telos-gov

telos-gov is a governance engine designed to enforce runtime oversight for AI agents, ensuring that their actions remain within defined boundaries through mathematical enforcement rather than mere instructions. The tool addresses the challenge of keeping autonomous agents aligned with specified purposes by scoring every tool call in real time and providing cryptographically signed receipts for each action. It operates entirely locally, with no cloud dependency, so all governance processes and audit trails remain on the user's machine.

The platform features a four-layer scoring cascade, incorporating keyword analysis, machine learning, embedding techniques, and contrastive evaluation to assess agent actions. Each action is evaluated in under 20 milliseconds, and the resulting governance receipts are signed using Ed25519 cryptography. telos-gov also maintains a hash-chained audit trail, retaining governance records for 90 days, and includes drift monitoring with CUSUM alerting to detect deviations from the intended agent behavior. Additional features include enforce mode to block unsafe actions in real time, authoring and commissioning of "Purpose Attractors" (which define agent boundaries), and counterfactual replay validation (CCRS) for retrospective analysis.

To begin using telos-gov, users define the agent's Purpose Attractor, sign it with an Ed25519 key to activate governance, and then all subsequent agent actions are scored against this purpose. The tool supports real-time enforcement, clarification, or escalation of actions, and automatically generates compliance evidence through its audit trail. Performance metrics indicate low false acceptance rates on safety benchmarks and high boundary detection rates, with transparency about areas where accuracy is still being improved.

telos-gov is distributed as a locally executed package, installable via pip, and can be initialized with a provided token. It is offered as a subscription service at $49 per month. The tool is positioned as a solution for those seeking rigorous, mathematically enforced governance for AI agents, and is suitable for both individual and enterprise use cases where agent safety and compliance are critical.