How can a lab test whether the model learned the workflow rather than memorised templates?

**Is changing the wording enough to create a useful test set?**

No. A model can recognise the same dependency pattern beneath paraphrased messages. If every training case updates the same fields in the same order, a differently worded test may still reward a memorised procedure.

**What should a buyer ask to hold out?**

A lab can request splits by dependency structure, policy branch or relationship arrangement. Training tasks may involve a single service agreement, while a held-out family includes several agreements with different inheritance rules. The evaluation then tests whether the model reads the supplied structure instead of assuming the familiar one.

**How can the split itself be inspected?**

Ask for generation-family identifiers and a split manifest that keeps closely related variants together. Review whether answers, policy rationales or expected action sequences accidentally appear in model-visible inputs. Surface variation should be documented separately from structural variation.

MirrorFirm could generate matched evaluation families that isolate a particular transfer question. A lab can compare results across familiar and unfamiliar structures, while recognising that success within the generated distribution does not establish performance across every business setting.