sengol-io/Sengol AI Governance

Run a Sengol evaluation suite and gate CI/CD on compliance obligations

Trust Signals

Scorecard Score: not yet scored
Maintenance Recency: Stale
License: None

Inputs

name	description	required	default
config	Path to sengol.yaml (per-org runtime config).	no	sengol.yaml
python-version	Python version to set up before installing sengol.	no	3.11
sengol-version	Pin a specific sengol release (e.g. "0.2.0"). Empty installs latest.	no	""
extras	Comma-separated sengol extras to install (e.g. "anthropic-judge,openai-judge,model"). The default ships both LLM-judge provider SDKs so Faithfulness/Hallucination-style judges wire up out of the box; add "model" for the statistical evaluators (PSI/AUC/Fairness), or set to "" for a bare install with deterministic evaluators only.	no	anthropic-judge,openai-judge
fail-on-gate-failure	Exit the action non-zero when the PromotionGate verdict is fail.	no	true
output-format	CI/CD output adapter: "github-summary" (default) or "junit". Workflow-command annotations are emitted alongside both modes.	no	github-summary
junit-output-path	Path to write JUnit XML when output-format=junit. Ignored otherwise.	no	eval-results.xml
export	Comma-separated result-exporter names to push results to (e.g. "mlflow"). Empty disables export. Options come from the sengol.yaml `exporters:` block; exporters are optional sinks and never change the gate verdict.	no	""
pr-comment	Post or update a sticky PR comment with gate results on pull_request events. Requires GITHUB_TOKEN with pull-requests: write. Degrades gracefully (::notice + step summary) when the token lacks permission — fork PRs never fail the gate over a comment.	no	true
repeat	Run LLM evaluators N times per input for flake tolerance. Deterministic evaluators always run once. Default 1 (no repeat). Pair with min-pass.	no	1
min-pass	Require M of N (repeat) LLM-evaluator attempts to pass per score. Default 0 means all attempts must pass (same as repeat value). All attempt records are written to evidence for auditability.	no	0
dataset-path	Path to a JSONL trace dataset to evaluate (passed to `sengol run-suite --dataset-path`). Empty uses the dataset configured in sengol.yaml.	no	""
dataset-id	Dataset identifier recorded on the evaluation run (passed to `sengol run-suite --dataset-id`). Empty derives it from config.	no	""

Outputs

name	description
passed	String "true" or "false" — overall gate verdict.
score	GoldScore (0.0–1.0) for the evaluation period. Absent when GoldScore is not configured.
report_url	Path or URL to the generated PDF evidence pack. Absent when PDF generation is not enabled.
no_agent_activity	String "true" when at least one obligation found no agent activity in the period. Absent otherwise.
regression	String "true" when current pass-rate regressed vs the configured baseline, "false" when at or above baseline. Empty string when no baseline is configured in sengol.yaml.