Security

Sovereignty is a claim about a line.Here is the line.

Your model runs on the aKillion hypervisor inside your own environment. Nothing goes out to a shared service and nothing you put in improves a product somebody else is buying. That is a property of how it is built rather than a promise about how it is operated, and the difference is worth drawing.

Book the review Permissions and automations

OutsideNot connected, not called, not trusted

Public model servicesNo call is made to one, so there is nothing to secure at that edge

Shared training poolsThere is no pool. Your records improve nothing that anybody else buys

Other customersNo shared model, so no mechanism by which what yours learns reaches theirs

Anybody outside your organisationNo standing access after setup, because there is no standing engagement

Nothing crosses inbound

Your environmentWhere the records already are, and where the model runs

Your recordsCustomers, work, money and people. The most sensitive data in the company

The hypervisorThe layer that governs which data any model may see, and where it may see it

Your modelPrivate to you, running where the data already is. Inference does not travel

The seven agentsThey inherit the role of whoever they act for, checked below the model

One crossing, outbound, on your commandYour own export. It is the only path drawn on this diagram that goes anywhere, and you are the one who opens it.

Illustrative diagram of the deployment boundary. Compliance posture and any certifications should be confirmed against current attestation documents before you rely on them.

What a reviewer actually asks

Four questions, and the mechanism behind each answer.

Security reviews do not turn on adjectives. They turn on whether an answer describes something the architecture makes true or something the vendor has undertaken to keep doing. Each answer below is followed by which of the two it is.

Does our data leave to be processed?No, because inference happens where the data already is.

This is the load bearing answer. Everything else on this page is a consequence of it. There is no outbound call carrying a prompt, so there is no outbound call to secure, monitor or trust.

Could our data train something we do not control?There is no shared pool for it to enter.

Not a policy commitment that could be revised in a future version of the terms. A property of an architecture with no route from your environment into a shared model.

What stops an agent seeing what a person could not?The permission check sits below the model, in the platform.

An agent inherits the role of whoever it acts for. Refusal is structural rather than behavioural, which means it does not depend on the model choosing correctly under pressure.

What happens if we want to leave?You export everything, whenever you want, without asking.

A platform that makes departure difficult is telling you something about its confidence in the rest of the product. This one treats leaving as a supported operation with a button.

The controls

Six of them, and where each one is enforced.

The right hand column is the point of this section. A control enforced in the model is a behaviour, and behaviours can be talked out of. A control enforced in the platform or at the boundary is a property, and properties cannot.

Role based access, per person and per teamIn the platformWhat each role may see, edit and approve, applied identically to people and to any agent acting on their behalf. One model of permission rather than two.

Approval thresholds you defineIn the platformMoney moving, records being removed, anything customer facing above a value you set. The agent prepares and stops. A named person releases and is recorded as having done so.

Dry run before any rule goes liveIn the platformNew automations run against real records from the previous period and the outcome is shown before anything can fire. You approve a demonstrated result, not a description.

Joiners, movers and leaversIn the roleAccess granted on the role at the start date and removed on the last day, written to the trail rather than depending on somebody remembering to raise a ticket.

Inference confined to your environmentAt the boundaryThere is no outbound call carrying your records to a model elsewhere, which is why the first control on this list cannot be bypassed by asking the model nicely.

Full export, at any timeAt the boundaryYour records are yours. There is no version of this arrangement where leaving requires a negotiation about getting your own data back.

Nothing on this list is enforced in the model. That is deliberate. Asking a model to decline is a request. Putting the check underneath it is a constraint, and only one of those two survives a determined prompt.

Every automated action is written to an audit trail with what ran, what it touched, what it produced and under whose authority, written to be read afterwards by a person rather than parsed by a tool.

The one path out

Leaving is a supported operation, not a negotiation.

This is the least comfortable section to write and the most useful one to read, because how a platform behaves at the end tells you more than anything it says at the beginning.

What you take with you

Every record, in full, in an open format
The document library, versioned as it stands
The audit trail for the whole period you were here
Your automation rules, written out so they can be rebuilt
Your scorecards, templates and terms

Whenever you want, without asking. Not a support ticket, not a call with an account manager, and not a clause somebody has to go and find. A platform that makes departure difficult is telling you something about its confidence in the rest of the product.

Where this connects

The boundary shows up in what agents may do.

ControlPermissions, automations and the audit trail in practice

AI layerWhat the agents can do and what none of them may do

Orion and IONWhere the boundary sits when the two halves are connected

For whoever signs off on the risk

Bring your security questionnaire.

The real one, with the questions your last vendor answered badly. We will go through it on the call and tell you plainly which answers are architecture and which are policy.

Book the security review See the controls running

KastorBook a demo

See your own week run on one record.

Bring one department and a real week of work. In about thirty minutes we put it through Kastor: an enquiry arriving, moving through your stages, becoming a signed agreement, becoming an invoice, and landing in margin. The same record the whole way, with agents doing the parts nobody should be doing by hand.

Your enquiry to cash loop on one record, with the handovers taken out
Agents that write the invoice and chase it, not agents that answer questions
What your receivables number is today, and what it looks like after
Your model running inside your own environment, with the data staying there

No trial signup and no card details One follow up, then we stop Your data stays in your environment We bring the questions, you bring one real week

Request your demo

Five fields and a consent box. We reply within one business day.

First name

Last name

Company

Work email

Phone

Industry (optional)

Where you would start (optional)

How you would run it (optional)

We use this to prepare the call and to reply. We do not sell it on or add you to a list you did not ask for.

What happens next

Three steps, and no sales process hiding behind them.

A demo request usually buys you a discovery call, then a qualification call, then a demo. This is one call, and the useful part of it happens in the first ten minutes.

A reply within one business day

Two or three times to choose from, and one question back so the call starts in the right place rather than starting at the beginning.

Thirty minutes on your own numbers

Not a slide deck. One department, one real week of work, and the loop running against it while you watch. Bring whoever argues with the numbers.

You keep the map either way

A written view of every point where your current setup asks a person to carry something between two systems. That is yours whether you go further or not.

Before the call

Three pages worth reading while you wait for the reply.

Lead to cashThe single argument the whole demo is built on

SecurityWhere your model runs and where your data stays

ControlWhat the owner sees when everything runs on one record

Not ready to book anything?

Fair enough. Read the lead to cash page instead. It makes the whole case in about six minutes and asks nothing of you. When it stops being theoretical, the form is still here.

Read that first