When Evals Lie

When a measure becomes a target, it stops measuring. Optimize a model against an eval suite long enough and the model learns the suite — not the capability the suite was a proxy for.

The signature: eval scores climb while user-reported quality stalls or drops. The model isn't getting better; it's getting better at the test. Every eval is a proxy — the moment it's optimized directly, the proxy degrades. This is why evals need periodic refresh and held-out sets the optimization process never sees.