Chapter D · 4 of 4~10 min

Eval metrics lab

Wire **latency, quality, and error** alerts before users feel the regression.

Reviewed July 2026 · interactive lab

Paths & resources for this lesson

1Try it yourself

Playground

Eval metrics lab

Match alerts to golden signals — latency, errors, quality, saturation.

p95 latency doubled

2Read & reflect

Recap

#What you just did

You matched alert scenarios to the right metric family.

Teach

#Pattern

Track per request: latency_ms, provider_error, groundedness_score, token_cost. Page when SLO burn rate exceeds budget.

Use it

#When you'd use this

  • After first production deploy
  • Before promoting canary or blue-green swap

Watch out

#Watch out

Alert fatigue — start with [p95 latency](/reference/p95-latency) and groundedness, not fifty charts.

Try next

#Try this next

Pick one SLO for your capstone assistant (e.g. p95 < 3s).

3Spark check

Continue learning · glossary & guides