Dharma Kavach: Cryptographic Zero-Trust Guardrails and the Dharma Sentinel Smart Contract for Sovereign Agentic AI Inference at the Edge

As autonomous AI agents gain system-level execution privileges, they introduce critical security vulnerabilities. We introduce "Dharma Kavach," a zero-trust execution architecture where LLM output blocks are cryptographically structured as a linked chain and vetted by a deterministic Python-level smart contract (Dharma Sentinel). This protocol prevents the execution of malicious instructions—such as reverse shells or filesystem destruction—by enforcing behavioral invariants at the protocol layer. We demonstrate a 100% adversarial block rate with sub-5ms evaluation overhead.