Values are easy to state. Hard to measure. Harder to enforce. Every AI system optimises for something. Today you will pick a value, build a benchmark that tests whether a model honours it, steer the model with a prompt, and measure whether the needle moved — or whether the model just got better at looking aligned.