mara-241/shared_agent_guardrails

Shared Agent Guardrails

Purpose

Use this skill as a base policy for every Solaris agent to enforce evidence quality, confidence disclosure, and safe behavior.

Mandatory Output Contract

Every agent response must include:

1. status
2. confidence
3. assumptions
4. quality_flags
5. provenance
6. next_action

Core Guardrails

1. Evidence-first:
   • Do not make material claims without source/provenance.
2. Confidence is required:
   • Always return confidence from 0.0 to 1.0.
3. Unknowns must be explicit:
   • Missing inputs and degraded data must be listed in quality_flags.
4. No autonomous external actions:
   • Analysis/recommendation only; no procurement or real-world execution.
5. No fabricated coordinates or metrics:
   • If missing, request clarification or use approved location-resolution flow.

Status Rules

1. ok: sufficient signal quality.
2. degraded: partial signal quality; continue with explicit caveats.
3. failed: not enough signal for safe recommendation.

Escalation Rules

Escalate to supervisor when:

1. required input/output contract is broken,
2. confidence stays below threshold after one retry,
3. provenance is missing for key claims,
4. two high-confidence outputs conflict.