White-Box State-Transition Governance for AI Agents: Candidate-First, Evidence-Bound Control of External Side Effects

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

AI agents can now invoke tools and initiate changes outside a chat interface. This technical research note proposes White-Box State-Transition Governance: a candidate-first, evidence-bound runtime model in which a proposed external change is evaluated as a bound lineage of evidence, pre-state, candidate transition, policy snapshot, governance decision, controlled execution boundary, observation, and replayable record. The accompanying public reference demo is deliberately minimal and deterministic. It demonstrates the stated governance semantics with fixed in-memory cases and makes no claim of operating-system isolation, real external execution, production readiness, or access to a model's hidden reasoning. The public reproducible implementation is fixed at GitHub commit 0e1853dfba6653ea5c7e48f40ef46bba704cb3ba:https://github.com/wan593315006/white-box-state-transition-governance-reference-demo/tree/0e1853dfba6653ea5c7e48f40ef46bba704cb3ba The paper also describes protected engineering context and intended directions for the Five-Body system. Those statements are bounded context, not public experimental claims. No protected Five-Body source, internal runtime state, method assets, credentials, or production configuration is included in this record.