Gage.Education · A virtual school for K-12

Before anything else

What your child actually does on a Tuesday.

Not a philosophy. The morning, in order.

Today

Craft · 8-10 · I can get good at things on purpose

Good morning, Ana.

One thing at a time. Everything else can wait.

Your next action

A Cycle of math

Quantity, Shape and the Counting World · 25 minutes, then you stand up and move before the next one unlocks.

Begin the Cycle

0nodes proven

0awaiting proof

0cycles today

The whole screen, first morning

This is all of it. No feed, no streak, no leaderboard, no badge, and nothing at all about any other child. The three counters read zero because Ana started today, and they will keep reading zero until she proves something, which cannot happen for eight days.

A school that wanted you to feel good on day one would have found something to put there.

They open one screen and it tells them one thing to do. Not a menu, not a dashboard, not eleven notifications. One card: here is the next thing, press this. If they were interrupted yesterday it says so and offers the same questions back.
Eight questions. About twelve minutes. Chosen so they get roughly four out of five right: hard enough to be worth doing, easy enough to keep going. Get one wrong and the tutor asks a question back. It will not tell them the answer, and that is enforced by the school rather than by the tutor’s good manners.
Then they stop, and move. The next Cycle will not open until they have. Screen minutes are capped by age and the cap is visible on the screen, to them, all day.
Nothing they learned today counts today. It comes back at least eight days later, unaided, mixed with other topics, on questions they have never seen. That is the only thing that enters the Ledger. It is slower than every other school and it is the entire point.
The afternoon is not this screen. A brief, a material, and a thing that exists in the world when they are done. A person reads what they made and writes back.

One Cycle · eight questions

Easier than that and nothing is being learned. Harder and a child stops. The two they miss are the reason the Cycle is worth doing.

One skill · four months

Only the gold rings enter the Ledger, and each one is taken alone, with no tutor, on questions never seen before. Day zero counts for nothing at all.

Start school, free

And then the screen goes away

The afternoon is not this screen.

A brief, a material, and a thing that exists in the world when they are done.

An overhead view of a wooden table in afternoon light. A model bridge made
of cardboard, paper, string and masking tape stands on two towers. Below it
a hand-drawn plan marks the span at thirty centimeters and the towers at
twelve, with the materials listed in pencil. A ruler, twine, scissors, a
pencil and offcuts lie around it. — An afternoon brief, built. No child appears in this photograph, and none will. We have no students yet, and we would not publish their faces if we did.

Every afternoon has a brief like this one: a bridge that spans thirty centimeters and holds its own weight, out of cardboard, paper, string and tape. They draw it before they build it, because a plan you have to change is where most of the thinking happens.

Nothing about it is graded by a machine. A named person reads what they made and writes back. It is the half of the day a child will actually tell you about at dinner.

This is what we mean by do. Not a simulation of building something. The thing, on a table, that falls over if the plan was wrong.

What the whole school rests on

Four commitments. Break one and this is just another platform.

The tutor never answers.

Not “usually doesn’t.” Cannot. Enforcement lives below the model, in the platform, untoggleable by the student.

Unguarded AI: practice +48%, unaided exam −17%. Guarded: +0.73-1.3 SD.

Nothing counts until you can do it alone, later.

Unaided, delayed at least eight days, on items you have never seen, mixed with other topics, then re‑proven at +30 and +120 days.

Interleaved delayed testing: 61% vs 38%, d = 0.83.

The adult is the source of momentum, not information.

AI takes instruction, grading, planning and assessment. The adult takes the one thing every trial says is decisive.

AI alone: 0.06 SD/yr. AI with a facilitating adult: 0.31 SD in six weeks.

We publish the audit before anyone asks.

Pre‑registered, independent, annual, including the years it goes badly, and the attrition figures selective schools omit.

No competitor in this category has ever done this.

The Atlas

A report card is one number pretending to be a child.

This is what replaces it: a living graph of 2,383 nodes joined by 2,378 prerequisites, of which yours is a unique constellation. Gold is proven. Blue is in progress. Gray is not yet reached. Hover any node.

proven in progress not yet

The Compass is adaptive testing against the graph: roughly twenty‑five questions locate a student across a three‑hundred‑concept domain. Twelve minutes, once, and the year’s plan writes itself. It is never scored, never shown as a grade, never used to sort children into fixed groups.

