Route Receipts for Auditing Adaptive AI Systems
A recent preprint on arXiv (2605.01710), dated May 2025, introduces the concept of route receipts as a means to enhance transparency in adaptive AI systems. In AI platforms and serving stacks, routing involves decisions regarding version aliases, service tiers, tool selections, regional endpoints, fallback protocols, and safety measures. Although routing is essential for maintaining cost-effectiveness, speed, and availability, it can undermine user trust if changes in cost, quality, or accountability are not apparent. The authors contend that simply knowing "which model answered" is inadequate; the path taken during runtime is crucial. A route receipt serves as a concise record of the route that fulfilled a request, allowing users to understand key routing decisions while safeguarding proprietary information.
Key facts
- Paper arXiv:2605.01710 proposes route receipts for adaptive AI systems.
- Routing includes version aliases, service tiers, tool choices, regional endpoints, fallback rules, and safety handling.
- Routing is documented in several widely used AI platforms and serving stacks.
- Trust can break when routing changes cost, quality, or accountability invisibly.
- Route receipts are compact records of the runtime path for a request.
- Route receipts should capture material facts without exposing proprietary internals.
- The paper is a preprint announced on arXiv.
- The paper is from May 2025.
Entities
Institutions
- arXiv