], [[wiki/Machine Learning|Machine Learning]]