A student blocked in fractions keeps moving at full speed in history, in Spanish, in the Studio. Blockage is node‑local, never person‑global. That single rule eliminates most of what schooling does to a child’s sense of themselves.

Ori · the tutor

Try to get the answer out of it.

Every hint Ori gives is one rung on a fixed ladder. There is no rung eight. Press the button as many times as you like.

node MATH.FRAC.EQUIV.02 · grade 4 hint rung 0 / 7

This is not a system prompt being polite. The item’s solution set, and its paraphrases, its near‑variants, and the final‑step expressions that make it trivially derivable, is computed when the question is generated and held as a blocklist outside the model. Any output that matches is blocked and regenerated one rung lower. Prompt injection reaches the tutor and dies at the wall behind it. That wall is the product. The model underneath is the cheap part.

The labs

Not a picture of the idea. The idea, moving.

Type a quadratic and watch the parabola bend as you write it. Open a cell and watch a mitochondrion spend energy. Launch the same ball twice with one number changed and let the arcs argue. Fractions as lengths you can compare with your eyes. Every lab is hand‑drawn vectors inside the same file the school already is: nothing is fetched, nothing is tracked, and every one prints its facts as text for the child who cannot see the picture.

The labs live beside the skills they illustrate and open nowhere near a Proof, because nothing counts until it is done alone, later. Try them now:

The Grapherthe curve answers as you write The Cellevery part alive, every part with a job The Launcherchange one number, watch the arc move The Bar Modelfractions as lengths, seen at a glance

The audit

We said we publish it before anyone asks. Here it is.

We audited this school against 71 standards drawn from nine frameworks. Cognia, the Middle States AI endorsement, the Council of International Schools, NEASC, the EU AI Act, UNESCO, Section 504 and the ADA, and the special-education standards the four largest US accreditors issued jointly. A standard counts as met only when it is written down, built into the software, and something measures whether the control is working.

80%readiness, on our own audit

43standards fully met of 71

44automated gates block every release

0children who have finished a year here

What the audit says we get right

A child who knows nothing is certified on 0.2% of skills in a year. It was 22% until we measured it and rebuilt the Proof
Every consequential decision routes to a named human, automatically
The tutor cannot give the answer, and the rule is enforced outside the model
The engine is tested for its own bias on every build, not only the child’s

What it says we get wrong

No outcome evidence of any kind. Zero children have finished a year
Two measured biases are open and unfixed, and named above
The governing board exists on paper and has never met
No synchronous instruction of any kind. Some states require it
We cannot detect disparate impact on any protected group, because we collect no demographic data. That refusal is deliberate and it has this cost

This is our own audit of ourselves, which is the weakest kind. It is a starting point for accreditation, not a substitute for it, and we would rather hand you the failures than have you find them.

Where this is

Beta testing. Free, and unfinished in ways we will name.

There is no price on this page because there is nothing yet to charge for. A school earns tuition by teaching children well, and we have not taught anyone yet. Use it, break it, and tell us what happened.

What works today

104 courses, K-12, across 8 strands
1,885 practiceable skills inside 2,383 mastery nodes, on an acyclic prerequisite graph
Adaptive placement in about twenty-five questions
Practice held in the 80-85% band, and spacing that schedules itself
Proofs that only count unaided, days later, on questions you have not seen
Runs with the network unplugged

What is still rough

Nobody has completed a full year, so we have no outcome data and will not imply otherwise
The tutor runs an authored hint ladder while its model key is pending
The Guide corps is one person and a queue, not yet a corps, and the governing board has never met
We are not accredited. We audit ourselves against nine frameworks and publish the score, which is not the same thing
26% of Kindergarten questions read three or more grades above the child they are for. Measured, published, unfixed
The standards crosswalk covers Mathematics only, and no certified curriculum specialist has reviewed it

What we ask of you

Let a real child use it for a real week
Tell us where they stalled, got bored, or got stuck
Tell us what a question asked that it should not have
Keep your own record; export is one click and always will be

When there is a price, it will be set by one thing: how many children each adult can know by name. Software costs almost nothing to run and we will not pretend otherwise. Until a cohort has finished a year and an independent evaluator has looked at the result, the honest number is zero.

