Who Gets to Interrupt?
Explore how current importance and urgency claims combine with task-specific historical reliability to allocate interruption authority among specialized AI agents.
Educational diagnostic · deterministic browser modelImportant: This interface demonstrates a proposed heuristic. It does not show that history calibration is better, optimal, safer, or fairer. Hidden outcome labels are not used in the online decision.
Candidate message
Coordinator decisions
History-calibrated policy
S ≥ 0.65 interrupt · 0.35 ≤ S < 0.65 queue · S < 0.35 digest
Same current claim, different historical reliability
The table holds I, U, N, and arrival minute fixed. Only the preset reliability changes, making the allocation of interruption authority directly inspectable.
| Agent history | R | Reputation score | Route |
|---|
Current-claim policy
Responds only to present importance and urgency. Historical reliability is displayed for auditing but does not affect this score.
History-calibrated policy
Discounts the current claim using task-specific historical reliability. Evidence defaults to 0.90, Planning to 0.50, Idea to 0.20, and cold start to 0.50.
Important warning
The multiplicative rule can reduce interruptions, but it can also suppress valid warnings and concentrate interruption authority.
Academic and evidence boundary
- This is an interactive demonstration of a proposed heuristic using authored synthetic examples.
- It is not participant data, real LLM behavior, field observation, deployment evidence, or safety validation.
- Yiqiao Liu selected the research direction and central model idea.
- AI/Codex assisted with implementation and interface construction.