Four checks no model ranking replaces
A “best AI for legal work” table answers an engineer’s question. A lawyer in Tripoli asks a narrower one: is the text on screen the text, can it be opened, and does its validity have evidence.
In two lines
- We are not publishing a crown for a model on someone else’s problem set.
- Every answer still has to survive four checks: the source opens, the identity matches, validity evidence is visible, and applicability stays a human decision.
Every few months a table appears. Ten models. Hundreds of tasks. A score out of thirty-five. Then a hard title: this is the best AI for legal work.
The table helps if you are picking an engine. It is almost empty if you are writing an appeal, reviewing a supply contract, or asking when a Libyan civil judgment can be challenged.
The distinction is simple, and a lot of marketing works to blur it: “a model answered” is not the same as “an answer you can put your name on.” The gap is not speed or polish. The gap is the source.
The wrong question
“What is the best AI for legal work?” has no jurisdiction in it. Best under which law? From which sources? And who carries the file if the citation does not say what the model claimed?
Cross-border tasks written in English, scored against American, European and Gulf statutes, tell you nothing about an article in the Libyan Code of Civil and Commercial Procedure, a Supreme Court ruling, or the Official Gazette issue that published the amendment.
We are not racing those tables on their field. Ours is narrower and harder: a Libyan text you can open, a validity state with evidence, and a matter file that does not unlock merely because someone belongs to the same organisation.
First: does the source open?
A citation that does not open the text is decoration. An elegant mention of “article such-and-such” with no path to the page leaves you where you started: trusting a sentence, not an original.
In Izuran a citation is a door. If a reference appears with the answer, you should reach the statute, the ruling or the gazette issue — not take the summary on faith. If it will not open, do not rely on it. Ask for the source.
Second: is the identity right?
Models summarise well. They also misidentify quietly: article number, year of the act, name of a regulation, page of a judgment. The sentence still looks sound. The pointer is to another text, or to a text that is not there.
The check is quick if the source is open: is this the law? this article? this page? If you hesitate, the answer is still a draft.
Third: what sits behind the validity badge?
In force. Amended. Repealed. Unconfirmed. The badge is useful. It is also the most dangerous line on the page if you read it as a final ruling.
We do not show the badge as independent proof of the applicable version. Sometimes the repealing or amending text is linked in the corpus. Sometimes it is named and not linked. Sometimes only the source badge exists. Those are different kinds of evidence, and the product says so.
Fourth: applicability is not the model’s job
Finding the article does not mean it governs this dispute. The competent court, the date of the facts, the special provision that qualifies the general one: that is a lawyer’s reading, not a search hit.
The answer is built to be reviewed, not filed. Export checks may catch a blank, a missing reference, or a validity badge that still needs an eye. They do not decide that the argument is right.
What we measure now — and what we will not claim
We test that a known document can be retrieved from the index. That exposes a missing record, not the quality of an opinion. We show the evidence level next to a validity badge. We keep the legal corpus apart from the files you upload, so a public original is never mixed with a client paper.
We do not publish a legal-accuracy percentage for answers. That needs a Libyan question set, reference answers labelled by practitioners, and blind scoring. We will run that when the set exists. Until then, a large number about a model “winning” on tasks that are not ours would be advertising, not measurement.
| What a model table shows | What a filing needs |
|---|---|
| Drafting quality on general tasks | A text you can open from the answer |
| Model accuracy in another jurisdiction | Identity of the act, article and page |
| Latency and unit cost | Validity evidence read with the badge |
| One winner for everything | Applicability decided on the facts |
How to read an answer here
- Treat the summary as working draft, not final advice.
- Open every reference that appears.
- Match identity: instrument, article, page, date.
- Read the validity badge and its evidence in the same breath.
- Keep applicability. If a source is missing, ask for it before you build on it.
The strongest model in the world, with no door to the original, is still a fluent writer on a draft. The lawyer needs the original.
If you want the screens after this argument, the user guide covers chat, research, matters and sources. This blog is not a second manual. It is why the manual is written this way.