Add the evaluation team when you need it.

Import exported OpenTelemetry / OpenInference / local-JSON traces today — no SDK, no network. A live pull from Langfuse, Phoenix, and LangSmith is an opt-in lane reachable with the connect --live verb experimental — off the required path, a pull is proposed and cannot gate.

Scaffold a proposed metric or judge for you to review — never an approved threshold.

A decided non-goal, not a backlog item — governed, not generated: no generator ships, by design. Any generated case would land proposed and could never enter as validated reference data.

The discovering-metrics skill reads your code, output schemas, prompts, and any traces — across any agentic framework — and proposes a typed MetricCatalog of what to measure. With no traces yet, your output schema alone drafts a contract-tier structural floor; everything lands proposed.

Route uncertain examples to human review; validated truth returns to your repo.

An LLM judge can’t gate your build until it’s calibrated to your labels.

An opt-in subprocess judge — the host provides a command EvalGlass invokes to score. It lands uncalibrated → informational and cannot gate until it is calibrated to your labels.

A generic OpenAI-compatible transport that imports no provider SDK; rubrics are host-injected, and the lane is authority-free and deletable. Opt-in, uncalibrated → informational.

See what regressed since your last run — only when the two runs are comparable.

Fail the build only when an approved gate is breached — never on a fresh, informational run.

Send scorecards outward — local export now, a hosted view is next. No telemetry; the repo stays the source of truth.

Send a copy of your scorecard to an optimizer — EvalGlass never edits your prompts.