anulum/Director-AI Guardrail Gate
Fail a CI run when the Director-AI hallucination guardrail's quality on your labelled eval set drops below a threshold. Token-level guardrail, run as a quality gate.
View on GitHubTrust Signals
- Scorecard Score
- not yet scored
- Maintenance Recency
- Stale
- License
- None
Inputs
| name | description | required | default |
|---|---|---|---|
| dataset | Path to a JSONL file of {prompt, response, expected} cases (expected: approve|reject). | yes | — |
| min-accuracy | Minimum overall accuracy (0-1). The gate fails below this. | no | 0.9 |
| min-catch-rate | Optional minimum hallucination catch rate (0-1) on reject-labelled cases. | no | "" |
| max-false-halt | Optional maximum false-halt rate (0-1) on approve-labelled cases. | no | "" |
| profile | Optional Director-AI config profile (e.g. medical, finance, legal). | no | "" |
| output | Optional path to write the JSON gate report (for CI artefacts). | no | "" |
| extras | pip extra to install with director-ai (e.g. nli, vector). Use 'none' to skip extras. | no | nli |
| version | director-ai version spec to install (e.g. '', '==3.18.1', '>=3.16'). | no | "" |
| python-version | Python version to set up. | no | 3.12 |
Outputs
no outputs