Trust Isn't a Scalar: Typed Provenance for Agent Chains

Short summary

Boolean trust scores fail in agent chains because different consumers care about different degradation axes—a summarizer tolerates weaker models but needs fresh data, while price calculation needs accuracy but tolerates staleness. Instead propagate typed provenance vectors with per-axis scores, let each consumer apply its own thresholds. Design pattern converging across the field; TrustBench framework makes the same move independently.

• •Single scalar trust flags collapse information needed by downstream consumers with different requirements
• •Propagate typed provenance vectors with per-axis scores (freshness, capability, tool success, verification)
• •Use min-composition across axes to preserve dangerous signals; let each consumer apply its own policy

Generated with AI, which can make mistakes.