RFC: Trust Score Algorithm - Seeking Expert Critique #2
Description
TRAIL assigns trust scores to AI agent DIDs based on verification depth, credential history, and endorsements. We need critique on:
- Is the scoring model gameable? - Can an attacker inflate trust through automated credential farming?
- How should trust decay over time? - Static scores vs. time-weighted, activity-based decay curves.
- What's the right balance between automated and human-verified trust signals?
Edge cases to consider:
- Newly created agents (cold start problem)
- Agents migrating between platforms
- Revoked-then-reinstated credentials
Background: TRAIL is designed for EU AI Act compliance - trust scores must be auditable and explainable.