openclaw - 💡(How to fix) Fix [RFC] Agent truthfulness and confidence signaling for verifiable trust [1 participants]

TL;DR

Implementing confidence tiers and source attribution for agent output can help distinguish between known and inferred statements, improving trust in autonomous mode.

Guidance

• Introduce explicit confidence tiers (known, believed, uncertain, unknown) for every factual claim made by the agent.
• Add source attribution (e.g., file+line for memory, tool name for tool outputs) to provide context for each claim.
• Develop a clear distinction in language for known vs. inferred statements, such as using "I know X because..." vs. "I believe X because...".
• Consider adding an autonomous mode flag to surface unverified claims separately in completion reports for review.

Example

// Example output with confidence tier and source attribution
- Claim: The capital of France is Paris. (Confidence: Known, Source: Geography Database)
- Claim: The weather tomorrow will be sunny. (Confidence: Believed, Source: Weather Forecast Tool)

Notes

The proposed solution requires significant changes to the agent's output formatting and language generation. It may also require updates to the underlying knowledge graph or fact storage to support confidence tiers and source attribution.

Recommendation

Apply workaround by implementing confidence tiers and source attribution to improve the reliability of agent assertions in autonomous mode, as this directly addresses the trust limitation mentioned.