This work develops an observable-only institutional framework for structural risk in competitive AI environments under a strict no-meta constraint: only public, auditable records are admissible, and no privileged central arbiter is assumed. The model is formulated as a repeated game on reachable public histories and provides deterrence conditions against both unilateral and coalition deviations within a unified framework, including a bounded hidden-channel extension. The manuscript proves a non-domination guarantee from explicit solvency dynamics, escrow-based reversibility, and contestable challenge procedures, with compatibility for non-monetary operational throttles. It further introduces finite-sample, one-sided certification under heavy tails and dependence (via big-block/small-gap coupling), with explicit confidence accounting and additive inflation terms for bounded record corruption. Certified structural inequalities are translated into a genuinely linear calibration layer (context-lifted MILP), with stated sufficiency conditions for strict deterrence under approximate column generation and extension beyond tested coalition templates under coverage conditions. The paper also provides voluntary-adoption incentive conditions, enabling self-adoption by no-meta agents rather than post hoc compliance only. Overall, the contribution is an end-to-end, verifiable design for decentralized safety institutions that remain implementable without privileged internal-state access or centralized arbitration.
K Takahashi (Tue,) studied this question.