How can a lab tell whether a model learned routing logic or generation artefacts?

**What shortcut signals should we look for?**

A destination may accidentally correlate with message length, a record identifier prefix or a repeated phrase. A model can exploit those cues without understanding why a team owns the request.

**How can we test that?**

Hold the business facts constant while changing wording, formatting and identifiers. The acceptable route should remain stable. Then hold the wording largely constant while changing a fact that determines ownership. The model should respond to the changed fact.

**What belongs in a split audit?**

Inspect whether closely related generation templates, request variants or shared case families cross training and evaluation boundaries. A nominally separate evaluation file may still contain near copies of patterns encountered during training.

**What should a buyer ask MirrorFirm to support?**

A pilot could include variant-family identifiers and controlled transformations, making these checks possible without exposing hidden generation machinery. Buyers can also request held-out policy structures to probe transfer. Such tests provide evidence about robustness within the defined task scope; they do not establish performance across every business setting.