Compliance needs a model of its own.

Every serious field has already built its own model.

The pattern is now old enough to be reported rather than predicted. When a field with its own language, its own documents and its own idea of a correct answer builds a model for itself, that model does better inside the field than a general model does. It has happened in medicine, in law, in finance, in science and in software, and it has happened often enough that the burden of proof has moved.

The research is public and it is consistent. Models pre-trained on biomedical literature read biomedical writing more accurately than models pre-trained on the open web. Models pre-trained on legal corpora do better on legal tasks, and the gain is largest precisely where the task is most distinctively legal, which is the opposite of what you would see if the advantage were a rounding error. The same holds for scientific text, for financial language and for code.

None of this says a general model is bad. It says a general model is general. Breadth is bought somewhere, and it is bought in depth.

Five fields, the same move. In each one the specialist model was built and then measured on the field’s own tasks rather than on a general test, because a field cannot claim an advantage it has no way to measure.

What a specialist actually gains.

The gain is not vocabulary alone, though vocabulary is where it starts. A field’s model learns the shapes its documents come in, what a complete answer looks like, and which of the many true things you could say are the ones that matter. Above all it learns the field’s standard of evidence.

That last one transfers least, and it is the one that decides this argument. Every field has a rule about what makes an answer good enough. In science it is reproducibility. In law it is authority. In compliance it is something narrower and harder, and it is worth stating exactly.

Breadth is bought somewhere, and it is bought in depth.

Compliance is not another vocabulary. It is a different test.

Most domains can be marked against something public. A legal question can be marked against statutes and decided cases. A medical question can be marked against the literature. Anyone who disagrees with the mark can go and read the same source.

Compliance cannot be marked that way. The right answer to a compliance case depends on the law plus the institution’s own framework: its risk assessment, its risk appetite, its rules and its thresholds. The same customer, with the same documents, can be correctly approved at one institution and correctly declined at another. Both institutions are compliant. Neither is making it up.

There is a second property, and people outside the work tend to underestimate it. A compliance decision is not finished when it is made. It is finished when it survives examination, by an auditor or a supervisor, long after the person who made it has moved on. The decision has to carry its own justification with it, because whoever judges it was not in the room.

What compliance work demands.

A private standard and a delayed examination together generate a specific list of demands. These are the seven properties Brayns LLM is built around, and the same seven ComplianceBench measures.

Read that list the way a supervisor would. Each line is a question you can be asked about a decision you already made, and for each one there is an answer that ends the conversation and an answer that does not.

The gaps are structural, not incidental.

A general model does not fail these demands by accident. It fails them because of what it was optimised to be. It was trained to be helpful and agreeable across every subject at once, and both of those pull against what the list above asks for.

Take grounding first, because it has the most direct evidence behind it. Work profiling legal question answering found that general models assert authority which does not exist, in a confident register, often enough to call it a property rather than a glitch. The compliance equivalent is a memo citing a rule the institution does not have, or a register entry that was never made. That is worse than a plainly wrong answer, because a plainly wrong answer invites a check and a well-dressed one does not.

Consistency fails more quietly. The same case, reordered or reformatted, can produce a different decision. Research on prompt ordering showed that permuting the same material moves the result, and research on long contexts showed that the position of a fact changes whether the model uses it at all. In conversation that is variety. In compliance it means two customers with identical facts can be treated differently because one file happened to be arranged differently, which is precisely what an examination is looking for.

Nothing about the customer changed between these two readings. Only the arrangement of the file did. A model whose decision moves with the arrangement cannot be relied on to treat two customers alike, and being able to show that you treated them alike is most of what consistency means in regulated work.

Escalation and policy adherence fail together, for the same underlying reason. A general model is trained to be helpful, on feedback from people who prefer answers to refusals. Research on sycophancy has documented the result: a drift towards the view the user has already expressed. A model shaped that way will produce a decision where it should stop, and will fold when an analyst pushes back even where the written procedure says it should not. Calibration research is the encouraging half of this. Models do carry usable information about how likely they are to be right, so knowing when to stop is a trainable property. It is simply not what general training optimises for.

A model that stops is an operational cost. A model that is confidently wrong is an operational risk, and the difference grows the longer the decision has to stand.

Auditability looks solved and is not. A general model will readily produce an explanation next to its answer. The explanation is generated text, though, and nothing guarantees that it describes the process which produced the answer rather than a plausible account of it. An auditor is not asking for a plausible account. They are asking what the decision actually rested on, and whether the same thing would produce the same decision again.

Decision accuracy and screening precision sit downstream of all of it. A model with no representation of your risk appetite has no way to hit your thresholds, and a model that cannot hold to your procedure cannot reach the decision your procedure prescribes. They are not separate failures. They are the visible surface of the five above.

A model built for compliance.

The conclusion is not that general AI is disappointing. It is that compliance is a domain, in the same sense that medicine and law are domains, and that every field which took its own work seriously ended up building for itself. Compliance is late to this rather than exempt from it.

Brayns LLM is that model. It is built for compliance work and nothing else, around the seven properties above, because those are the properties the work is judged on rather than the properties that are convenient to demonstrate.

A claim like that carries one more obligation, and it is the reason ComplianceBench exists. 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. Some of the fields above did build that for themselves. Software shipped its model and the test it wanted to be judged on in the same paper. Law came to it later and collaboratively, where Guha and colleagues (2023) assembled a legal reasoning benchmark with the profession rather than around it. Compliance had neither. It now has the benchmark, published in full and open to challenge, which is what turns the rest of this argument into something you can check rather than something you have to take on trust.