New deterministic FSM runtime treats LLM as bounded compute for enterprise reliability

Most AI agent frameworks let the LLM control loops, select tools, and mutate execution flow—fine for demos but disastrous for regulated environments like KYC, billing, or DevSecOps where auditability and formal reasoning are non-negotiable. A new architecture solves this by embedding the LLM inside a deterministic Finite State Machine (FSM), stripping it of orchestration control. The runtime governs transitions, governance topology, and execution graphs; the LLM only computes bounded steps. This prevents the model from seeing governance metadata or policy internals via a ProjectionLayer ("evaluator blindness"), avoiding semantic contamination and adaptive behavior.

Beyond context isolation, the runtime uses a constrained AST engine for condition evaluation—no arbitrary Python exec() or dynamic method calls—limiting semantic surface area like Rego or IAM policy DSLs. Transition entropy monitoring detects structural instability (e.g., path variance >2.5 bits) and flags unstable execution. The accompanying failure laboratory includes deliberate attack injectors (tool injection, GDPR erase simulation) to test adversarial failure handling. A development agent follows the same governed principles with staged, validated code mutations. All 51 tests pass with zero mypy errors. The project challenges whether autonomous agents are the wrong abstraction for production systems.