Deterministic AI Harnesses: Bounding Probabilistic Drift, Statechart Oscillation, and MCP Execution Risk in Autonomous Systems

Executive Abstract: Production artificial intelligence agents fail when software architects wire unconstrained probabilistic large language models (LLMs) directly into enterprise APIs, shell environments, and databases using linear prompt-and-retry loops (ReAct). Across multi-step execution trajectories, per-step token context grows quadratically while compound trajectory reliability decays exponentially. This working paper introduces the 5-Plane Cognitive Runtime Architecture and formalizes three non-negotiable engineering invariants from Beyond the Prompt: The Architecture of Autonomous Systems: (1) 7-Stage Directed Statecharts with Sliding-Window Oscillation Damping, (2) Prefix-Pinned Context Compaction for KV-Cache Stability, and (3) Rootless OS Process-Group & Network Namespace Sandboxing for Model Context Protocol (MCP) Servers. Companion Volume & Open-Source Code: Book (Amazon Kindle / Print): https://www.amazon.com/dp/B0HJZ1BY9P (ASIN: B0HJZ1BY9P / Paperback: B0HJZ9TGP5 / Hardcover: B0HK1CNB5Y) Open-Source AST Harness Auditor & MCP Server: https://github.com/digbijoylabs/beyond-the-prompt Publisher Portal: https://digbijoylabs.com