How this quiz decides what it decides, and everything it cannot tell you.
Scoring model 2, question bank 2, 46 questions.
It is a similarity quiz. It compares the forms you say you use against a set of features taken from published descriptions of Arabic varieties, and reports which descriptions your answers are closest to. It is not a measurement of your speech and not a test you can pass.
Arabic varieties overlap enormously. Most of the features here are shared by several varieties, which is why almost every answer in the question bank counts for more than one and why the result names a group before it names a variety.
Individuals differ from every profile. A description of Lebanese Arabic is a generalisation over millions of people, and no single person matches it on every point.
Migration, media, marriage, school and moving house all produce speech that no single variety describes. A mixed result is not a failed measurement, it is usually a more accurate one.
The result does not identify your nationality, your ethnicity, where you were born or where you live, and nothing in this quiz tries to work any of those out. The map describes where a dialect is spoken, not where you are.
Two levels: a family, and where the evidence allows it, a variety inside that family. A variety exists here only because the question bank carries features that can single it out. Size, prominence and completeness are not reasons.
6 families over 17 varieties.
Each answer carries evidence at whichever level it actually supports. Saying shu identifies Levantine and says nothing about which Levantine, so it scores the family and leaves the varieties alone. A variety’s final score is its own evidence blended with its family’s, so a variety can never be reported strongly unless its family is.
Each score is divided by the most that group could have scored on the questions you were actually asked, rather than by a single global maximum. That is what stops a gap in the question bank from counting against anybody: if a question offers no strong option for Gulf Arabic, answering it adds almost nothing to the Gulf denominator either.
So a percentage means “how much of the available evidence for this group did your answers give”. It is not a share of your speech, not a probability, and not a confidence interval.
Every result carries the version of the scoring model and the version of the question bank that produced it, and a shared link carries both. An old link keeps the numbers it was made with rather than being quietly reinterpreted when the model changes.
Everybody gets the same eight opening questions, in the same order. They separate the families and nothing finer, and running an adaptive selector over an empty answer set would only rediscover them.
After that, the quiz scores what you have said so far, works out which groups are still plausible, and asks whichever remaining question best tells the leading candidates apart. Questions about groups already ruled out are never asked, which is why a clearly Egyptian answerer is not asked to distinguish Moroccan from Tunisian.
It stops when one family is clearly ahead and one variety inside it is clearly ahead, or when it reaches its question limit. If two varieties are still close at that point it offers a few more questions rather than picking one.
There is no language model anywhere in this. The same answers always produce the same questions and the same result, which is what makes it possible to explain and to test.
The radar has one axis per kind of evidence, not one per dialect. Each point is how much of the available evidence of that kind your answers gave for your closest match. A kind you were never asked about is left off rather than drawn as a zero.
The composition ring normalises the variety scores against each other so they add to 100. They are quiz match components. They are not proportions of anybody’s speech, and very small ones are grouped rather than drawn as slivers.
The speaker figures are estimates from published sources, with their years shown. They are assembled from national populations, they overlap, and they do not sum to the Arabic-speaking world, so they are shown as a ranked comparison with visible ranges rather than as a distribution. Nothing here is a statistical distribution and nothing here should be read as one.
The map uses real public-domain country outlines, because approximate shapes drawn by hand would be a fake map. Countries are the finest geography available at this size, and they are not language boundaries: Arabic varieties run into one another across every border. The shading describes the dialect and is not derived from anything about you.
Every one of these was researched and then left out, because a category the evidence cannot support is worse than an honest gap.
Every answer in the question bank carries a reference in the source code. They are not shown beside the questions, because a citation under each option would ruin the quiz, so they are collected here.
Wiktionary is used heavily because it tags Arabic varieties with ISO language codes rather than with prose labels, and a code is a claim somebody can re-check. It is treated as supporting evidence: no fine-grained distinction in the bank rests on it alone.
A free account is required to start the quiz. That is a decision about who the feature is for; it is not a way of collecting anything. Premium is never required and never will be.
No attempt is written to your account and no answer reaches a server. The scoring runs entirely in your browser. Nothing about your quiz is stored against your user record, which also means there is no history to review and nothing to delete.
A result is private until you share it. Copying the link or saving the image is the only thing that makes it public, and the link carries the group scores and the variety it named, never your answers.
A longer version of this page, written for a linguist or a developer reviewing the assumptions, is in the repository at docs/dialect-quiz-methodology.md.