HOW IT WORKS
How Lokorium works
Lokorium is the final check before your data. This page explains what it is made of, what you have to trust, and what you do not.
The architecture
On the route
Lokorium sits on each protected route, in your own environment, and holds a request before it reaches your system.
The check
Deterministic: the same request and the same rules always produce the same decision. There is no machine learning anywhere in the decision path.
Sealed and attested
The check, its signing keys and the record run in a hardware protected environment, isolated from the systems and the people around it, including ours. It attests to what is running before it is trusted with keys, and you can check that attestation.
The evidence
Every decision is written as an encrypted, signed entry, hash chained to the entry before it, so any deletion, insertion or edit breaks the chain.
What a decision looks at
| A rule can refer to | For example |
|---|---|
| Who or what is acting | The complaints agent, a vendor’s tool, a person |
| The action | Read, change, export, pay, send |
| The scope | One customer record, or every record |
| The values | Amount, recipient, account |
| The stated purpose | Responding to complaint 4471 |
That is why a rule can say “one record, for an open complaint”, and not just “this agent may use the database”.
How rules work
House rules
Apply on every protected route, to everyone, and always win. Never export the whole customer base. Never change bank details without approval. Never pay out above the limit.
Agent rules
Sit inside the house rules: what each agent may do, on which routes, within which limits, and when a person must approve.
Only people change rules
Every change is approved by a named person, and the approval is itself recorded. The system can recommend a change; it never makes one.
Watch mode
A new route can start by recording everything and blocking only the house rules, so you see real traffic before you tighten the rules.
The trust model
You do not have to trust
- The AI model or its vendor
- The agent’s own logs
- The cloud provider
- Your own administrators
- Lokorium’s staff
You do trust
- The hardware’s protection and attestation
- The published rules you approved
- Standard cryptography you can verify independently
And one thing worth saying plainly: Lokorium does not build models, resell them, or belong to anyone who does. A check owned by the company whose AI it is checking is worth less, whatever is underneath it.
When something goes wrong
If a request breaks a rule
It is blocked or sent to a person, and the reason is recorded.
If someone tampers with the record
The chain breaks and verification fails.
What it protects against
| Threat | Without Lokorium | With Lokorium |
|---|---|---|
| Stolen login used by an AI agent | The login is valid, so the request goes through | The action is checked against the rules, whoever is logged in |
| Hidden instruction (prompt injection) | Your AI follows it | The resulting request is checked like any other and blocked if it breaks the rules |
| Over broad request | Freeze all accounts runs | Blocked: the scope breaks the rule |
| Insider changes a rule quietly | Nobody notices | Rule changes need recorded approval |
| Logs edited after the fact | Evidence lost | The chain breaks, and verification shows it |
| A vendor’s agent overreaches | It does what its access allows | It does only what its agent rules allow |
How a deployment goes
One route at a time, and watching before blocking. Your data stays where it is throughout.
01
One route first
The route into your most sensitive data. Nothing changes for the people and systems using it.
02
Watch before you block
It records every request on that route and blocks only your house rules. You see your real traffic, including the requests nobody knew were being made.
03
Tighten when you are ready
You turn on rules when you want them, with the record already showing what each one would have stopped.
04
Then the next route
The same four stages again. Each route stands on its own, so none of them waits on the rest.
What it asks of you: someone who can say what the rules are, and access to the route. Your agents do not change, your systems are not rebuilt, and the data does not move.
See it on one of your own routes.
Bring the workflow you cannot sign off. We will show you the decision and the record it writes.