© 2026 Vallar LLC

Safe AI in education

Vallar is an AI research lab focused on education. Our vision: AI that strengthens how young people learn, with safety as its foundation, not a feature. Milana, our learning platform for schools, is almost here.

Focus

Not shortcut AI: AI for learning

In a classroom, a right answer is not enough. You ask to see the work. You check the reasoning. You notice when someone is guessing.

Our research is about AI that does the same: it helps people understand, not just finish.

An answer is easy. Understanding is the work.

Why it matters

AI can write the essay. It can solve the equation. It can answer almost anything. But an answer is not understanding. The gap between them is where we work.

“AI has become humanity’s biggest comfort zone. I believe it is quietly becoming the reason we stop thinking for ourselves.”

Hendrik Abicht, founder of Vallar
  1. We searched.

    We went to a library, hunted the shelves for the right book and read our way to an answer.

  2. We chose.

    Google found the pages for us. We still read them, weighed them and decided what to trust.

  3. We ask.

    We type a question. A finished answer comes back.

Each step saved us effort. People who expect information to stay saved tend to remember less of it1, and searching once makes people likelier to search again, even for easy questions2. Learning from AI summaries instead of web links leaves people reporting shallower knowledge3. In a randomised trial, students who practised maths with unrestricted GPT-4 later scored 17% lower on an unaided exam; a tutor version that gave hints instead of answers showed no such drop4. The effort is the point: actively recalling what we learn strengthens memory5, and the adult brain measurably adapts to demanding training6.

  1. Sparrow, Liu & Wegner (2011), Science; Schooler & Storm (2021), Memory
  2. Storm, Stone & Benjamin (2017), Memory
  3. Melumad & Yun (2025), PNAS Nexus
  4. Bastani et al. (2025), PNAS
  5. Roediger & Karpicke (2006), Psychological Science; Rowland (2014), Psychological Bulletin
  6. Woollett & Maguire (2011), Current Biology

Open questions

No term here is settled. That is why it is research.

Question 1 of 5. Calibration: Can it say I don’t know?

How we work

Five principles. Everything we build has to pass all five.

  1. Safety as foundation
    not feature

    Built in from the first line, not added after something goes wrong.

  2. Strengthen
    don’t outsource

    Every exchange should leave a student thinking more, not less.

  3. Adults in the loop
    never out of it

    A teacher or a parent always sees, steers and decides. The AI assists and reports up.

  4. Compliance as floor
    not ceiling

    The law sets the minimum. We build to a standard nobody enforces yet.

  5. Transparency as a
    publishable position

    What binds us should be public, for schools, parents and regulators to read.

The wall

In ancient times, a vallar was a protective wall, built to guard what mattered most.

Six layers, built to surround every student.

  1. Layered models

    Built to run only on models with strong safety records, shaped for learning and for each student’s age.

  2. Checked both ways

    Built to check every question in and every answer out, with a teacher or parent in the loop.

  3. Thinking, not shortcuts

    Built to strengthen how students think, not to do the thinking for them.

  4. Calm by design

    No streaks, no pure white: an interface built for young eyes and attention.

  5. Data that stays

    Built to keep student data in the region the school signs for, and out of model training.

  6. Rules, to be published

    Our commitments, to be published as the Vallar Charter for schools, parents and regulators to read.

Vallar is the wall. Milana is the guide it protects.

From the lab

Fig. — Withheld until release

Milana

A learning platform for schools, by Vallar. Almost here.

Contact

Got a question we should be asking?

Questions welcome. Answers in progress.