The hierarchy matters
Learned models estimate uncertain future outcomes. Deterministic logic reconciles evidence and protects hard constraints. Optimisation compares feasible alternatives. Workflow establishes authority and completion. Measurement reconciles outcomes.
Changing that order can create model decoration: predictions are generated after an action has already been selected and cannot materially influence the work.
Reference and tenant evidence
A reference product should execute the same mapper, artifact resolver, inference, optimisation, persistence and API path intended for a tenant. Fictional data can prove mechanics and decision materiality inside its declared scope.
It cannot prove live performance, realised client uplift or external superiority. Tenant mapping, calibration, approval and shadow evidence remain necessary.
Safe failure is visible failure
Expired artifacts, stale sources, out-of-envelope inputs and unavailable integrations must activate explicit fallback or abstention. Silent substitution destroys operator trust and makes later measurement meaningless.
The operator surface should expose the consequence in business language while keeping technical evidence available for stewardship.