EvalGlass

Use case · CI for AI applications

A release needs a gate that fails honestly.

CI for AI should block only when a check you approved truly falls short — never on a number nobody reviewed. Ask your coding agent to wire CI on your approved checks; the build reads one typed verdict, and a fresh, unapproved run stays informational.

Set up a CI gate for the checks we have reviewed and approved.
You Set up a CI gate for the checks we have reviewed and approved.
Your coding agent · Claude Code plugin Wired. CI now runs the vendored runtime and branches on the typed verdict. Your approved exact-match gate passes at 0.96; the un-approved policy check stays informational and does not affect the build. A gate fires only after your deliberate approval — I can type the config edit, but I can’t grant the authority.

What it checks

the verdict from your approved gates · exit code 0 (pass / informational) or 1 (fail / blocked). A gate fires only after a deliberate host approval — there is no gate / approve / certify verb. A fresh run gates nothing and never fails the build.

The scorecard

ci run · approved gates pass
exact_match (gated)0.96 pass
workflow-policyinformational
ci_should_failfalse

Verdict pass — a specific approved gate met its threshold. The policy check is informational and does not gate. Illustrative example, not a measured result.

experimental The shipped gate reads the point threshold; the epistemic core hardens it to decide on the lower confidence bound — so a tiny-sample perfect score is not enough to pass the default gate (a named point-smoke policy is the explicit escape hatch). And authority is capability-typed: a fake or uncalibrated judge produces evidence, not authority, so it can never move a gate. Confidence & intervals →

What it will not claim

Exit 0 is not “quality verified” — it covers both pass and informational. Green means a specific gate you approved met its threshold, never “safe AI.” The agent can wire the pipeline; only you can approve what gates it. No false green →

Ask your coding agent.

Evaluate my agentic app using EvalGlass.

Related

All use cases
every change moment
Eval-driven development
where CI fits the change-then-check loop
Scorecards
what each run hands back
Get the plugin
two commands in Claude Code
Docs
the detailed reference