Deterministic runtime governance for agentic AI. A multi-stage policy pipeline, per-claim verification, signed policy bundles and tamper-evident audit evidence, enforced at every step of the agent loop.
Critical policies are enforced outside probabilistic LLM reasoning: same input, same decision, every time. The policy pipeline runs rule match → graph lookup → plausibility → contradiction, per claim — not per response.
Domain rules authored by compliance experts as configuration, matched deterministically against extracted claims.
Three gates on every tool call: well-formed, policy-permitted, semantically plausible.
Workflow state machines that prevent step-skipping and enforce required behavior.
Policy bundles are signed, support multi-signature approval, and changes are attributable.
Runtime monitors watch agent behavior for drift against its validated baseline.
Tamper-evident logs as queryable triples, with integrity verification on retrieved facts.
Formal policy verification — proving a policy set is internally consistent and complete before it is enforced — is under development as a Governance capability. Consistent with our claims policy, it will appear here as available only when it ships with evidence.
Governance ships inside the knowlytix package. Install it and everything below works now.
# policy engine, contracts, tool gateway, drift monitoring, audit import knowlytix.harness.governance
