Sovereign cybersecurity, within everyone's reachSovereign security for allSOC AI on your servers · free edition Request a demoDemo

Why LUAXERR

Triage never ends. And it is rarely verifiable.

A security team spends its days dismissing alerts that threatened nothing, then explaining why it dismissed them. LUAXERR does that work alongside them, on their own machines, and leaves a trail an auditor can read six months later.

The starting point

Three difficulties that keep coming back

None of them comes from a lack of rigour. They come from volume, and from the fact that a decision taken quickly leaves little behind.

More alerts than hours

A properly tuned SIEM produces more alerts than a team can read. The sorting happens anyway, and nobody knows exactly which ones were skipped.

The proof arrives too late

An auditor asks why a given alert was dismissed. The console export says who clicked and when. It does not say what was weighed.

What goes silent never surfaces

A source that fails produces no alert. It is the one incident that triggers nothing, and it passes for a quiet day.

Our choices

Five decisions, and what they cost

None of them is free. Writing them down means the first meeting is about the right subject.

Local on principle

The model reasons on your machine. No alert leaves, and your firewall counts it at zero. The price: you need hardware, and a verdict takes from a few seconds to a few minutes.

Where the AI reasons →
Deterministic first

Prioritisation, coverage and hardening are computed, not guessed: the same input returns the same result. The model steps in only where it adds something, and its verdict carries the fingerprint of the prompt that produced it. Replaying that verdict identically, however, depends on the machine and on the version: on a GPU, two identical runs can diverge.

The boundary, brick by brick →
Auditable, not merely logged

Every decision enters a chained journal. The Enterprise edition turns it into a dossier signed with Ed25519 and ML-DSA-65, verified offline, without us and without the host. The signature proves the integrity of what was signed, not that the input data was complete.

An auditor's four questions →
Honest about what it does not know

A silent control, an asset outside the inventory, missing data: LUAXERR names them instead of concluding. With no declared inventory, it never shows “0 gaps”, it shows “not compared”.

Source continuity →
The decision stays human

No alert closes without an analyst, and every irreversible act waits for validation. The deterministic guardrail can only raise an alert, never set one aside.

The anti-injection guardrail →

What is measured

The figures we publish, with their protocol

A percentage without a protocol is not a measurement. Each of these figures carries its protocol and its scope, and we also publish the ones that do not flatter us.

0bytes

Network egress, verifiable at your firewall

320cases

Control set the model was never trained on. Public measurement method

−12/ 12

False positives removed by a tuning replayed on the measured history, without setting aside a true positive

Hybrid

Signed audit dossier, Ed25519 and post-quantum (ML-DSA-65), verified offline without us

SHA-256

Prompt fingerprint in every verdict: it identifies the content, it does not replay the verdict

The test bench in detail, the control set and what is left to measure: see the results →

The scope

What LUAXERR is not

Said before signature, because a bad surprise at install time costs more than a lost sale.

  • It is not a SIEM. LUAXERR reads yours, it does not replace it. If you have none, Wazuh is free and feeds it directly.
  • It is not an EDR or a firewall. It sees only what your sources give it, and it says when they give nothing.
  • It is not a SOAR. It proposes actions and prepares rules. It does not apply them on its own.
  • It is not an analyst replacement. It prepares the work, the analyst decides, and the correction they make is folded back into the tuning.

The full list, with the known limits: the scope →

Auditability

The fundamental questions

“Which model, which version?”

Named, versioned, running on your machine. No third-party provider to question.

“Where does it reason?”

On your server, or in your cloud. Intelligence arrives as a signed pack, loaded offline. Check it at your firewall.

“Can you replay a decision?”

Every verdict carries the model and the prompt fingerprint that produced it. On a processor, prioritisation is deterministic: same input, same score. On a graphics card we measured four cases out of sixty flipping between three identical runs. The card speeds triage up, it does not replay it.

“What is your error rate?”

We measure it on 320 alerts the model was never trained on, 160 of noise and 160 of attacks. In the measurement of 10 November 2026, 144 noise alerts out of 160 were recognised and 137 attacks out of 160 were escalated by the model alone.

These results describe the performance observed under the conditions of the protocol. They are not a guarantee that no error occurs in real conditions: that is why no alert is closed automatically.

The measurement conditions and their limits are described in our measurement method.

DiagramThe auditability dossier, and who verifies it
The auditability dossier, and who verifies it A period dossier lists the models, prompt fingerprints, decisions and human control, plus its reservations. It is signed twice, with Ed25519 and with ML-DSA-65, a post-quantum signature. The auditor verifies it offline with the public key, with no return to the vendor. AUDITABILITY DOSSIER · PERIOD SEPT 01-30 · EXAMPLE modelmistral-7b · q4 · sha256 a91f…3c prompt fingerprintc41e…9a (v4) decisions1,284 · human control 96% accessfull log, 3 accounts reservations“traceable, not bit-for-bit reproducible” SIGNEDTWICE public key Your auditorverifies offline, alone LUAXERRno return to us AUDITABILITY DOSSIERPERIOD SEPT 01-30SIGNEDTWICEmodelmistral-7b · q4 · sha256 a91f…3cprompt fingerprintc41e…9a (v4)decisions1,284 · human control 96%accessfull log, 3 accountsreservations“traceable, not bit-for-bit reproducible”public keyYour auditorverifies offline, aloneLUAXERRno return to us
The figures in this diagram are an example. The dossier lists what it does not establish. A document that only states what it proves leaves the auditor to find the holes alone, and loses its credibility exactly when it matters.

What is different

The difference is measured by what you can verify

Installing on site is not enough to guarantee transparency. LUAXERR makes its decisions traceable, lets you verify the records, and documents the cases where the available data does not allow a conclusion.

The questionWhy it is hardWhat LUAXERR does
Where the AI reasonsA remote computation cannot be checked from your own premises.✓On your server, verifiable at your firewall
What stopped emittingAn absent alert does not draw attention to itself.✓Silent controls, unknown assets, never audited: all named
The error rateA percentage without a protocol is not a measurement.✓Published, with protocol and sample size
Proof for the auditorAn export is debatable; a signature is verifiable.✓Dossier signed with Ed25519 and ML-DSA-65, verifiable offline without us
OT networks and connected devicesA probe that emits can disturb a production line.✓Passive probe, IEC 62443 and ETSI EN 303 645, that says what it does not hear
The trialA trial that has to go through Legal does not happen.✓Free on your servers, no time limit

Contact

See LUAXERR judge an alert from your environment

A demo on your context, not an idealised scenario. And if the answer is “not now”, we will say that too.

This site uses cookies. Technical cookies are required for the site to work and stay on at all times. Functional and audience-measurement cookies wait for your consent.

Customise