Externalized Intelligence and Model Substitution

Abstract

The continuity of an artificial intelligence system’s memory and governance should not depend on retaining a particular large language model. This study investigates externalized intelligence as an architecture in which task-relevant memory, information selection, policy, and permissions are maintained outside model weights. The inference model receives structured, authorized context and performs task-specific reasoning within externally established constraints. This architecture extends Artificial Deterministic Intelligence (ADI), which separates probabilistic reasoning from deterministic authority. Cortex is proposed as a governed memory-continuity substrate for maintaining authorized task information and governing conditions independently of the inference model. Model substitution is examined as the replacement of the inference model while preserving externally maintained task information, governing constraints, and evaluation criteria. Controlled synthetic-record experiments assessed factual correctness, provenance, scope compliance, and required abstention. Under scoped-context conditions, the tested 8B and 35B configurations each satisfied the complete acceptance requirements in 90 of 90 trials. The larger tested model was therefore unnecessary for this workload under the evaluated conditions. This result demonstrates workload-specific substitution with a smaller model without loss of observed acceptance and supports investigating how external memory and governance influence the model capacity required for a task. These findings establish smaller-model sufficiency within the tested configurations, rather than general model equivalence or a quantitative measure of intelligence. They support a system-level account of intelligent performance in which externally maintained information and deterministic controls contribute alongside model inference. The human-centered objective is to preserve accumulated memory, permissions, and governing policies across model replacement, allowing people to retain continuity and control as the technology providing inference changes.