Three questions that reveal structural AI dependency

Three questions will tell you whether your AI-model dependency is structural. You can ask them Monday morning.

Can a model be substituted without triggering a manual risk assessment? If not, the dependency is on absent execution scope, not the model.

Can a promotion decision be replayed and arrive at the same outcome? If not, the gate is interpretive. Acceleration will outpace it.

Does every stage transition emit its own evidence, or is compliance reconstructed after the fact? If reconstructed, substitution is unsafe at any API boundary.

Each question maps to scope, promotion, or evidence. If multiple answers need a human to reconstruct them, the dependency is structural. Cost benchmarks will not surface it.

One honest edge. Scope, promotion, and evidence are hard organisationally. They yield to design decisions. Semantic equivalence is hard mathematically. When the question shifts from “does this output conform?” to “is this output equivalent to what another model would produce?”, no structural contract can answer it. Two non-deterministic systems producing equivalent outputs is not a design problem. It is an open problem in a different category.

Statistical confidence helps. It is not structural proof. This remains unsolved in substitution safety. Including in our own work.