RedOS / Sovereignty

Your own instance. Not a row in someone else's.

Each RedOS is installed as a closed environment for one business - cockpit, knowledge, agents and storage, provisioned for you, operated for you, isolated from every other client. Delivered as software: you never install or operate anything.

Sovereign by construction The first question a serious buyer asks - and the hardest one to retrofit
The difference

"Multi-tenant" is a word doing a lot of work.

Almost every AI platform you will be shown is one application serving every customer, where your company is a row in a shared table and a workspace ID in a query. That is not the same thing as having your own system, and the difference only becomes visible on the day it matters.

The usual model
A tenant in someone else's app
  • One application instance serving every customer at once.
  • Your data sits alongside your competitors', separated by a column in a query.
  • Isolation depends on every developer remembering to filter correctly, every time.
  • A single application bug is a cross-customer data incident.
  • "Delete my data" means trusting a flag was set somewhere.
  • You inherit every other customer's noisy neighbour problem.
RedOS
A closed environment for one business
  • Your own instance of the whole system - cockpit, knowledge, agents, storage.
  • Row-level security forced on at the database, in the same change that creates each table.
  • The application connects as a least-privilege account that cannot bypass it.
  • Proven continuously by an automated cross-tenant test harness, not by assurance.
  • GDPR-grade tenant erasure - the environment goes, not a flag.
  • Provisioned and operated for you. You never install or run anything.
Architecture

What "your own instance" actually means.

Visual to come Instance topology diagram Two panels side by side: the pooled model (one app, many customers, one database, isolation by query filter) against the RedOS model (one environment per business, its own data plane, isolation enforced beneath the application). This is the single most persuasive image on the page and it is worth commissioning properly.
Structural, not bolted on

Six properties that cannot be added later.

Each of these is a consequence of how the platform is built. None of them is a setting, a policy document, or a promise about process.

Isolation at the database

Every tenant-data table ships with row-level security forced on in the same change that creates it. The application connects as an account that cannot bypass it. Even a bug in application code cannot leak data across clients.

One permission authority

A single fail-closed service answers "can this principal do or see this?" for every principal - human and agent - across the whole platform. There is no second, weaker system to audit.

Private mail cannot leak

The three-scope mail wall is restrictive at the database layer. Private mail structurally cannot enter a shared view or the Knowledge Brain - not "is filtered out of", cannot enter.

No silent failure

A ratified hard rule: no silent fallbacks. If something is broken the system says so loudly rather than substituting degraded or invented data. Enforced in CI, so it cannot rot.

Humans gate the dangerous things

Deployments, go-lives, budget top-ups and high-risk actions require explicit human approval. Who is allowed to approve is itself a governed grant.

What we say is what is true

Internal specification documents carry freshness stamps checked against the code by CI. A claim that has drifted from the build fails the build.

Visual to come The gate, drawn once A single diagram showing a request from a person and a request from an agent arriving at the same fail-closed authority and being answered by the same rules. It carries headline 2 as well as this page, so it is worth building to be reused on the Team page.
Diligence

What your IT and legal teams will ask.

Written to be forwarded. If a question here is answered vaguely, that is a defect - tell us.

Where does our data live?
In your instance. Not in a shared application database partitioned by a customer ID. The environment is provisioned for your business and operated for it.
What stops another client seeing our data?
Row-level security forced on at the database, applied in the same migration that creates each tenant-data table, with the application connecting as a least-privilege account that cannot bypass it. It is not application logic, so an application bug cannot defeat it.
How do you know it works?
An automated cross-tenant test harness runs against the isolation guarantee. It is a continuously proven property, not an annual assertion.
Can an AI agent see more than the person it works for?
No. Agents are principals like people - same rank, same departments, same explicit grants, same fail-closed authority. Finance's agent cannot read HR's files any more than finance's people can. There is no separate "bot permissions" model.
Does our confidential email end up training something?
Private mail structurally cannot enter shared views or the Knowledge Brain - the mail wall is restrictive at the database layer. Human-to-human direct messages are never ingested into the company knowledge base.
Who holds our credentials?
Not the cockpit. Red Command is the policy authority - it decides what is turned on - but deliberately holds no passwords or tokens. OAuth is handled by a dedicated subsystem.
What happens when we leave?
GDPR-grade tenant erasure. The environment is removed, rather than a deletion flag being set on rows that stay where they are.
Are we locked into one AI vendor?
No. Model routing is tiered and configurable per tenant and per agent. A better model next quarter is a config change, not a rebuild.
Sovereignty

Send this page to your IT lead.

If it does not answer their questions, we would rather hear that than find out later.

Book a Red Brief