Ground it or don't ship it

Pericls gives compliance and regulatory guidance to fintech companies. It has ingested global fintech regulation across every jurisdiction we cover and extracted the obligations, the penalties, the vendors and the licences into a corpus we build in advance.

The corpus is the part people skip past, and it is the part that matters. Ask a general assistant the same question and it will analyse what it believes is most relevant. Pericls analysed everything. That is the difference between a plausible answer and a complete one, and in this domain the gap between those two is where the whole product lives.

The stakes set the engineering. Get a compliance recommendation wrong and it costs your customer more than money. It can cost them a licence, the ability to operate, and in some jurisdictions somebody's liberty.

Now the technical crux. The questions Pericls answers are binary. Which vendor product satisfies this obligation. Which licence satisfies this obligation. Which obligation is already satisfied by default because you satisfied a different one in another jurisdiction. Yes or no, this one or that one.

By definition, neural nets do not return binary. They return a point on a spectrum which you then threshold into a yes or a no. Regression models return a curve. That is not a flaw, it is what they are, and for an enormous number of problems it is exactly right. For a question where being wrong means you believe you are compliant and you are not, it is the wrong instrument.

Everyone reaches for RAG as the fix for hallucination, and RAG is genuinely useful, but it does not solve this. Retrieval still ends with a generative model deciding what to say. Two failure paths survive it. Retrieve the wrong references and the answer is confidently wrong. Retrieve the right ones and the model can still hallucinate on top of them, which happens most where the information is simply absent and something has to fill the gap. You have narrowed the input. You have not eliminated the guess.

So the rule I work to is this. If a wrong answer is unrecoverable, the final decision cannot come from a model. The model retrieves, ranks and explains. Something deterministic makes the call.

I want to be clear that this is not an argument against language models, because they earn their place in the same system. Interpretation, comparison, semantic work across regulations written by different bodies in different decades in different registers: that is genuinely what they are best at, and nothing else I have tried comes close.

It is worth being precise about how good, because the claim gets overstated in both directions. On standard sentiment benchmarks with labelled training data, fine-tuned smaller models like BERT still tend to edge out large language models. Where LLMs win decisively is zero-shot and few-shot, without task-specific training, and on the harder cases that need actual reasoning, like implicit sentiment and sarcasm. Comparable or better without training, clearly better on the subtle cases, not better across the board.

Which is the argument in one line. They are the best interpretation layer available, and that is precisely why they should not be handed the verdict.

“If a wrong answer is unrecoverable, the final decision cannot come from a model.”