Formal Proof of Governance for AI Workflows Without Loss of Expressivity
A new study published on arXiv (2605.01030) reveals that it's possible to apply effect-level governance to AI workflow architectures without losing their ability to perform complex computations. The researchers introduce a governance operator named G using Interaction Trees in Rocq 8.19, which oversees actions like memory access, external calls, and queries to oracles (LLMs). This work consists of 36 modules and around 12,000 lines of Rocq code, along with 454 theorems and no admitted lemmas. They highlight seven important properties, including governed Turing completeness and oracle expressivity, a boundary for decidability with total governance predicates, and the preservation of goals during permitted executions.
Key facts
- Paper arXiv:2605.01030 presents formal proof of governance for AI workflows.
- Governance operator G mediates all effectful directives including LLM queries.
- Proof uses Interaction Trees in Rocq 8.19 with 0 admitted lemmas.
- Development includes 36 modules, ~12,000 lines of Rocq, and 454 theorems.
- Seven properties established including governed Turing completeness and oracle expressivity.
- Decidability boundary: governance predicates total and closed under Boolean composition.
- Semantic program properties remain non-trivial and undecidable by governance.
- Expressive minimality of primitive capabilities: compute, memory, reasoning, external call, observability.
Entities
Institutions
- arXiv