21.5 A minimal experiment contract

A credible experiment separates four questions:

1. Does the registered probe actually change the declared obstruction?

2. Can the method localize that change without reacting to null controls?

3. Does the repair improve held-out structural audits?

4. Does the repaired system preserve the task semantics and operational constraints?

Retained failures are part of this contract. The LINCS-KET joint-norm result, for example, explains why blockwise admission exists. A negative experiment that identifies a missing gate contributes more to the pattern language than an unqualified reduction in a training residual.