GenOS: A Probabilistic Semantics for Safe AI Code Generation
A new paper on arXiv (2608.03588) introduces GenOS, a probabilistic operational semantics designed to ensure semantic robustness in AI code generation. The research addresses the challenge of safely replacing prompts, contracts, generators, or programs within agentic workflows, where small changes can alter program behavior. GenOS models each layer as a Markov kernel and introduces observer-relative equivalence to prove that equivalent prompts yield equal probabilities for downstream events, including verified commits. The paper is categorized as a cross-type announcement and focuses on the theoretical foundations of compositional certification for AI coding agents.
Key facts
- GenOS is introduced as a probabilistic operational semantics for AI code generation.
- The paper is available on arXiv with ID 2608.03588.
- It addresses the problem of safely replacing components in agentic workflows.
- Each layer is modeled as a Markov kernel.
- Interfaces carry observer-relative equivalence.
- The paper proves that equivalence-compatible kernels descend to quotient classes.
- Quotienting commutes with distributional extension and sequential composition.
- Equivalent prompts induce equal probabilities for all downstream equivalence-closed events, including verified commit.
Entities
Institutions
- arXiv