Skip to content

RAI & observability

Responsible AI (RAI) policy

Oris applies a default PolicyEnforcer at pipeline boundaries:

  • Input checks — Blocked keys (e.g. password, secret, token), basic prompt-injection heuristics, and simple PII-shaped patterns in string values.
  • Output checks — Blocked terms, plus stub hooks for toxicity/hallucination markers used in tests.

These checks align with InputGuard / OutputGuard hooks in the executor so behavior stays consistent whether you run a full pipeline or SafeRunner.

Layer your defenses

Default policy is a baseline, not a complete safety program. Combine Oris with org policies, model safeguards, and human review appropriate to your domain.

Tracing

Each run produces a RunTrace with ordered StepTrace records: timestamps, status, per-step latency_ms, flags, and optional metadata.

The stable JSON-oriented view is PipelineResult.to_run_summary()—used by the CLI and ideal for logs. See Runs, output & traces.

Audit logging

The tracing package includes audit helpers for redaction-aware logging—see oris.tracing.audit in the repository.