The order of execution matters
Models estimate uncertain future outcomes. Fixed logic reconciles records and enforces hard constraints. Optimisation compares the feasible alternatives. Workflow controls approval and completion. Measurement reconciles the result.
If the model runs after the action has already been selected, its prediction cannot affect the work. It is decoration, regardless of its accuracy.
Controlled product evidence and client evidence prove different things
A controlled product run should use the same mapping, inference, optimisation and operating path intended for a client. It can prove that the software executes correctly and that model output can change the decision within the stated scope.
It cannot prove accuracy on a client’s operation, realised financial improvement or superiority in the market. Those claims require client mapping, calibration, approval and observed results.
Failure must be explicit
An expired model, stale source, unusual input or unavailable integration should trigger a visible fallback or abstention. Silent substitution makes the decision impossible to interpret and the result impossible to measure.
The operator needs the business consequence in plain language. Technical evidence can remain available to the people responsible for maintaining the engine.