Frequently asked
What we get asked first. With the figures.
Ten questions come up in every conversation. The answers give their measurements, their protocol and what they do not cover.
Do I need a GPU?
Not to start, and here are the two figures to decide for yourself. On a standard processor we measure 170s per alert on average, 573s in the worst case. On a graphics card rented by the hour, the same chain returns its verdict in 2.3s, 74 times faster, and the worst case falls to 3.5s.
The card speeds things up, but the verdict does not replay identically on it. The hardware choice therefore also depends on your audit requirements. Sizing is worked out on yours.
Do my logs leave my infrastructure?
No. The solution runs on your infrastructure, without depending on a remote AI service. Intelligence comes in as a signed pack, loaded offline. Check it at your firewall: that is a verification, not a statement.
Which model, and can I change it?
The model is shipped, named and versioned. The engine reads weights in GGUF format, so you can replace it, subject to compatibility with your configuration. Every verdict carries the model and prompt fingerprint: a change shows in the history, and it is documented.
I do not have a SIEM yet.
Wazuh is a free option to centralise your events, and it feeds LUAXERR directly. If you already run Splunk, Microsoft Sentinel, QRadar, Elastic or Google SecOps, integration depends on the product and its version; it is confirmed when we study your environment.
The code is not published. Why trust you?
Part of it can be checked on your side: outbound traffic at your firewall, the identity and version of the model, the traceability attached to each verdict, the reproducibility of the deterministic processing, and test results with their protocol.
What you cannot do: read our guardrails line by line. So we separate what can be checked from what requires trusting us. What can be checked anyway →
Is it real-time?
No. Processing takes from a few seconds to a few minutes per alert, depending on the hardware, the model and the load. We favour traced, replayable analysis. Performance is measured on a set representative of your environment.
Does an AI triage better than a detection rule?
Detection and prioritisation are two different jobs. A rule decides whether an event fires; it does not say whether that firing deserves an analyst's attention this morning. That second job is ours.
A queue that is too large leads to disabling the chatty rules. Bringing the noise down helps keep the existing detection coverage.
Downstream of a SIEM, your AI will never see an attack the rules did not detect.
That is correct. From the SIEM's alerts alone, we are bounded by its coverage.
What we add is triage, justification and traceability: a rule will not tell you why an alert was set aside, by whom, and on which facts. A separate local detection evaluates its own rules when the SIEM is unavailable; it does not replace full coverage.
What is your false positive 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.
Logs contain fields filled in by the attacker. Injection, hallucinations?
A log field is untrusted data, so the model is never its own guardrail: five deterministic checks frame it, outside the model. Injection detector before the call, data isolated in the prompt, constrained output, coherence and grounding checks, closed action catalogue.
One of them verifies that the verdict does not contradict the reasoning above it: a text that names an attack cannot conclude “noise”. The results and their limits are documented in the repository, including the payloads that get through.
These mechanisms reduce the risk without removing it. What remains: no automatic closure, the decision belongs to the analyst. The guardrail in detail.
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.
