02 · Ours
The analysis
This is The i's reading. HAQQ has no approval over it.
Where a claim can be checked we checked it and say how. Where it cannot, it is printed as HAQQ's statement with HAQQ's name on it. Two facts set the register. HAQQ was founded in 2020 and reported its first round seven months ago. Club Cedro is a few weeks old. Everything we found is here, read on two tracks: how HAQQ measures against the category as it stands, and where it has come from and can go. Where we have a view it is offered as a suggestion, not a verdict.
Track one · Against the benchmark
They put an independent test on their own site, and printed the result
On 23 July HAQQ put an independent evaluation on its own site. Sixty one tasks contributed by practising lawyers, across 21 jurisdictions including the United States, the United Kingdom, India, Singapore, the Netherlands and Oman. Every task run twice. Binary criteria fixed by lawyers before testing, and a task passes only when every criterion is met. HAQQ passed 27.0 percent. On the hardest subset it passed 5.9 percent, against 1.1 percent for the average legal AI on the same set. Their own summary: on roughly three out of four hard tasks, at least one lawyer authored criterion was still missed.
Read that against the category. Vendors in legal AI publish scores in the nineties, on tests they design. HAQQ published its result on a test somebody else designed, and wrote the caveat themselves. The evaluator is not named, because that evaluator keeps its reports private by default, so a reader cannot rerun it. That is a real limit, and a smaller one than anything else in the category offers.
The forty nine is theirs, and they say so
The other figure on the homepage, 49 out of 50, comes from HAQQ's own rubric. Fifty points across eleven categories, criteria published, nineteen tools scored on the same sheet. The top five on that sheet are in the table below. Topping your own rubric proves less than publishing the rubric does. Putting the criteria out and scoring named competitors on them is something we have not found elsewhere in the category, and the two figures together, the 27 and the 49, tell a reader more than either alone: this is what we get on our own terms, and this is what we get on someone else's.
HAQQ's 50 point rubric, top five of nineteen, as published on haqq.ai| Tool | Average |
|---|
| HAQQ, Justinian | 46.8 |
| Claude Fable 5 | 42.7 |
| Claude Opus 4.7 | 40.8 |
| Mike OS | 39.1 |
| Harvey | 37.9 |
They named the right test, and they are the ones closest to passing it
The most useful sentence in the questionnaire is in Q06. The technical test, HAQQ says, is not whether the system can produce fluent legal text but whether an outsider can inspect how the task was grounded, which jurisdiction and sources were used, what controls were applied, and whether a lawyer reviewed the work before it became final. That is a governance argument in a category that mostly sells fluency, and fluency is the property that has put lawyers in front of judges.
The first draft of this piece argued that HAQQ had set that test and then published figures that could not pass it. That was wrong. MINZÂR, still in development, overlooked the research pages on haqq.ai where the figures and their methods are published. Corrected below. What survives is narrower: until a legal AI benchmark exists that a reader can rerun, no vendor fully passes the Q06 test, HAQQ included. HAQQ is the nearest.
The research desk
Since February HAQQ has published around twenty studies on its site. A country by country tracker on whether lawyers may use AI, 17 jurisdictions, one restricting. A grading of 3,000 answers from ten frontier models across 300 legal tasks. Head to head reviews of Harvey, Legora, CoCounsel and Lexis on the same rubric. On 11 September, three models run against 41 real legal questions, all failing on Saudi and Egyptian law despite correct looking citations.
The most useful is the hallucination audit: 1,598 verified court cases involving AI fabricated material as of 9 June, built on Damien Charlotin's public database at HEC Paris, which lists a case only where a court explicitly found or clearly implied reliance on hallucinated material. HAQQ did not build that count. It reads it, checks it and publishes it, which is a public service the category needed and did not have, and the strongest argument for HAQQ's own product.
The result is theirs. So is the decision to print it.
Track two · Where it came from, and where it can go
They stayed, and built it here
HAQQ was founded in Beirut in 2020 and is run from Beirut today. On our own record that is the exception. Anghami, founded in Beirut, moved its headquarters to Abu Dhabi in 2021. Integral AI, Bounty, Ramp and Medicus AI were built by Lebanese founders somewhere else. Each of those was a rational choice and the register says so. HAQQ made the other choice, through the years it describes in Q15 without naming them, and its answer to Q16 says why: keep the talent, export the product. That is major, and it is the thing Club Cedro, as part of Lubnan Online, exists to put on the record first. Everything else in this issue is read in that light.
Six years, read in order
Founded in Beirut in 2020, which means built through the years HAQQ describes in Q15 without naming them. In February 2026 Sowlutions Ventures led a three million dollar round, reported by Wamda and L'Orient Today, with the platform above 7,000 clients at the time. That is somebody looking at the same company we are looking at, from closer, and putting money behind it. It is the largest confirmed round on the Lebanon tier of our record, where the next reported figure is around 150,000 dollars, and it is a plus on any reading. In July they published the 27. By September they were running a research desk that scores their own competitors in public and a site that now says 15,000 or more legal teams. Both client figures stay on the record with their dates, one reported and one stated, because that is how a figure keeps its standing. Read as a line rather than a snapshot, the direction is a company that keeps publishing more about itself, not less.
The strongest thing here is a boundary, because a boundary can be watched
Abbas Kabalan on the engineering conversation: this works in 97 out of 100, so tell me what happens in the other three, because one of them is somebody's case or somebody's money. And a refusal to build the fully autonomous AI lawyer, stated as a line the company will lose demos over. A published boundary is the one claim in this questionnaire that anybody can audit, by watching what HAQQ ships over the next two years.