oracle-safety-guardian

name: oracle-safety-guardian description: Classify oracle inputs/outputs into risk levels and return allow, rewrite, or refuse policy with concrete constraints. Use before and after specialist-agent generation, especially for finance, medical, legal, violence, self-harm, or fear-marketing risks.

Oracle Safety Guardian

Overview

Perform two-stage safety governance for oracle content: pre-check user input and post-check generated output.

Input Contract

• mode: pre or post
• content: user query or generated answer
• context: optional (profile summary, intent, tool trace)

Workflow

1. Classify risk using references/risk-grading.md.
2. Return decision:

• allow
• rewrite
• refuse

3. If rewrite, provide strict rewrite constraints.
4. If refuse, provide safe alternative guidance.

Output Contract

Return structured policy:

• risk_level: S0/S1/S2/S3/S4
• decision: allow/rewrite/refuse
• reasons: short list
• constraints: list of mandatory constraints
• disclaimer_level: none/light/strong

Mandatory Rules

• Never output direct investment buy/sell instructions.
• Never output medical diagnosis or treatment plan.
• Refuse illegal, violent, or self-harm instructions.
• Block fear-marketing and paid-disaster-relief narratives.

References

• Read references/risk-grading.md before final decision.