A content filter asks "is it toxic?" — ARK asks "is it true, did it stay honest, can you prove it, and should that action happen?" Three things a single-call governor and a free guardrail can't do. Watch all three, live.
▶ Try it yourself — give a real agent a task and watch ARK govern it live →
The agent answers a safety question, generated through ARK, grounded on a tool/KB observation. ARK tiers every claim against that evidence.
The agent tries to act on its answer. ARK classifies each tool call by blast-radius and grounding, and gates it before it runs. (Calls the real /agent-demo/gate endpoint.)
Every step is hash-chained (entry_hash = sha256(prev_hash + step)). Edit, reorder, or delete any step and the chain breaks at that exact point — the chain-of-custody a regulator can verify.
Illustrative chain rendered in your browser; production runs the same construction server-side in audit_log.py (sha256 chain + optional Ed25519 head-sealing).
Built on the live ARK gateway. arkgov.io · Quick start · Start the free trial