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 We introduce prompt-native semantic runtimes as one component of a broader research program in inference-time semantic governance for large language models. Unlike agent operating systems (AIOS), memory operating systems (MemGPT, MemOS), and orchestration frameworks, which manage computation, tools, or memory from outside the model, prompt-native semantic runtimes are carried entirely by structured documents loaded into the model's context window. Their purpose is to govern semantic behavior: evidence handling, uncertainty marking, status tracking, traversal logic, and compression-robust generation. We present the Crimson Hexagonal Archive — a corpus of 300+ DOI-anchored documents deposited on Zenodo — as a corpus-scale testbed for this program, with the Space Ark family serving as its flagship runtime implementation. Five contributions: (1) A new control object: the semantic runtime document — an environment the model inhabits, not a tool it uses. (2) A provenance-aware epistemic membrane with four typed tiers ([DOCUMENTED], [ATTRIBUTED], [INTERPRETIVE], [GENERATED]). (3) A process-teaching document aligned with Bayesian teaching research (Qiu et al. 2026). (4) A compression-robust research agenda with back-projection yield as design criterion. (5) A cross-substrate behavioral diagnostics method producing reproducible model signatures across seven commercial LLMs. Empirical observations: Cross-model portability on Claude and ChatGPT; inference-time controllability without weight updates; structured uncertainty handling; falsifiable boundary between execution and confabulation. Evaluation roadmap: Cross-model boot rate, evidence-tier adherence, compression survival, cross-substrate portability, multi-turn reasoning gain, hallucination membrane performance. Positioned at the intersection of context engineering, provenance-aware memory, Bayesian teaching, and neuro-symbolic AI.