Education · SolEdu · A success story

Why we built SolEdu — and how it’s going

In 2025 Türkiye rebuilt its national curriculum; in March 2026 the ministry’s 123-page question-writing guide redefined what a good exam question is. We saw a production problem hiding inside an education problem — and engineered a system for it. This is the story so far.

Why we took this on

The Türkiye Yüzyılı Maarif Model didn’t just reorder topics. It changed the contract of assessment: questions must measure whether a student can use knowledge — interpret data, decide, produce — not merely recall it. The ministry said it plainly:

That single sentence ages out every existing question bank in the country. And the clock is real: the first cohort raised under the model sits its central exams in 2028. Whatever a school’s mock exams measure, its classrooms will practice — measurement drives teaching, not the other way round. Someone had to produce tens of thousands of guide-compliant questions, fast, without sacrificing quality. That is not an authoring problem; it is volume, consistency, and quality control — a production problem. Exactly our kind of problem.

Whatever the question measures, the classroom practices

There is a mechanism that makes assessment the real enforcer of any curriculum — educators call it the backwash effect. Whatever type of question appears in mock exams is what teachers drill; whatever teachers drill is the habit students graduate with. If the question bank measures recall, the classroom trains recall, whatever the curriculum says. That is why the ministry’s guide is so precise about what makes a question genuinely context-based:

Long passages, decorative images, everyday names — none of these makes a question context-based. The guide’s single functionality test asks: could the student answer without reading the context at all? If yes, the context is scenery — and the question must be rebuilt.

The “just have AI write it” trap

Generating one plausible question with AI is trivial. Generating a thousand that all satisfy fourteen criteria at once, without repeating each other — that is another business entirely. Ask a chat tool for twenty context-based questions: they look fine one by one. The flaws only surface under systematic audit:

1. 01

   Decorative context

   Remove the fancy passage and the question still solves — the guide’s core test fails.

2. 02

   Unverified answers

   The same mind writes the question and its key; nobody solves it independently.

3. 03

   Alignment drift

   Asked to measure reasoning, it measures recall. Looks right, misses the target.

4. 04

   Weak distractors

   Wrong options are random; they attract nobody, so the item stops discriminating.

5. 05

   Pattern collapse

   Twenty questions, one opening formula, the same scenario family throughout.

6. 06

   Fabricated metadata

   It invents curriculum codes that don’t exist. The label looks right; the referent doesn’t.

What we built: a production line for questions

SolEdu treats a question the way a factory treats a part. AI generates candidates aligned to the official learning outcomes; every candidate then passes through independent audit stations covering the guide’s fourteen criteria — context selection, ethics and neutrality, language, stem and option quality. A blind solver that never saw the intended answer solves each item independently; disagreement sends it back with a reason. Each new question is checked against the entire pool produced to date, so repeating scenarios and stock openings are eliminated before delivery.

Every accepted item ships print-ready in ÖSYM booklet format with its answer key, a künye whose every code derives from the official curriculum database — it cannot be invented — and a karne summarizing its audit history. Every decision is timestamped; even rejected questions stay on the record.

Production follows the guide’s own five-step loop: fix the learning outcome, construct the context, write the stem and options, engineer each distractor to represent a specific misconception, then run the functionality test — if the item solves without its context, it goes back to step two. Failed items are never simply deleted; they are re-engineered, and the loop count is itself a quality signal.

Sample item record (künye) — Kolay · COĞ.10.1.3.a · SB2 · KB2.8 · OB4 · SDB2.1

How it’s going

The system runs at scale today. The numbers below are live production figures, not projections — and they grow with every cycle:

10,000+

Questions produced and audited in the system to date

14 criteria

The guide’s checklist, audited item by item on every question

8 subjects · 9–11

All high-school grades that moved to the Maarif Model

Karne

Every question ships with a report card of its audit history

Publishers and schools receive audited sets in the format their typesetting systems expect, and can trace any question’s history on demand. “Quality” stopped being a claim and became evidence we can show. The road to 2028 is long — but the line is running, and every cycle it gets faster and more diverse. For us, SolEdu is the proof of a thesis: paired with engineering discipline, AI can carry real responsibility in a demanding, regulated domain.