Should a model treat a missing record as proof that a control failed?

**Should a model treat a missing record as proof that a control failed?**

No. The appropriate conclusion depends on what the model knows about the evidence source. A search that returns no approval is different from a complete approval register showing that no approval was recorded.

**What should the model say instead?**

It should describe the supported finding precisely. In healthcare administration, an appointment change may require a notification record. If the notification archive is unavailable, the evidence position is unresolved. If an authoritative event history explicitly records that notification was skipped, the model has a basis for a different conclusion.

**Why does this matter for training?**

Models can learn to fill gaps with plausible process narratives. They can also overcorrect and report every missing attachment as a control breach. Both behaviours weaken an audit bundle because they replace the evidence position with an assumption.

A generated task pack could separate records that are absent, withheld, unavailable or explicitly negative. The expected response can require an appropriately qualified conclusion and a targeted next request, allowing labs to assess uncertainty as an operational skill.