What Separates a School Redesign That Lasts From One That Collapses?
From the Future Schools: Case Studies in School Innovation collection
Every few years a new school is announced as the answer to what is wrong with the old one. The launch is well covered: the founder's interview, the architect's drawings, the first cohort's early results. What happens five years later is covered far less, and that imbalance shapes how school leaders think about innovation. They see the openings and rarely the endings.
The usual debate is about the idea. Is project-based learning better than direct instruction? Should AI tutors take over part of the teaching day? These are fair questions, but the record of the last fifteen years suggests they are not the deciding ones.
Innovative schools rarely succeed or fail because the idea was bold or timid. They succeed or fail on which parts of the school they changed together, in what sequence, on what evidence, and whether parents, teachers and funding could absorb the change. That pattern can be read in advance, which is what makes one school's experience useful to another.
A school is a set of levers
Any school, conventional or experimental, can be read as settings on eight levers: time, space, grouping, curriculum, assessment, the teacher's role, technology, and the money-and-governance model. Two schools with similar slogans can sit at very different settings. Two schools with very different slogans can turn out to have moved the same few levers. Reading a case lever by lever turns a prospectus into a list of design choices, each of which can be examined.
The same reading separates redesign from rebranding. A structural change shows up in three documents: the weekly timetable, the staffing plan and the budget. An "AI lab" added to an unchanged timetable is a new room, not a new model.
Finland's 2016 curriculum reform is a useful example from the other direction. Headlines abroad reported that Finland was abolishing subjects, while the curriculum kept subject teaching and required schools to add at least one multidisciplinary, phenomenon-based unit each year. The reform was an addition layered onto an intact structure, and it deserves to be judged as one.
Results need filters before belief
Once the levers are clear, the next question is what the results actually show. Innovative schools are unusually good at telling their own story, and headline figures often pass through five filters that inflate them: self-reported gains, selective admissions, family wealth, survivorship and launch-day press. The practical test is to ask who measured, who enrolled, who paid and who left, before asking what the school did.
Two examples show the range. High Tech High in San Diego, open since 2000, reports an 82% college acceptance rate; that is the school's own figure, from a network that admits through a zip-code-weighted lottery rather than open enrolment or academic selection. Alpha School publishes test results it describes as beyond the 99th percentile nationally; those are also the school's own figures, from families paying around $40,000 a year at its Austin flagship, and critics note there is no control group or independent audit behind them.
A result from a selective, high-fee school is not the same result as one from an open-enrolment public school, even when the percentages look alike. None of this makes the figures false. It makes them answers to a narrower question than the one a school leader is usually asking.
Survivorship is the quietest filter. Schools that close stop publishing, so the visible record tilts towards the survivors. That is why closures and reversals are the most informative data in school innovation: they show where a model's weaknesses lay, often more plainly than any success story.
Two clocks
The closures in the record share a pattern. AltSchool, founded in 2013 by a former Google executive and reported to have raised about $175 million from investors including Andreessen Horowitz and Mark Zuckerberg, began closing sites in 2017. By 2019 it had handed its remaining schools to another operator and become a software company, Altitude Learning.
Summit Learning, personalised-learning software built with Facebook engineers, spread to more than 380 schools by its own count. Students and parents in Connecticut, Brooklyn and Kansas protested against it between 2017 and 2019. An analysis by the education news site Chalkbeat found that about 18% of schools left the programme each year, against the 10% implied by Summit's own retention claim.
Governments are not immune. South Korea launched AI digital textbooks nationwide in March 2025 at a reported public cost of over 1.2 trillion won. Media reports citing ministry and industry data put school adoption at 37% in the first semester and 19% in the second, and in August 2025 the National Assembly removed the software's official textbook status.
The common thread is scaling before proof. Venture investors expect growth within a few years, and political mandates run on electoral terms; a school model needs several cohorts of students before anyone can say what it does. When the faster clock sets the pace, growth outruns evidence, and children, parents and teachers carry the cost of finding out.
Adoption is the real test of a national bet
Where a government moves more carefully, the problem shifts from design to uptake. Estonia's AI Leap gave nearly all of its upper-secondary schools access to AI tools and teacher training from September 2025. By March 2026, reporting put student activations at about 7,700 of roughly 20,000 eligible, while around three-quarters of teachers had activated accounts.
Adoption data tells a school leader more than a launch announcement, because it measures what teachers and students chose to do rather than what was offered to them. A national bet is only as strong as its weakest link between the ministry and the classroom.
The historians David Tyack and Larry Cuban, in Tinkering Toward Utopia (1995), called the persistent structures of grades, periods, subjects and one teacher per room the grammar of schooling. Those structures endure because parents, universities, timetables and teacher training all depend on them. A reform that treats this pull as a failure of will tends to snap back; a reform that plans for it has a better chance of lasting.
The money is part of the model
A teaching model also stands on a funding model, and the second can fail first. Make School, which offered a computer-science degree financed through income-share agreements, closed abruptly in July 2021 and left 167 students mid-programme. Venture capital, income-share financing and state education savings accounts have all paid for new schools, and each can change faster than a school can adapt.
How a model will be paid for in five years' time therefore belongs in any assessment of it, alongside its pedagogy. A strong teaching idea tied to a fragile funding source carries a risk that families and staff eventually bear.
What lasts
The cases with long records point the other way. High Tech High grew over many years and built its own graduate school of education to train teachers in its approach. That patience is also why its results can be examined at all, including the questions about the depth of its mathematics teaching that its own staff have acknowledged.
Durable models tend to move a few levers in sequence and measure as they go. Failing models tend to move many at once and scale before the results are in. The difference is rarely one of vision; it is one of order and pace.
For an existing school, that points to a modest form: the school-within-a-school pilot. One grade, one studio or one programme with protected time; a clear hypothesis about what should improve; a comparison group of similar students; and an exit rule agreed in advance. A pilot built this way can fail cleanly, which is exactly what allows it to succeed honestly.
Read this way, school innovation stops being a test of courage or imagination and becomes a matter of design discipline: which levers, in what order, on what evidence, at a pace the people involved can bear. The conventional school is not the obstacle in that account. It is the base any new idea has to be added to, and the standard it has to beat.