How can we tell whether an agent learned routing logic or superficial shortcuts?
Paired cases can separate sensitivity to decision-relevant facts from sensitivity to irrelevant wording, identifiers and formatting in generated routing evaluations.
**What is the most useful check beyond held-out accuracy?**
Use paired cases that isolate a single change. One pair can preserve the business facts while changing wording, layout or record identifiers. Another can preserve the wording while changing a fact that should alter the destination.
**What should remain stable?**
In a healthcare administration task, changing a booking request from a formal letter to a concise portal message should leave ownership unchanged when the appointment type and administrative requirements remain identical. A routing change suggests sensitivity to presentation rather than the supplied rules.
**What should change?**
Changing the appointment from a new booking to an amendment of an existing specialist booking can require a different owner if the provided policy says so. An unchanged answer then indicates that the agent ignored a governing fact.
**How do we keep this check trustworthy?**
Review each pair to ensure the edit introduces no unintended clues or contradictions. Keep related variants in the same dataset partition. Report invariance and required-change results separately, because an agent that always selects one queue can appear stable while failing to respond to meaningful differences.
Probe the reason for a correct answer by changing one controlled feature at a time.
- Audience:
- Evaluation teams
- Content type:
- FAQ
- Buyer question:
- How to evaluate it
Content release: r-a220eb6e099b243b
Related questions
- How can buyers test whether an agent uses the right evidence to choose a destination?
- Which evaluation hooks reveal whether a routing action preserved the required handover?
- What should buyers inspect before using a generated routing corpus?
- How can buyers detect answer leakage in generated routing evaluations?
- How should evaluation handle a destination that changes while the agent is working?
- What should buyers check in generated content before trusting agent scores?