Skip to content

Explainable and user-friendly guardrail decisions #32

Description

@glaborie

Summary

Guardrail and safety responses should not be silent or opaque. Users and operators must see why a request was blocked or flagged, with reason, confidence, evidence, and remediations for trust and debuggability.

Problem to solve

Current guardrail/injection/PII blocks are not always explainable. This erodes developer/operator trust and hinders incident learning or auditing.

Proposed solution

  • Structured error messages: rule fired, evidence, confidence, recommended fix
  • Guardrail hit explanations in logs and UI (optionally in user response)
  • Audit log of blocks and triggers (for incident review, see related ticket)

Alternatives considered

  • Silent or minimal errors
  • External custom tooling for log analysis

Who benefits?

  • AI engineers
  • Platform teams
  • Product / operations teams
  • Maintainers
  • Other

Priority / impact

  • Improves trust and adoption
  • Enables root-cause and proactive engineering

Additional context

Best practices: explainable AI/ML, GDPR/art. 22 requirements for explainability.

Metadata

Metadata

Assignees

No one assigned

    Labels

    enhancementNew feature or request

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions