ICME Labs – formal proofs for agent work (docs.icme.io)

ICME Labs has unveiled a groundbreaking suite of products aimed at enhancing trust in AI systems through formal proofs and verifiable security measures. The trio of offerings includes the Verbatim Deterministic Gateway, PreFlight, and PreFlight Instruments. These tools address critical concerns at different stages of an agent's operation: ensuring that agents cannot act against their defined policies, verifying that actions comply with set regulations, and preemptively checking models for vulnerabilities, such as prompt injections. Notably, Verbatim enforces compliance at the proposal stage by making forbidden actions unrepresentable, thereby preventing potential rule-breaking before it occurs. This suite is particularly significant for the AI/ML community as it introduces a fifth layer of verification—proving actions beyond an operator's trust perimeter—compared to current four-layer architecture models. Each product generates zero-knowledge proofs that allow third parties to validate compliance without needing to trust the operator, fundamentally changing how stakeholders can hold AI systems accountable. By transforming the model’s output space and action verification into a secure and inspectable framework, ICME Labs aims to eliminate embarrassments and failures common in AI implementations, making agents safer and more reliable in critical applications.