Deployment
On your servers, or as sovereign SaaS.
Two deployments, one core, one chain of proof. The code is the same, the audit file is signed with the same key, and the error rate comes from the same bench. What changes is who holds the machine. And in both cases, one instance per customer: never a shared console.
One instance per customer, no shared database, no shared console
Datacentres and jurisdiction for the managed option, French operator named in the contract
Your alerts train no shared model. The fine-tuning corpus is handwritten and its hash is published
Signed audit file, Ed25519 and post-quantum (ML-DSA-65), verifiable offline without us and without the host
The same code in both cases: same verdicts, same signed dossier
The two deployments
Who holds the machine, and what that changes
Our solution deploys either on site, with full control of network flows through your own firewall, or as a managed option, depending on each organisation's needs and infrastructure.
On-premise, autonomous
- The model reasons on your machines. Outbound traffic is verified at your firewall, without having to take our word for it.
- Works on a network cut off from the internet. Intelligence arrives as a signed pack, loaded offline.
- What it requires: a machine that carries the model, and someone on your side to operate it. We train, we do not pilot.
- Updates: your call. No version installs itself without you.
Sovereign SaaS, managed
- One instance per customer. No shared database, no shared console: no data travels from one customer to another, because no path exists for it to travel along.
- Hosted in France by a French operator, subject to European law alone. The host's name and the datacentre are written in the contract and in the audit file, not on a brochure.
- We operate, you keep the verdicts. No automatic closure, no irreversible act without an analyst. Here as on-premise.
- The way out is written before the way in: one export, three hashes, and the same stack restarts at your place.
The right questionsEight questions worth asking of any security hosting
Sovereignty is not just the address of the servers. Here are the eight questions we suggest you ask, of us as much as of any other vendor, what each one changes for you, and the answer we write into the contract.
| The question | Why it matters | Our answer, managed |
|---|---|---|
| Whom you share the database with | A shared database creates a path between two customers, even when operations are impeccable. | ✓One instance per customer, no shared database and no shared console |
| Where the model reasons | Where the computing happens decides who can reach it, and which law applies to it. | ✓In your instance, in France, on a local model |
| What becomes of your alerts | A service improvement clause is enough to pull your logs into a shared model. | ✓They train no shared model. The fine-tuning corpus is handwritten, and its hash is published |
| The error rate | Without a protocol and a sample size, a percentage is not a measurement. | ✓Published, with its protocol and its sample size |
| Proof for the auditor | An export can be argued with, a signature can be verified. | ✓File signed with Ed25519 and ML-DSA-65, verifiable without us and without the host |
| The day you leave | Reversibility is written on the way in: on the way out, it is too late. | ✓The same stack restarts at your place, images and volumes included |
| What you can check yourself | What cannot be measured has to be taken on trust. | ✓No outbound call from your instance, replayable verdicts, prompt fingerprint in every decision |
| Applicable law | An extraterritorial law reaches a vendor through its ownership, not through the address of its racks. | ✓French operator, subject to European law alone |
ChoosingOn-premise or managed: what actually changes
Both columns come out of the same repository: this table sets no two levels of quality against each other, it describes two divisions of labour and two legal perimeters.
| Criterion | On-premise | Managed |
|---|---|---|
| Sovereignty | ||
| Where the logs liveAnd who can reach them physically. | At your place, on your hardware | In your dedicated instance, in France |
| Network cut off from the internetAir-gap, when accreditation demands it. | Possible, end to end | No: the instance is reachable over the network |
| Who sees the operationsTechnical monitoring, backups, system logs. | You alone | Our operations team, within a perimeter written in the contract |
| Operations | ||
| Who installs and updates | You, when you decide | Us, in an agreed window |
| Who is on call | Your team | Ours, at the level set in the contract |
| Hardware to plan forThe model needs memory, not a farm. | A server that carries the model | None |
| Proof | ||
| Signed auditability file | Yes | Yes, identical, same key |
| Verifiable without the vendor | Yes | Yes, and without the host |
| Published error rate | Yes | Yes, the same bench |
| Lead time and cost | ||
| Getting started | Depends on your hardware and your sources | Shorter: there is no server to order |
| Cost model | Licence, the hardware is yours | Subscription, the instance is included |
| The amountsThey depend on alert volume and on your sources. | Set in the quote | Set in the quote |
Who forFour situations, and the deployment that fits
The right deployment depends on three things: your accreditation, your hardware and your on-call rota. The first two situations call for on-premise, the next two for the managed option.
You are accredited, or working towards accreditation
- A framework that requires an isolated network accommodates no hosting at all, however French it may be.
