Suppose an order service must accept a caller-supplied rule deciding whether to release an order or hold it for review. Pass the rule explicit inputs such as the order amount and identity-check result. Have it return decision data containing the selected action and a reason. The host can inspect that result before deciding which operation to perform.
In this hypothetical integration, a returned release decision does not establish that the order was released. The host might reject an unknown action, apply a later authorization check or update the wrong order. Recording the input and returned data helps locate a disputed decision. Establishing the resulting action requires a separate observation of the host.
The evaluator matters as much as the rule interface. In Lispex 1.21, default source evaluation runs trusted code in-process. That is a reasonable choice when the supplied source is trusted and bounded execution is unnecessary. It is not a resource guarantee for arbitrary caller source.
Native embed and embed full are distinct bounded evaluators using import-free WASM and fresh instances. A resource claim must identify which evaluator and profile supply it. Preparation and evaluation also need separate accounting against the chosen profile. A limit stated for evaluation alone does not describe all preparation work.
For this service, I would record the evaluator, profile, explicit input and returned decision together, then associate any observed host action separately. That keeps a rule's calculation available for inspection without treating its output as proof of enforcement.