Fail-Closed by Design: SHACKLE Releases SP/1.0 Conformance Standard for AI Agent Runtime Governance
Key Takeaways
- ▸SHACKLE introduces SP/1.0, a verifiable conformance standard for AI agent runtime governance with three decision outcomes (ALLOW, DENY, HITL) and a mathematical model for validity
- ▸The standard includes a reference Python implementation and 15 independently-reproducible conformance test vectors enabling independent verification of compliance
- ▸SP/1.0 enforces fail-closed enforcement where recorded actions are never proof of authorization, preventing agents from automatically resuming denied actions
Summary
SHACKLE has published SP/1.0, a conformance standard for runtime mediation of autonomous AI agents. The standard defines three decision outcomes (ALLOW, DENY, HITL) and a mathematical conformance model where a transition is valid if everything required is within what the system provably supports.
The standard includes a reference Python implementation with 15 hash-verifiable conformance test vectors that can be independently verified, moving agent governance compliance from assertion-based to verification-based. Systems can now be independently certified as SP/1.0-conformant through reproduction of the published fixture set.
The core innovation is enforcing fail-closed governance: recorded actions are not proof of authorization, so previously-denied actions that resume must be re-evaluated, not automatically allowed based on history. This addresses a critical vulnerability in existing agent frameworks and transforms safety governance from informal best practices to formally auditable compliance.
- Systems can now be independently certified as SP/1.0-conformant, enabling auditable governance and trust in autonomous AI agent behavior
