An outcome metric measures beneficial change in user or organisational reality. It differs from an output metric such as number of generated answers.
A leading metric gives an early signal of progress; a guardrail metric detects unacceptable side effects such as error, exclusion, cost or workload transfer.
The build-measure-learn loop delivers the smallest useful experiment, collects evidence and uses it to adapt the problem, solution or quality bar. Decision records preserve product and technical judgement for humans and agents.
