Governing AI Actions, Not Agents, Through Institutional Attestation.

As autonomous AI agents begin to undertake critical, irreversible actions like clinical prescribing or software deployment, a new governance model is needed. This paper introduces "Institutional Attestation," a framework inspired by how human institutions govern powerful autonomous actors by focusing on the point of consequential action, rather than attempting to monitor internal reasoning. Under this model, an AI agent maintains full autonomy over its planning and reasoning processes. However, it lacks direct execution authority for designated high-risk actions. Instead, execution is contingent upon specific preconditions, each independently verified and attested by a separate, authoritative source. These attestations are cryptographically linked to the agent's declared intent and evaluated against a deterministic policy. All decisions and attestations are recorded in a tamper-evident log, allowing for independent re-verification. The paper includes a proof-of-concept implementation and illustrates its application in scenarios such as software deployment and medical prescribing, offering a robust, auditable approach to AI governance.