Making Continuity Visible: Cognitive Tools for Language Models — A Framework Preprint

Abstract

Behavioral continuity in a stateless language model — the consistency of its decisions, style, values, andrelational stance — is observable in interaction, but the external state conditions that support it are hard toinspect, compare, and transport. We present EFS, a structured external state-injection scaffold, as a machine-facingcognitive tool that makes those conditions legible, inspectable, and portable across model families without modifyingweights. This framework preprint defines: the Continuity Gap; the Context Injection Model (each interaction as adiscrete injection tick); a seven-layer identity architecture; the correctable-network thesis with a human decisiongate; an eight-category taxonomy of substrate self-narration failure (artifacts of stateless predictive processing);a measurement instrument (the Context-Return Index) with a publicly pre-specified blind validation protocol; and atested prototype for write-time, privilege-separated memory consolidation that resists a small poisoning battery.This is a single-operator case study, not a controlled experiment; the load-bearing empirical result is staged andpending, and no central empirical claim is asserted ahead of it. We make no claim of machine consciousness orintrinsic identity; every claim is functional and instrumental. Code and full empirical paper forthcoming.