Why we built it this way

Core Legal Rules — a statement of conviction. Last updated 2026-06-18.

This is an opinion piece, clearly labeled as such. The methodology document states what the system does and how it can fail; this one explains why we believe an AI-built, AI-verified, openly documented approach to legal rules is not a compromise but the better path. You are free to disagree. We'd rather earn your trust by argument than ask for it by assertion.

The honest worry, named first

Anyone using this site is thinking some version of: how much confidence can I place in rules written by a machine? That is the right question, and we put it on the table rather than hoping you won't ask. One of the most useful habits we know — learned over a long career making arguments — is to walk straight at the weakness in your own position and see whether it is actually a strength wearing a disguise. So: the rules here are AI-generated. We think that, done in the open and checked against primary sources, that is a reason for more confidence, not less. Here is the case.

Experts are weaker than we wish, and we've known it for a while

The intuition that a seasoned human expert is the gold standard is comforting and, for a surprising range of tasks, wrong. The finding is old and unusually well-replicated. In 1954 the psychologist Paul Meehl reviewed study after study pitting expert clinical judgment against simple statistical models and found the models routinely equaled or beat the experts. The result has held up across decades and domains.

Yale Law professor Ian Ayres collected the modern version of this story in Super Crunchers (2007). The example that stays with people: a panel assembled to predict the outcomes of a U.S. Supreme Court term. A statistical model that ignored the legal merits entirely — looking only at features like which circuit the case came from and whether it was civil or criminal — correctly predicted about 75% of the affirm/reverse outcomes. The assembled legal experts, reading the actual briefs and law, managed about 59%. The machine that didn't read the law beat the professors who did.

Note what those winning systems had in common, because it is the whole point. Every one of them was validated against reality. The diagnosis was later confirmed; the Court actually ruled; the model's accuracy was a measured fact, not a promise.

Even the best journals get things wrong — and so does the law

It is worth sitting with how unreliable our most trusted sources of knowledge already are. Stuart Ritchie's Science Fictions shows that fraud, bias, and ordinary error reach the pages of Nature and Science; in medicine, systematic reviews find that roughly one in six cited "facts" does not actually support the claim made for it, around half of those seriously. We did not want to wave that away, so we asked the obvious question: is the law any better? It is, at least, more disciplined about citation-checking — student cite-checkers make law reviews a reputed exception to the high citation-error rates of other fields. But that is mostly a claim about process: tellingly, legal scholarship has rarely measured its own error rate the way science has, so we cannot even say with confidence how accurate it is — and the absence of a number is not the absence of errors. Careful citation form is not the same as truth, and the legal academy's own librarians have shown that its weakest link is correction: when an article turns out to be wrong, readers frequently cannot tell that it was fixed, which version is authoritative, or what changed.

We find that clarifying rather than discouraging. It means the goal was never the impossible one of zero errors; it is the achievable one of a system that surfaces, sources, and openly corrects them. And it means being just as hard on ourselves: asked to produce citations from memory, today's best legal AI models fail badly and confidently — which is exactly why this system never lets a model invent authority, but has it read the controlling text instead. The discipline we are proud of is not that we never err. It is that, unlike most of the systems people already rely on, our errors are meant to be visible and fixable in the open. (Sources for these findings are listed in our methodology page.)

Why that doesn't mean "trust any AI blindly"

A large language model is not, by itself, one of those validated predictors. Left to free-associate, it produces fluent, plausible text and can be confidently wrong. The magic in the studies above was never "the machine knows." It was "the machine's output was checked against ground truth and won."

So the way you earn the right to trust an AI-built rule is to do to it exactly what made those models trustworthy: tie each claim to a primary source, check it with more than one independent system, try hard to break it, and measure how often you fail. Grounded and verified, a language model becomes something like the validated predictor. Ungrounded, it is the opposite. The verification architecture described in our methodology document is not a hedge against our belief in AI — it is that belief, made operational.

Trust is shifting, and it has to be earned, not claimed

There is a larger current here. Faith in hierarchical, institutional authority has been declining across the developed world for half a century — the trend predates any one politician or slogan and shows up in survey after survey. It is tempting to conclude that AI simply inherits the authority that experts are losing. It does not. The same erosion that makes people doubt experts makes them doubt AI; ask a room of ordinary people how much they trust it and you'll hear the wariness directly.

That is precisely why we build the way we do. In a low-trust world, the only durable currency is verifiability. We don't ask you to transfer your trust from a human expert to a machine oracle. We ask you to trust neither on faith — and instead to look. The sources are linked. The method is published. The error rate is something we measure and intend to show. You don't have to believe a name or a credential, because we've removed the name and the credential from the equation and put the process in their place.

What we are, and what we are not

We are not claiming these rules are infallible. We are claiming something we can actually back: that this is a system which shows its work, grounds its claims in authority you can read yourself, checks itself in ways a lone human cannot, learns from every error you report, and tells you honestly what it has and has not yet verified. Against the real alternatives — commercial outlines that ship with errata, or a single tired human marking thousands of rules with an error rate no one measures — we think that is the more trustworthy thing, and the more honest one.

Take a look behind the curtain. That's the whole idea.

Counterpoint we take seriously: a reader could argue that grounding and cross-checking still can't catch errors that are common to the training data of every model, or that "majority rule" by machine consensus can quietly canonize an oversimplification. Both are real. They are why primary-source grounding, deliberately planted test errors, lower confidence ceilings for contestable doctrine, and your error reports all exist — and why this page is an argument for a method, not a guarantee of a result.