Turn behavior into checks
Metric Builder now
Ask /evalglass add-metric and your agent scaffolds a proposed metric for you to review — never an approved threshold.
What it does
The add-metric verb appends a metric to your evalglass.yaml — a built-in, or one backed by a scorer you write — so you edit instead of starting blank. It lands proposed and informational: it can report, but it cannot gate until you validate it and approve a threshold in your own config. A judge metric (add-judge) lands uncalibrated and can't gate until you calibrate it.
Say this to your agent
Add a faithfulness metric to my evals for me to review.
A new metric lands proposed — it can't gate until you say so.
The agent populates host-owned files; it never fills in authority. There is no verb that approves a threshold or activates a gate — that stays your edit. The authoring verbs →
In the docs
Read the detail →
what your agent reads for this