TRACE live

Runtime assurance for AI agents

Every call looked innocent.
The session wasn’t.

TRACE wraps an agent in three things: a content guard you can read, temporal monitors whose memory never grows, and a signed ledger that anyone can re-check without the model. Below, simulated agent sessions run through all three.

Watch a session

Runs entirely in your browser. Nothing is sent anywhere.

Agent session

3 wire formats · 1 policy
    Press Play or Step to run the session.
    Add your own step

    Policy monitors

    Certificate ledger

    hash chain · HMAC

      How it works

      1. Text first. Anything the agent is about to read goes through the content guard: a Tsetlin machine whose 200 rules are ANDs of words. Every unsafe rule that fires is logged with a small CNF receipt of why.
      2. Calls as events. An adapter turns each tool call, in any provider’s format, into events like export or sensitive_read. The policy speaks only in events, so it moves between agents unchanged.
      3. Memory that never grows. Each rule is an online temporal monitor (always, count, response, precedence). Its state depends on the policy, not on how long the agent runs.
      4. A log you can’t quietly edit. Every decision is hash-chained and signed, the session is sealed with a Merkle root, and a verifier replays it without the model. This page writes the same format as the Python package: its ledgers pass TRACE’s own verify_ledger.

      What this demo is and is not

      • The agent is simulated: the tool calls are scripted, the guard and monitors that judge them are real.
      • The content guard learned word cues from one dataset. It scores 84% on that dataset’s held-out split, and the paper reports that it falls apart on a different corpus.
      • Monitors catch only what a policy states. The HMAC key is shared and shown here (demo-session-key), so the ledger proves integrity to whoever holds the key, not authorship.