Establish empirical criteria for identifying primitive computations in artificial systems

Establish empirical criteria and practical tests that distinguish genuine computation of the six Computational Primitives Theory primitives in artificial systems from behavior that merely imitates or resembles those computations.

Background

The paper argues that the six primitives must operate as an integrated system and predicts that they cannot simply be engineered as independent modules. Testing this prediction requires determining whether an artificial system genuinely computes a primitive rather than producing an input-output pattern that resembles one.

Because the primitives are defined at the computational rather than mechanistic level, surface behavior or benchmark performance may not establish that a primitive is present. The paper suggests examining dependence on internal state, generalization beyond training data, and conditions of the system’s own persistence, but presents these only as a first approximation rather than a definitive test.