- Intelligence arrives as a signed pack, loaded by hand. The outbound flow does not exist, and that is verified at your place.
You have a team and servers
- You already run a SIEM and an on-call rota: the managed option would charge you for work your own team already covers.
- The hardware is yours, updating is your call, and none of your logs reach us.
You have neither a server nor an on-call rota
- An SMB, a local authority, a practice: the constraint is not willingness, it is the server to buy and the night to cover.
- You keep what matters: the verdict stays advisory, and the audit file verifies without us.
You run security for other companies
- Each customer in their own instance, and the auditability file is delivered per customer, signed.
- What does not exist: a single console for your whole estate. You operate as many as you have customers: that is the counterpart of having no shared database, and we set it out at the scoping stage.
Hosting
One instance. One tenant.
No shared database, no shared console, no neighbour down the hall. What does not exist cannot leak.
The invariants
What does not change from one deployment to the other
- The deterministic guardrail can only escalate an alert, never set one aside.
- No alert closes by itself. An analyst always decides.
- The auditability file is signed with Ed25519 and ML-DSA-65, and verifies offline.
- The core is the same code in both deployments, and it does not leave your instance.
- The model is named and versioned, and every verdict carries the hash of the prompt that produced it.
- The error rate is published, with its protocol and its sample size.
What the managed option is not
The managed option has four limits, and we would rather write them down before you sign. Here they are, so you can hold them against your own constraints before going further.
- It is not an air-gap. The instance is reachable over the network. If your accreditation requires an isolated network, on-premise is what you need, and we will say so at the first conversation.
- It is not a single console for all your customers. A service provider operates as many instances as customers. That is the price of having no shared database.
- It is not “no third party”. A host enters your subcontracting chain. It is named in the contract and in the audit file, so that the decision is yours to make with the facts in hand.
- It is not self-service. A dedicated instance is ordered and commissioned: your sources are scoped, then the instance is accepted, before the first alert is judged.
The rest of what LUAXERR does not do is written in the usual place: the scope, stated before you sign →
Reversibility
Switching, in both directions
Reversibility is part of the offer, not of the options: the way out is written into the contract you sign on the way in. It takes four steps, because both modes, managed and on-premise, run the same stack.
Export
Verdicts, signed files, your analysts' feedback and the configuration leave as one archive. Nothing is converted: it is the format the tool already reads.
On requestRecord the hashes
You record the sha256 of the three things that matter: the weights being served, the prompt, and the audit file. Identical on both sides, or it is not the same solution.
HumanRestart at your place
The same stack restarts on your server: same images, same volumes, same model. The first verdict rendered at your place can be compared with the last one rendered at ours.
AutomaticClose the instance
The instance is destroyed, the backups purged, and you receive the deletion certificate. It is signed, like everything else.
Automatic · signedFrequently asked
What we are asked about hosting
Who exactly is the host?
A French operator, in datacentres located in France. Its name, the datacentre and the subcontracting chain are written in the contract and repeated in the audit file, because that is where they are of use to you. A brochure is not a contractual commitment: we would rather put the name where it protects you.
Who can read my logs in the managed option?
Your analysts, and our operations team within the perimeter written in the contract: technical monitoring, backups, system logs. No silent access, no reading “to improve the service”. And no customer's data is visible from another customer's instance, since no path exists between the two.
Do my alerts train your model?
No. The model being served is fine-tuned on a handwritten corpus whose size and sha256 hash are published with the bench results. If you order fine-tuning on your own analysts' feedback, it produces your model, on your instance, and it feeds no shared model.
What if you disappear?
What matters survives without us: the export archive is in the format the tool already reads, signed files verify with a public key you hold, and the stack restarts on your server without us. What our disappearance would take away is the updates, and the Enterprise edition writes its reversibility into the contract. It is the question we ask our own suppliers: it would be odd not to answer it.
Do extraterritorial laws apply?
A law such as the Cloud Act reaches a provider through its parent company, not through the address of its racks. A datacentre in Europe operated by a company under United States law remains reachable; that is why the criterion used here is the operator's ownership and applicable law, not geography alone. On-premise does not raise the question at all.
Can we start managed and move on-premise later?
Yes, and that is the path we recommend when the server is not yet budgeted. Switching is an export followed by a restart, not a migration: same images, same volumes, same model. The hashes recorded on both sides prove it. See the four steps →
Watch LUAXERR judge an alert, then choose where it runs
A demonstration on an alert from your own environment, and we will tell you which of the two deployments fits. Including when the answer is “neither, for now”.