The model

Brayns LLM.

The only compliance model built by compliance people. It is built for compliance work and nothing else, around the seven properties that work is judged on rather than the ones that are convenient to demonstrate.

General AI is built to be adequate at everything. Compliance is not an everything problem. It is decided against one institution's own framework, and examined years later by someone who was not in the room.

Brayns LLM is built for compliance work and nothing else. Everything on the platform runs on it: one behaviour to learn, one to test, and one to answer for in front of your regulator. General models guess at compliance, and a compliance function should not run on guesses.

The seven properties below are not a summary of the model's qualities. They are the list of questions a compliance decision has to survive, each one stated in a form a practitioner recognises and a researcher can build against.

The seven properties a decision is judged on.

Brayns LLM is built around these seven, and ComplianceBench measures the same seven. Each one is a question your regulator will ask of a decision, put in the form your own procedure answers it.

PROPERTY 01
Decision accuracy

The decision matches the one a senior analyst reaches on the file.

PROPERTY 02
Screening precision

What should be caught is caught and the noise stays out.

PROPERTY 03
Evidence grounding

Every conclusion points at something in the case file.

PROPERTY 04
Decision consistency

The same facts produce the same decision, every time.

PROPERTY 05
Policy adherence

Your written procedure wins over the model's own reading.

PROPERTY 06
Uncertainty and escalation

When the file is incomplete, it stops and says so.

PROPERTY 07
Auditability

Every decision's reasoning is recorded and can be followed.

Held to the standard the domain sets.

The same seven are what ComplianceBench measures. A model built for a field has to be measurable in that field, by people who do not work here, against answer keys they can read for themselves.

ComplianceBench is the public form of that standard. It publishes the institution, the cases, the answer keys and the rubrics, so the definition of correct this model is built around can be read and contested by people who do not work here.

That release contains the instrument and no scores, which is deliberate. A benchmark released together with its maker's winning score is marketing material. Released before any score exists, it is an instrument, and an instrument is what this field lacked.

Built by AI researchers and compliance professionals.

Brayns LLM is built by two kinds of people in one team: AI researchers from frontier AI labs, and compliance practitioners whose careers are measured in decades across banks, payment institutions and supervision. Neither advises the other. Both build the model.

Compliance professionals decide what correct means
Practitioners with decades in banks, payment institutions and supervision settle the answer keys: what the correct decision is on a given file, under a given procedure. That definition is theirs, not the model's.
AI researchers hold the model to it
Researchers from frontier AI labs settle how the model is measured against that definition, and how it is held to it.
One team, not one advising the other
Both disciplines sit inside the build. Neither is an advisory board to the other, and the decisions about correctness are made in the work, not around it.
Checked by people who do not work here
The same seven properties are what ComplianceBench measures, with the cases, answer keys and rubrics published, so the definition of correct can be read and contested from outside.

Enterprise-grade security.

Brayns operates under the standards your regulator and vendor due diligence expect, independently audited and kept current.

GDPR CompliantPersonal data handled under EU law, by design and by default.DETAILS →
SOC 2 Type IIIndependently audited controls, shown to operate over time.DETAILS →
ISOISO 27001:2022Certified information security management.DETAILS →
ISOISO 27701:2019Certified privacy information management.DETAILS →
ISOISO/IEC 42001:2023The international standard for AI management systems.DETAILS →
EU AI ActRisk management and transparency built to the Act's requirements.DETAILS →