CLOVER: 検証可能な実行と推論のためのコンテキスト管理OS

抄録

Modern large language models possess rich latent representations capable of supporting structured reasoning, yet their interaction environment provides no mechanism to preserve or govern it.
Model-execution frameworks such as llama.cpp standardize inference but do not organize context, leaving most LLM usage bound to stateless, prompt-centric interfaces. This mismatch leads to familiar failures in long-term coherence, including semantic drift and unstable multi-turn reasoning.

We identify the missing architectural layer between model execution and user-facing applications: a context operating system. To address this gap, we propose CLOVER (Context Layered OS for Verified Execution & Reasoning), a four-layer system that externalizes structure rather than modifying model internals. CLOVER integrates a hierarchical context model, block-based memory with provenance tracking, anomaly detection for contextual drift, priority retrieval, and a session state machine governing interaction flow.

This architecture enables predictable, reproducible, and model-agnostic long-term reasoning, providing a stable foundation for coherent and
interpretable AI systems.