ميزانMizanكله بميزانه

كله بميزانه

Ask about education. Follow the evidence.

A conversation that keeps its sources close.
You could start here
العربية / English
Mizan can make mistakes. Check sources and context. The scenario engine is in development.Connecting to sources
The current buildميزان

Welcome to Mizan.

One chat interface. Evidence search works today; testing decisions is the next stage.

Available now · evidence assistant

Ask in Arabic or English, continue the conversation, and open sources behind the answer. The current MVP does not calculate outcomes for a particular school.

In development · institutional digital twin

School data connectors, causal and operational models, agent-based simulation, Monte Carlo uncertainty, and validation against observed outcomes are being built.

The model used now and the target model

Checking the active model connection…

Domain fine-tuning of Gemma 4 is under development. It is not answering this demo.

How your messages are handled

Checking the current deployment's privacy information…

Explore the source code and project notes on GitHub ↗.

كله بميزانه
The complete technical picture

How Mizan works

The chat is the only control surface. An agent routes the question to the right backend path, keeps the assumptions visible, and explains the result.

QuestionAgentic routerEvidence or scenario engineChecked answer
Working now

01 / Understand the request

The planner reads the latest question and short chat history, detects research, a school decision, or an unclear choice, and prepares a self-contained search query. Ambiguous operating decisions trigger a clarifying question.

Working now

02 / Retrieve evidence

Public documents are cleaned, deduplicated, split into passages, and indexed. BM25 and multilingual E5 find candidates; reciprocal rank fusion and a multilingual cross-encoder rerank them. The library currently holds 2,648 distinct works.

Working now

03 / Ground the response

The answer model receives selected excerpts and metadata. The server checks cited IDs, quoted text, and numerical tokens against retrieved passages. This checks provenance, not whether every interpretation is causal or locally transferable.

In development

04 / Build the institution twin

Permissioned connectors will map school or university records into a baseline of students, teachers, rooms, timetables, budgets, attendance, and learning outcomes. Data quality, privacy, and missingness must be checked before modelling.

In development

05 / Run a decision scenario

The agent will turn a decision into a defined intervention and constraints. Causal models estimate defensible effects; agent-based and operational models represent people, capacity, schedules, and costs; Monte Carlo runs express uncertainty.

In development

06 / Validate and communicate

Backtests and prospective pilots will compare predicted ranges with real outcomes and recalibrate the twin. The response will separate measured evidence, assumptions, simulated outcomes, subgroup impacts, resource needs, and uncertainty.

The model's role

Checking the active model…

Fine-tuning and hosting Gemma 4 for this domain are under development. The language model will route and explain; it will not invent simulation results.
What exists and what comes next

MVP & roadmap

The status of each layer is explicit so a sourced answer is never confused with a tested prediction.

Live in this MVP

Evidence conversation

Arabic and English chat, automatic routing, retrieval over 2,648 distinct works, source inspection, citation checks, and an honest answer when evidence is insufficient.

Checking the current model…

Under development

Institution data + decision engine

Structured school and university connectors, baseline digital twins, causal estimates, operational constraints, agent-based simulation, Monte Carlo uncertainty, and evaluation against real outcomes. No school-specific forecast is available in this MVP.

Under development

Domain-tuned Gemma 4

Fine-tuning and GPU hosting for Gemma 4 are under development. The tuned model is not trained, hosted, or answering here yet.

Validation gate

Before decision support

Run historical backtests and prospective pilots, document where the models fail, calibrate uncertainty, and check equity and privacy before presenting institution-specific predictions as decision support.
Behind the conversation

Source library

Research, policy, and data you can trace back to the original.

—distinct worksLoading the source index…

Loading sources…

Source records may include versions or volumes of a work. Findings apply to the original publication's context.
Follow the evidence