What you can't do without evals

• Can't detect regressions when you change a prompt
• Can't compare prompt versions objectively
• Can't know if a new model is actually better
• Can't run quality gates in CI