Why a Qatar edition
A tutor that talks to a child all day should think in the child's language and from the child's world. Fanar was trained on Arabic first, with the Gulf's dialects and references in its training data, and it was built in Doha. For a school in Qatar, that is the right model for the part that asks the questions and the parts that judge whether the questions are good ones.
It also answers the question every ministry asks about AI in schools: where does a child's conversation go, and whose model reads it. In this edition the answer is Fanar, on servers the ministry or the school controls.
What is inside
The tutor proposes each line before it is shown. Separate checks read that line, each with its own instructions, and Sheaf's coherence engine measures whether they agree. In the Qatar edition every one of these runs on Fanar.
The tutor
Asks a small question, waits, and asks the next one. In Arabic or English, whichever the student uses.
The rule check
Reads the tutor's next line against the rules the ministry and the school have set.
The learning check
Asks whether the line moves the student's own thinking forward, or hands over the answer.
The wellbeing check
Reads the student's messages for signs they need a person, and tells the teacher at once.
The gate
If the checks disagree with the tutor, the line waits for the teacher instead of going out.
The record
Every line, every hold and every release, kept where the school keeps it.
The model is Fanar-2-27B-Instruct, released by QCRI under the Apache 2.0 licence. Because the weights are open, the school or the ministry can run it on its own machines.
Runs inside Qatar
Every part of this edition, the model, the checks, the record of conversations and the teacher's screen, runs on servers inside Qatar: in a ministry data centre, at a national provider, or in the school itself. No student message leaves the country, and no AI lab, in Qatar or elsewhere, sees it. A school's own server keeps running when the connection to the ministry is down, and nothing a student does is lost.
The ministry sets the curriculum, the textbooks the tutor works from, the languages, and the values it teaches by. Sheaf Labs provides the system and the record of what it was set to do. A school adjusts within the limits the ministry sets.
What it is like in a classroom
The teacher starts a class and shares a code. Each student joins in Arabic or English. The tutor asks small questions and waits. When a student is stuck, the teacher's screen says so at once. Before a test, the teacher switches the tutor to test mode with one click. Parents get a weekly view of what their child worked on, in the family's language.
Data handling
A student's conversations and the small amount the tutor remembers about them are stored encrypted, in the school's own records. Nothing is used to train any model. In an in-country deployment, no message is processed outside Qatar. See the privacy policy.
The mathematics behind it
The checks are the same in every edition. The panel that reads each line, the rule check, the learning check and the wellbeing check, is checked with sheaf cohomology. It asks whether the three readings fit into one whole, and a line goes out only when they do. The rules of conduct go through formal logic, which checks that a new rule can hold alongside the rules already in force.
A tutor that thinks in Arabic, in a school that keeps its own records.
Talk to us about the Qatar editionAll editionsSheaf School runs on the same engine as Sheaf One.