9.8 Evaluating transformational creativity

Predictive accuracy on the motivating data is necessary but weak: a more expressive theory can often fit what exposed the failure. Evaluation should instead form a vector of criteria:

Resolution

Does the extension remove the localized obstruction?

Minimality

Could a less disruptive change do the same work?

Conservativity

Which earlier results and skills survive?

Novel consequence

Does the extension entail a test not used to propose it?

Discrimination

Can an experiment separate it from rival extensions?

Unification

Does it organize several failures through one construction?

Persistence

Is the new object or operation reused in later tasks?

Calibration

Does the system defer admission when evidence is inadequate?

No scalar combination of these criteria is canonical. The vector should remain visible so that a compact but destructive theory is not silently ranked above a conservative one, or a novel vocabulary above an empirically grounded extension. Unification may reduce description length, but compression alone does not discharge the other admission obligations.

. Admit a transformation only with a versioned package comparison—including a theory map when the presentation changes—a localized obstruction, comparison against in-language repairs, transport account, independent consequence, and explicit epistemic status. A fluent proposal without these components remains a hypothesis, not an admitted creative jump.