Every intervention should identify the rule, evidence graph, uncertainty, affected scope, and available recourse.
Why this question matters
A risk score without explanation cannot support debugging, appeal, or policy improvement. Multi-agent systems make explanation harder because no single action may be forbidden. The governance decision depends on a pattern across participants and time.
Evidence-linked governance presents the minimum subgraph that supports the decision and distinguishes observed facts from inferred relationships. It should also record which policy version and threshold were active.
Signals worth observing
- Operators receive only a score or generic safety label.
- Inferred relationships are presented as established facts.
- A policy cannot be reproduced from retained evidence.
Practical control direction
- Store rule, version, features, and evidence references together.
- Label observations, inferences, and confidence separately.
- Provide appeal and re-evaluation paths for consequential decisions.
AgentCollusion lensExplainability is not decoration; it is the interface between collusion research and operational governance.Sources and further reading
Next field note: Incident Response Must Follow the Agent Graph
