Ship

Promote a gate

Make a metric enforce CI — deliberately.

A metric gates only when four host-owned conditions all hold: a validated dataset, an approved threshold, the metric set to gating, and — for a judge metric — a calibrated judge. Promotion is an explicit act, never a default — and it is a host YAML edit, not a command.

There is no promote / gate / approve / certify verb Nothing under /evalglass can turn a metric into a gate. Promoting a gate is a host-owned YAML edit you make — setting metric_status: gating, threshold_approval: approved, and the dataset to validated — done after the host-owned preconditions are met. The promoting-a-gate skill only guides this. An agent may type the edit at the host’s explicit direction, but no verb and no agent can grant the authority — only the host can.

After First Run, every metric is informational. To make one enforce CI, you clear each reason that keeps it informational. Do them deliberately — each step is you, the host, taking responsibility for a claim.

The four conditions

Step	Clears the reason	Owned in
1 · Validate the dataset	`dataset_proposed`	The dataset’s host-owned status (proposed → validated).
2 · Approve a threshold	`threshold_proposed`	A host `ApprovedThreshold` record.
3 · Set the metric to gating	`metric_status=informational`	The metric’s status in `evalglass.yaml` / authority records.
4 · Calibrate the judge (judge metrics only)	`judge_uncalibrated`	A host calibration record.

The host YAML edit

Promotion is this edit, made by the host. The metric declares its intent, but the configuration cannot approve itself — an under-specified metric is forced back to proposed / informational. The approval lives in host-owned records the config references:

yaml

metrics:
  - name: exact_match
    evaluator_ref: exact_match@1
    lens: reference
    score_type: binary
    dataset: datasets/arithmetic.jsonl   # must be a validated dataset
    metric_status: gating                # intent — honored only with the records below
    threshold: 0.95
    threshold_approval: approved          # backed by a host ApprovedThreshold record
    direction: higher_is_better

Re-run, and the metric’s authority resolves to can_gate:

value meets the approved threshold → pass (exit 0).
value below it → fail (exit 1).
could not be measured, or coverage incomplete → blocked (exit 1).

Promote one gate at a time Each promoted gate is a promise that the dataset, threshold, and (if applicable) judge are sound. Promote deliberately and review each record — a gate you cannot defend is exactly the false confidence EvalGlass exists to prevent.

Next steps

Ship

Wire CI

Make your pipeline act on the verdict.

Trust Model

Authority

The ladder these four steps climb.

Reference

Verdicts & exit codes

How the new verdict maps to an exit.