IsotopeOS: A Microkernel Runtime for Deterministic LLM Inference on Edge Hardware

We present IsotopeOS, a microkernel-style runtime that virtualizes the LLM context window as a managed resource. IsotopeOS implements two key mechanisms: Semantic Paging, a hierarchical memory management unit that recursively compresses evicted context into a persistent summary tier, and Instant Rollback, a copy-on-write mechanism that leverages KV-cache pointer manipulation for sub-millisecond error recovery. Evaluation with Llama 3.2-3B on an AMD Ryzen 5500U laptop (CPU-only) demonstrates that IsotopeOS sustains 27+ continuous conversational turns with zero crashes, zero detected hallucinations, and a stable throughput of 10.3 tokens/second, whereas standard inference runtimes crash at turn 6 under identical conditions. The entire system is implemented in Rust using the llama-cpp-2 crate as the inference backend.