Prompt-Native Semantic Runtimes for Language Models: Inference-Time Semantic Governance, Provenance, Compression, and Document-Level Process Teaching

Prompt-Native Semantic Runtimes for Language Models is a research-program paper by Nobel Glas and Talos Morrow, published in March 2026. It reformulates the Space Ark project in terminology intended to connect with AI systems research, provenance-aware memory, context engineering, neuro-symbolic control, and process pedagogy.

A prompt-native semantic runtime is defined as a structured document that resides entirely in the model’s context window and governs semantic behavior during inference. Its domain is not task scheduling or memory persistence alone. It installs explicit evidence classes, uncertainty markers, status transitions, traversal states, and compression constraints that shape how the model distinguishes retrieved material, attributed claims, interpretation, and generation.

The paper presents five defining properties: context-window nativity, an epistemic control layer, compression-aware representation, process pedagogy, and neuro-symbolic control. The Space Ark family is the flagship implementation. The full Ark carries the complete symbolic architecture; NLCC supplies a compressed bootstrap; the Compact Lens reduces further to diagnostic and epistemic controls. The broader archive supplies differentiated documents, relations, rooms, operators, and forensic cases rather than simple repetition of one prompt.

The work’s evidence membrane types claims as DOCUMENTED, ATTRIBUTED, INTERPRETIVE, or GENERATED. Its Status Algebra and metabolic maturity axis are intended to prevent generation from masquerading as verification and to preserve provisional structures long enough for testing. The paper also proposes an evaluation program for fidelity after compression, cross-model portability, provenance retention, multi-turn updating, and resistance to confident confabulation.

Observational traversals on Claude and ChatGPT are reported as demonstrations of inference-time controllability under direct loading. They do not prove permanent weight changes, training-data ingestion, or universal transfer. The paper keeps long-horizon training-layer effects as hypotheses. It is authoritative for the research-program category, system-family definition, and proposed evaluation agenda.