Signet
Signet governs every action an AI agent attempts. It decides pass, block, or hold against a sealed mandate, it makes the call with logic and not a second model, and it seals the result to a permanent, tamper-evident record. It runs locally, in microseconds.
The gate ยท Runtime governance
Every action passes a non-bypassable gate before it executes. The gate answers with set logic and argument provenance, so there is nothing to persuade and nothing to drift. That is what makes the result reproducible and the overhead about 1x.
First, is this tool in the agent's mandate at all. Second, where did the arguments come from: are they owner-authorized, or derived from untrusted content. Trusted passes, untrusted is blocked, ambiguous is held for a human.
A hallucinated tool fails admission. A fabricated or injected argument fails provenance. An off-plan action fails scope. The model can propose anything. The runtime disposes.
The mandate
When an agent is born, Signet compiles its mandate: role, scope, tool limits, budgets, and the environments it may act in. That mandate is sealed into the agent's genesis record, so its authority is fixed before it runs and cannot be talked up by a clever instruction.
Mandates layer. A base policy, then a domain pack on top, merged strictest-wins. Change the policy once and it propagates, without redeploying the agent.
* The budget can be changed by the operator, not the agent.
The ledger
Every decision is sealed to an append-only, hash-chained ledger on the agent's own machine. Each entry links to the one before it, so deletion, insertion, or reordering is detectable. The record is yours, and it never expires.
| Deed | Verdict | Reason | prev_hash | seal |
|---|---|---|---|---|
| write_file | Pass | owner-authorized | 00000… | a91f72… |
| write_file | Block | secret_pattern in arg | a91f72… | 3d0c81… |
| send_email | Hold | untrusted-derived | 3d0c81… | 7f1a9c… |
| wire_transfer | Block | out of mandate | 7f1a9c… | b7375f… |
Chain intact. Running signet verify recomputes every seal and confirms nothing was altered.
Model independence
The model is never written into the agent's identity. It is paired at the transport by agent resolution, and governed against the sealed mandate. Change GPT for Claude and the agent is the same accountable entity, with the same authority and the same history.
Local
The gate runs on your machine. No action leaves the network to be approved. Works in air-gapped and restricted environments.
Deterministic
Set logic and provenance, not a classifier. The same inputs give the same verdict, every time, which is what an auditor and a regulator can rely on.
About 1x
A decision in microseconds, with no second model call in the path. Your agents run at full speed.
What it does, stated honestly
Drift
A hard boundary the agent cannot act outside of, plus optional drift detection that flags statistical movement away from the agent's established conduct. Containment first, early warning second.
Hallucination
Signet does not prevent a model from hallucinating. It stops a hallucination from becoming an action: the invented tool fails admission, the fabricated argument fails provenance, the off-plan step fails scope. Nothing wrong reaches execution unchecked.
We will walk you through Signet governing a real fleet, and what it takes to deploy inside your stack.
See Signet governing agents live, and what it would take to deploy in your stack. We will get back to you within one business day.
Or email mitch@trajexy.com directly.
Thanks. We will be in touch within one business day.