There is a seductive shortcut available to anyone building a reasoning system today. You need training examples — cases where an expert examined a situation and explained their judgment. Expert time is scarce and expensive. But you happen to have access to a very capable language model. So you let the model generate the reasoning, label the cases itself, and feed its own output back in as training data. It is fast, it is cheap, and it is one of the most corrosive things you can do to a system whose entire value is the quality of its judgment.

We’ve drawn a hard line against it, and we enforce that line not as a policy memo but as a structural constraint in the system itself.

The distillation trap #

The shortcut is a form of distillation: a model teaching itself by grading its own work. In some domains this is harmless or even useful. In forensic judgment it is poison, for a specific reason. The thing we are trying to capture — the ability to look at a financial contradiction and judge whether it reflects a genuine problem or has an innocent explanation — is precisely the thing a general-purpose model does not reliably have. That judgment comes from years of forensic experience: from having seen how real misstatements are constructed, how they’re disguised in the language of routine accounting, how an aggressive arrangement differs from a customary one.

If you let a general model manufacture that reasoning, you are not capturing expert judgment. You are capturing the model’s imitation of expert judgment — fluent, confident, and wrong in exactly the ways a non-expert is wrong. And because it is fluent and confident, you won’t notice. The system will fill up with reasoning that reads like expertise and isn’t, and every model you train on it inherits the counterfeit. You have built an echo chamber and called it a knowledge base.

We saw early how easily this happens. A corpus of reasoning that looked like a valuable training asset turned out, on inspection, to have been generated by a general model rather than by forensic experts. The right response was not to clean it up or use it cautiously. It was to freeze it entirely and refuse to train on it until real expert annotations existed. Better to have no training data than to have counterfeit training data, because counterfeit data doesn’t just fail to help — it actively teaches the system to reason like a non-expert while feeling like it’s learning.

Why we distinguish real and fluent failure #

A subtle point underlies all of this. The danger of a fluent imitation is not that it is obviously bad. It is that it is plausibly bad. A general model asked to judge whether a structure is a genuine controlled-entity arrangement or merely boilerplate accounting language will often produce a confident, well-written answer — and will systematically over-call, treating routine language as if it were a real red flag, or mistaking a regulatory risk discussion for a company’s actual structure. The reasoning is articulate. The conclusion is wrong in a way only a domain expert would catch. This is exactly the gap that experience fills, and exactly the gap that self-generated training data cannot.

The implication is that who produces a piece of reasoning is not a metadata footnote. It is the single most important fact about that piece of reasoning. A judgment authored by a forensic expert and a judgment authored by a general model may be indistinguishable on the page and worlds apart in value.

Enforcing it in the schema, not the style guide #

Because the distinction is so important and so invisible, we don’t trust ourselves to remember it. We enforce it at the data layer. The reasoning attached to a case carries the identity and the role of its author, and the system structurally rejects any reasoning whose author is not a verified forensic expert. It is not a guideline that a careful engineer is supposed to follow. It is a constraint that makes the unwanted state unrepresentable: a non-expert simply cannot have their reasoning accepted as expert training data, because the database will refuse the write.

This has a cost. It means we cannot bootstrap our way to a reasoning corpus with a clever prompt. It means the system’s growth is gated on something genuinely scarce — the time and attention of people who actually possess forensic judgment. We consider that the correct trade. The bottleneck is real expertise, and we would rather respect that bottleneck than route around it with a counterfeit. A reasoning engine is only as good as the judgment it learned from. We refuse to let it learn from anything less than the real thing.