Type: article
Abstract: Long-lived reasoning systems face a continuity problem that is distinct from persistent storage, retrieval, or context projection. A system may preserve an authoritative state outside the model while still failing to make the consequential meaning of that state operational for the next reasoning instance. We call this semantic-state reinstantiation: a task-conditioned behavioral property of an execution protocol, defined as the extent to which a fresh reasoning instance can recover and use predefined consequential relationships from externally represented state. IEM studies a governed implementation of this more general research object. Persistence asks whether state remains available; reinstantiation asks whether its consequential meaning becomes operational again in a fresh reasoner. We present the Implementation and Execution Model (IEM), a governed semantic-orientation architecture that externalizes semantic state from any particular reasoning instance. IEM represents evidence and interpretations under provenance and authority, distinguishes exploratory and provisional interpretations from governed standing, compiles task-conditioned executable context, and supports scoped correction while preserving unaffected history. The paper does not claim invention of persistent memory, provenance, temporal state, graphs, belief revision, runtime-independent persistence, authorized state, or context projection individually. Those functions have substantial prior art. The primary publication contribution is the operationalization and measurement of semantic-state reinstantiation across fresh reasoning instances, with IEM presented as one governed implementation. We evaluate the architecture through the Northstar experimental sequence. In a 30-day longitudinal synthetic corpus, both IEM and cumulative-Raw conditions preserved the audited consequential distinctions while IEM used 87.9% fewer input tokens. A prospectively scored context-pressure experiment again found no semantic separation - 120/120 distinctions were preserved in both conditions - but IEM used 59.25% less aggregate input context, reaching a 74.04% reduction at the highest checkpoint. This negative result is retained as evidence that acumulative Raw remained semantically adequate within the tested range. A later bundled runtime-completion comparison evaluated the complete adaptive IEM protocol against a one-pass complete-Raw baseline. The IEM condition tested whether a fresh reasoner operating on task-conditioned context could detect semantic insufficiency, request the smallest supported governed context reconstruction permitted by the protocol from a separate Context Engine, discard the incomplete reasoning instance, and continue through a fresh reasoner. Under the frozen 45-unit scoring key, the IEM runtime preserved 44/45 units (97.78%) and the complete Raw-source comparator preserved 26/45 (57.78%), a 40.00 percentage-point difference. Both conditions preserved all 15 upstream units, while downstream preservation was 29/30 for IEM and 11/30 for Raw. Blinded scorers agreed on 89/90 judgments before one third-scorer adjudication. The bounded result is not that IEM had information unavailable to Raw: Raw received the complete source corpus. Rather, the experiment records a task-specific behavioral separation between state availability and downstream use of predefined consequential relationships under two different execution protocols. Because v0.776 compares the complete adaptive IEM protocol with a one-pass complete-Raw baseline, it does not isolate component causality or establish superiority over rolling summaries, graph/provenance projections, matched adaptive Raw pipelines, or other compiled-state architectures. Paper 1 therefore treats semantic-state reinstantiation as a separate research object from persistence, retrieval, provenance, belief revision, and context delivery.
Author: Webb, Nicholas
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7212210510