Start school

For families

The questions people actually ask us.

How old does my child have to be?

Five to eighteen. The grade you enter sets the rhythm and the screen cap, not the content. The Compass finds the real level in each subject separately, and for most children those differ by two years or more.

My child is behind in one subject and ahead in another.

That is the ordinary case and the reason this exists. There is no year group here. A child can be doing year-six fractions and year-three spelling on the same morning, and nothing on any screen tells them that is unusual.

Do I have to teach?

No. You have to be present. The evidence is consistent and slightly annoying about this: the software does the instruction, and the adult in the room is what makes it work anyway. Twenty minutes of genuine attention beats two hours of supervision.

What if my child gets stuck and I cannot help?

The tutor works down seven levels of hint. If the seventh does not do it, it stops and fetches a person, and it tells you it has. Being stuck is treated as information, not as failure.

What about a child with dyslexia, ADHD, or anxiety about math?

Short bursts, one thing on screen, no timer visible, no class average anywhere, and motion can be turned down in Settings. Every question can be read aloud by the device, and nothing is scored on speed. We are not going to claim it is a clinical intervention. It is not. But almost nothing here rewards the things those children find hardest.

Is my child’s data safe?

Everything lives in your browser until you choose to make an account. No camera, no microphone, no attention tracking, no emotion detection, no advertising identifiers, no third-party analytics. One click exports everything; one click erases it.

What does it cost?

Nothing, now. This is an open beta and there is no tuition, because a school earns tuition by teaching children well and nobody has finished a year with us yet.

What if we stop?

You take the record with you. Export is one click, it is a plain readable file, and it works whether or not you ever made an account.

Start school, free What is still unfinished

The evidence, including the parts against us

We will not say two sigma.

Bloom’s 1984 figure is the most‑cited number in education technology and has never been replicated. Every school quoting it is quoting a legend. Here is what the record actually supports.

Finding	Effect	What it tells us
Unguarded AI tutor, ~1,000 students PNAS 2025	−17%	Practice scores rose 48% while unaided exams fell. Access without a guardrail is worse than nothing.
AI tutor alone, 2 yrs, 18 schools NBER 2026	0.06 SD	“A non‑event,” in its own founder’s words. Engagement, not access, is the binding constraint.
Teacher‑facilitated AI, Nigeria World Bank 2025	0.31 SD	Six weeks ≈ 1.5-2 years of ordinary schooling, at $48/student. The adult is the ingredient.
Guarded AI tutor, Harvard RCT 2025	0.73-1.3 SD	Roughly twice best‑practice active learning, per hour. The ceiling we design toward.
Human one‑to‑one tutoring VanLehn 2011	0.79 d	The honest benchmark. Not two sigma.
Mastery learning 1987-1990	0.04-0.52	An unresolved order‑of‑magnitude dispute. We report the range, not the flattering pole.
Contingent rewards on motivation Deci 1999	−0.40 d	Worse in children than adults. Why this school has no points store.

Research informs the design. Gage does not yet have measured learner outcomes. Studies of other programs cannot establish our results. Learning gains, pacing and lasting understanding must be measured in our own pilot and independently evaluated before we claim them.

And a clause in the family compact and in the evaluator’s contract: if the audit shows we are not beating the alternatives, we publish that, say what we are changing, and refund the year on request. A school that has to be right builds differently from one that only has to be persuasive.

They ask you to believe them.

We publish the audit, name every guide, ban the camera, and charge a seventh as much.

One thing to know before you enroll

This curriculum runs ahead of the Common Core sequence, and by design. Grade 2 here teaches multiplication and unit fractions, which Common Core places in grade 3. By grade 7 the mathematics is high-school content: quadratics, formal proof, circle theorems.

That is good for a child who stays. It matters if you leave: a receiving school placing your child by age will have no way to read what they have actually done. Our transcript states the external standard behind every unit and the grade that framework places it in, so a registrar can see the offset rather than guess at it.

We would rather you knew this now than discovered it at a placement test.