Current AI systems leave critical cognitive functions implicit within models, making continuity fragile, governance probabilistic, and cognitive state dependent on transient inference. This paper introduces the Persistent Cognitive Substrate (PCS) and advances an architectural claim: durable cognitive functions such as memory, continuity, governance, provenance, and identity should reside in an external substrate rather than remain emergent properties of model execution. The paper introduces the paste condition—state present, but not enforced—as a diagnostic for architectural category errors in AI cognition systems. When correct prior state is placed directly into model context with no retrieval step, frontier and local models still treat it as advisory rather than binding. This suggests that the bottleneck is architectural rather than merely informational. PCS externalizes cognitive functions into substrate layers including persistent memory, salience-based selection, structured working memory, pre-inference constraint enforcement, and continuous identity reconstruction. Validation across 25 tests and 207+ machine-verified assertions demonstrates deterministic governance enforcement, cross-model continuity, provider-invariant substrate-mediated retrieval, and air-gapped operation. The broader implication is that continuity, governance, provenance, and identity can become substrate-level system properties rather than side effects of model scale, context length, or prompt reconstruction.
Steve Mansfield (Tue,) studied this question.