The bet everyone makes
Every wave of compliance technology makes the same bet. The model is almost good enough. A little more scale, a little better prompting, some fine-tuning on regulatory text, and the system will finally be able to read the law and apply it. The bet assumes that the thing standing between us and automated compliance is the quality of the model, and that closing that gap is the work.
It is a reasonable assumption and it is wrong, and you can see that it is wrong by asking a simple question. If a perfect model arrived tomorrow, one that never hallucinated and understood every regulation exactly, would the problem be solved? It would not, because there would still be nothing for it to apply. The regulation it understood would still live in PDFs and guidance notes and the heads of experts, and not in any form a system can execute, check, or audit. The model would be brilliant and the engine would still have no rules to run.
What is actually hard
The deterministic part of this, the engine that evaluates rules against facts, is not where the difficulty lives. That problem is old and well understood. Formal verification, expert systems, decision tables, the whole lineage goes back decades. Running a rule is not hard. We have known how to run rules for as long as we have had computers.
The hard part is that the rules do not exist. Not in the form that matters. Regulation exists as prose written for humans, scattered across primary law, secondary guidance, regulator letters, and case history, revised constantly, and full of the ambiguity that human language carries. To make any of it executable, someone has to read it, decide what it actually requires, encode that requirement as explicit conditions, attach it to its source so the decision can be traced, and have a person who knows the domain confirm the encoding is faithful. That is the work. It is slow, it is expert, and there is a great deal of it.
Why the model does not move it
Here is the part that the bet keeps missing. A better model does not make that work meaningfully faster, because the work is not bottlenecked on reading. The model can help read the regulation, and that helps. But the judgment about what the rule means, and the sign-off that the encoding is correct, are the slow steps, and those are human by necessity. A rule that a model drafted is not a rule you can deploy. It is a candidate that still has to be checked by someone who will be accountable for it, because the moment a wrong rule makes a decision, "the model wrote it" is not a defense.
So the model speeds up the cheap part and leaves the expensive part exactly where it was. You can generate ten thousand candidate rules in an afternoon now, and you have not moved the bottleneck an inch, because the bottleneck was never generation. It was validation, and validation runs at the speed of expert attention, which no model increases.
The implication
If the constraint is the rules and not the model, then the investment that matters is not a bigger model. It is the corpus: the slow accumulation of encoded, sourced, validated rules, one regulation at a time. That is an unglamorous asset. It does not demo. It is not a breakthrough you announce. It is closer to building a map, or compiling a dictionary, a long patient act of writing things down correctly that compounds into something no competitor can shortcut, precisely because there was no shortcut for you either.
The model is necessary. Without it the inputs stay expensive and the whole thing is uneconomic. But necessary is not the same as scarce, and the model is not the scarce thing. Everyone has the model. The rules are the scarce thing, and they are scarce because they are hard, and they are hard in a way that throwing a larger model at them does not touch.
The point
The reason no one has yet built the compliance engine that the industry keeps promising is not that the models were not good enough. The models passed good enough a while ago for the job they are actually suited to. The reason is that the regulation was never written down in a form a machine could reason over, and writing it down is slow, careful, expert work that looks nothing like the breakthroughs the field celebrates. The bottleneck was never the model. It was the rules, and it still is, and the only way through it is to do the unglamorous thing and write them.