Understanding and Developmental Uptake Are Not the Same Judgment: A Dual-Axis Diagnostic Framework for Artificial Systems

Debates over whether an artificial system “understands” often oscillate between wholesale denial and unqualified attribution. Debates over whether a persistent agent “learns from experience” are similarly prone to treating context access, memory retention, or behavioral change as development. Both debates conflate judgments made at different timescales and over different evidentiary units. A system may use target relations to handle novel combinations and counterfactuals during a current interaction without allowing that episode to change its future. Conversely, an episode may establish a persistent and revisable action constraint without supplying independent evidence that the system understands the mechanism at issue. This paper proposes a dual-axis diagnostic framework designed to block cross-axis inference errors. It does not offer a unified definition of understanding or development. Instead, it separates the evidentiary burdens for current target understanding and experience-specific Developmental Uptake and specifies that evidence for either judgment may not license the other without additional tests. The first axis concerns current target understanding. For a system S, target T_U, and predeclared boundary B_U, it instantiates a candidate functional account through four conditions: relational organization, causal use, non-accidental generation, and accuracy and scope calibration. The second axis concerns the Developmental Uptake of a target experience e_D. Using the previously proposed Developmental Uptake Criterion, it asks whether the experience-specific effect is selectively used according to its current standing, preserved across standing-equivalent surface variants, reorganized after counterevidence, recombined in new situations within its scope, and causally attributable through experience-level interventions. To prevent trivial dissociations between unrelated objects of judgment, any comparison must first pass an outcome-blind target-coupling gate. Researchers must predeclare the minimum understanding subtarget T_U^e that e_D actually bears on, and the uptake endpoint must fall within the judgments or actions that this relation could affect. Only after this admission condition is satisfied do the axes yield a testable hypothesis of diagnostic non-entailment. A shared thermoregulation microworld illustrates four candidate evidence combinations and a 2 × 2 cross-diagnostic protocol that separately manipulates current relational organization R and cross-episode experience update M. Each axis uses three evidence states: SUPPORTED, VALID-TEST-FAILED, and INCONCLUSIVE. Only the first two may enter the four quadrants. No experimental results are reported, and the two axes are not claimed to be ontologically independent. This is a conceptual diagnostic framework and evaluation agenda. It has not been peer reviewed.