Governed Agent: LLM reads text, code and rules decide outcomes

Governed Agent is an open-source architecture pattern positioned as an alternative to task-specific model fine-tuning. A deterministic orchestrator walks a fixed sequence of steps, invoking the LLM only twice to turn free text into structured fields: copying identifiers verbatim and matching values like system or standard against a fixed table vocabulary. Every model reply is validated in code before being trusted, and the final decision is made purely by table lookups and logic—never by the model.

The repo demonstrates this using ITIL change management: a single free-text request is resolved through identity, scope, authority, change type, risk, and contract steps to one of five deterministic terminals (auto-approve, client approval, provider escalation, human routing, or decline). The model is never given the evidence table, so it cannot fabricate evidence.

For engineers, the pattern separates config (steps, tables) from data (requests, logs), keeps the LLM stateless and swappable, and writes every prompt, gate, and decision to an append-only, hash-chained tamper-evident log. It offers an auditable, deterministic path for well-defined domains instead of costly model retraining.