The mind behind the face

How VAL thinks.

A frontier AI model is extraordinary, and entirely wrong to hand a child unmediated. VAL’s mind is the layer in between: the engineering that turns raw, adult-built intelligence into a companion a seven-year-old can safely talk to.

Why not just a chatbot?

You can’t hand a child a raw AI.

Frontier models are built for adults. Point one straight at a child and the problems are immediate, and familiar to anyone who has worked with them.

Built for adults

Their default tone, assumptions, and content are written for grown-ups, not seven-year-olds.

Sycophantic

They tend to tell you what you want to hear, the opposite of what a child forming judgement needs.

They hallucinate

Confidently wrong answers are a problem for anyone; for a child who trusts the voice, worse.

No memory discipline

A raw model has no sense of what to keep, isolate, or forget about this particular child.

A black box

Point one at a bedroom and parents get no visibility into what’s being said.

Rented safety

If your safety lives inside a model you don’t own, it shifts with every vendor update, and it can be jailbroken.

VAL answers every one of these with an engineered component, not a keyword blocklist bolted onto a chatbot. The architecture is the product.

The lifecycle of a single turn

Follow one “why?” all the way through.

Every time your child speaks, this happens, start to finish, fast enough to feel like a normal conversation.

  1. 01/ 06Who’s speaking

    Bound to your child

    The turn opens in your child’s own sealed, encrypted space. Nothing is shared, nothing is global.

  2. 02/ 06Context

    Only what’s relevant

    It gathers the right context, recent conversation, a longer understanding of your child, their stage, and nothing that isn’t.

  3. 03/ 06Think

    A frontier model drafts

    A best-in-class model generates a candidate reply, calibrated to your child’s age and to VAL’s character.

  4. 04/ 06Check

    Nothing spoken until it clears

    The candidate passes through several independent evaluators. Your child hears nothing until it clears. This is the part we’re proudest of.

    check · then · voice
  5. 05/ 06Speak

    In VAL’s voice

    Cleared speech becomes VAL’s warm voice, with a face that matches the meaning, streamed fast enough to feel live.

  6. 06/ 06Parents

    You stay in the loop

    Meaningful and sensitive moments are surfaced to a parent-facing layer. VAL is never a black box.

What makes it trustworthy

Clever is easy. Safe and kind is the hard part.

Owned, not rented

The intelligence is rented from the best frontier models and is swappable. The safety, character, and memory are VAL’s own. Models get better; the guarantees never regress when a vendor ships an update.

Per-child isolation

Every child has their own sealed, encrypted space. One child’s conversation can never surface in another’s. Privacy here is structural, not a policy promise.

Disciplined memory

VAL remembers what makes it a better companion to your child, and is deliberate about what it keeps and what it lets go. That continuity is what turns a device into a friend.

Parent transparency

A dedicated layer surfaces meaningful moments to you. You stay close to the friendship without hovering over it, and nothing important happens in the dark.

Built to feel live

All of this runs on every single turn, and still the conversation feels natural and responsive. Latency is treated as a first-class engineering discipline.

Componentized

Each concern is its own tested component. That’s exactly why the safety claims hold: they’re separable, verifiable pieces, not something tangled up inside a prompt.

The part we’re proudest of

Nothing is spoken until it clears.

VAL’s reply doesn’t pass through one safety check. It passes through several independent evaluators, each responsible for a single, different property, and each able to stop or revise the response before your child hears it. We call it check-then-voice: VAL never speaks a word to a child that hasn’t already passed safety.

Safety

Physically and emotionally safe for a child, recognising distress and responding with care, not just blocking words.

Accuracy

Is it actually true? A dedicated guard against the confident-but-wrong problem every frontier model has.

Honesty

The truth, not just what a child wants to hear, engineered against the people-pleasing default of raw models.

Age-fit

Right for this child’s developmental stage, not “safe for kids” in the abstract.

Independent of the model

The checks sit outside the model that writes the reply, so a model behaving badly never gets a free pass to your child.

Risk-adaptive

It grows more careful when a conversation calls for it.

Fails safe

When anything is uncertain, it errs toward caution.

Tech specs

The device, in detail.

The face

Type
Low-fi expressive character display
Purpose
Expresses only — no touch, no feed
Refresh
Real-time, synced to voice

Sound

Speaker
Full-range, tuned for speech
Microphones
Far-field array, physical mute
Voice
One warm, consistent voice

The mind

Intelligence
Best frontier models, rented
Safety
Owned by VAL, model-agnostic
Feel
Engineered to feel live

Connectivity

Wi-Fi
Dual-band 2.4 / 5 GHz
Bluetooth
LE for setup
Setup
Parent app, minutes

Power

Battery
All-day, rechargeable
Charging
USB-C, overnight
Nightlight
Warm, dimmable ring

In your hands

Materials
Soft-touch, child-safe
Size
Fits small hands and a bedside
In the box
VAL, USB-C cable, quick start

Privacy & safety

Mic control
One-touch physical mute
Data
Encrypted, yours to clear
Oversight
Parent app, always

Made for

Company
The most curious years
Grows with them
Deepens as your child does
Screens
None. Ever.

Specifications shown are indicative and evolving ahead of launch.

The best way to understand VAL is to talk to it.