Synthetic School: the white paper

A white paper. Every member of staff and every family in the school is an AI agent, and you lead it.

Summary

Synthetic School lets a school leader run a complete school for a full year, in software, before doing it for real. Every member of staff, every family and the board or authority above the school is an AI agent.

Most school leaders learn the job on real students and staff. Only 27% of new principals in New Zealand felt prepared when they started. In England, only 64.7% of new secondary headteachers under 50 were still in the same post or higher three years later.

The leader works from an inbox and a diary, within the hours a real week allows. The end of every week is a save point, so a leader can go back, decide differently and watch both versions of the school play out. A debrief shows what each decision caused, and nothing is scored or ranked.

The engine decides and the agents narrate, so every outcome traces back to a rule. A working prototype of the engine is calibrated to published attendance figures in two school systems. On 200 fresh synthetic schools in each, the fit holds within 0.2 points of every published figure.

Published research supports letting the engine decide while agents narrate, and it backs debriefs and shared random draws. No study yet links simulation-based leadership preparation to leaders' practice or to students' outcomes. A two-stage randomised pilot trial starts with the first: leaders' practice on scenarios they have not seen.

One engine runs every school system, and each system's rules live in a pack. The first two packs, for a public school system and for a regulated market of international schools, are in build. Five more systems, on four continents, have their sources mapped.

The same school serves leadership training first. It also works as a test harness for AI products meant for schools, and as a policy lab that runs a policy across hundreds of schools.

International schools alone number 15,075, with US$69.3bn in fees a year, and 38% are affiliated to groups. A programme seat is priced at US$450, and a hard model budget keeps model spend at or below 27% of that price.

The timing suits it. Rules for AI in schools are arriving, and some ask for testing before use. England's product safety standards expect testing with real users. EU obligations for high-risk education uses apply from 2 December 2027.

The ask: £250,000 in a SEIS-eligible equity round, which funds about 18 months of work to Gate 3.

The problem

School leaders learn the job on the job, with real students and real staff, and many say they started unprepared.

The job matters to students, though by how much is debated. A 2021 synthesis compared a 75th-percentile elementary school principal with a 25th-percentile one. The gain was 2.9 months of maths and 2.7 months of reading a year (Grissom, Egalite and Lindsay, 2021).

The Wallace Foundation now notes that those methods have been challenged. A 2024 re-estimate puts the gain nearer three weeks of maths and a little over a week of English (Wallace, 2025).

Preparation programmes teach the knowledge and rehearse single moments, such as a coaching conversation or a branching scenario. Project ALL comes closest, staging "a year in the life of a school principal" through more than 400 story elements (Mann and Shakeshaft, 2013). None that we found lets a new school leader branch a year-long school and compare both versions side by side.

What Synthetic School is

Synthetic School is a complete school in software, staffed by AI agents, which one person runs as its leader for a full school year.

The staff, the families and the board or authority above the school are all agents. Each has a role and a history, and remembers what the leader said to them last month. Every student has a full record, from attendance and attainment to additional support needs.

The leader works from an inbox and a diary, within the hours a real week allows. Every action costs time, and anything left unread grows older and harder to fix.

A trainee plays the weeks that matter and can delegate the rest, and the school keeps running either way. The end of every week is a save point. A leader can go back, make a different decision and watch both versions of the school play out.

Each term ends with a debrief. It shows how the people in the school saw the term and what followed from each decision. It also maps the leader's choices onto the leadership standard the programme uses.

Nothing is scored or ranked. The debrief is for learning, and the trainee decides who sees it.

Every pack ships with its own fictional school. The first two are Somewhere High, a state secondary in a public school system, and Somewhere International, a private secondary teaching a national curriculum from abroad. Neither is modelled on a real school or a real person.

How it works

The engine decides and the agents narrate. No model call changes the school's true state: agents propose actions, and the engine resolves them by rule.

The leader's chair, the facilitator console and the test harness are the only ways in. All three talk to one engine service, and models are reached only through its runtime, which checks and meters every call.

The leader's words are read into moves from a fixed list of 16, such as a promise or an overrule, and the leader sees every reading. Rules then turn each move into effects, starting with the trust of the person it was said to.

The leader sees the school only through its channels. Registers, reports, emails and conversations arrive late, partial or wrong, as they do in a real school. Reading through that gap is part of the job the product trains.

Child protection and other procedures set by law or national guidance are fixed steps in each system's pack. Agents can follow them well or badly, but no agent can skip, add or reorder a step.

Every outcome traces back to the rules, random draws and stored model replies that produced it. Each run starts from a seed, saves a snapshot at the end of every week and logs every event in between. Any run can therefore be replayed exactly, and branched from the end of any week.

Agents think only when something reaches them, so running cost grows with conversations and barely moves with the size of the school. Each moment goes to a model tier by how much it matters, and a cost meter holds every seat to its model budget.

The engine is calibrated against each system's published statistics. It is fitted to spread and relationships as well as averages, because a simulation can match a mean and still get the school wrong. Model versions are pinned for each run, and the validation suite runs again before any switch.

Built for every school system

Synthetic School runs one engine for every school system, and each system's rules live in a pack. The first two packs, for a public school system and for a regulated market of international schools, were chosen as proving grounds. Every later pack runs on the same engine.

Every school shares a skeleton: a timetable, a register, students with additional needs, staff on contracts, a budget, families, an inspector and someone the school leader answers to. A pack holds what differs:

• the calendar and terms, and the language the school runs in

• attendance codes and the published statistics the engine is calibrated to

• child protection procedure, and the law on suspension and expulsion

• the inspection or accountability framework, and the leadership standard the debrief maps to

• the funding model, qualifications and local job titles

• personas written from public, published material about that system, never from identifiable people

Every fact in a pack carries its source. A pack checker fails any procedure step with no owner or no route from the first step, and lists every open item.

The public frameworks a pack draws on already exist in every major system. Five more systems, on four continents, have their sources mapped so far, checked in October 2026. All five publish an inspection or accountability framework and a child protection law or policy with set reporting steps, and three publish a regular attendance series the engine can be calibrated to.

Where a system publishes no attendance series, a design partner supplies the target and the pack marks it as partner data.

There are four families of pack:

Writing a new pack is research work, with no change to the engine's code. The build plan estimates 15 to 25 hours of authoring and 3 to 5 hours of review by a serving school leader from that system. These are targets, to be replaced with timings from the first two packs. Before Gate 2, a child protection lead reviews each pack's storyline library, and repeats the review every year.

A cultural reviewer also reads the international-schools personas before Gate 2. Each pack ships with its own fictional school, and every one is called Somewhere.

One school, three uses

The same synthetic school serves leadership programmes, companies building AI for schools and the people who set education policy.

Training is the first market. In a survey of 177 international school leaders, published in 2025, 49.7% were in their first headship and 73.4% came from outside the school they lead (Morrison, 2025).

An incoming school leader can play a term in the same system and type of school before starting. A board session puts the leader and the board in the same school together, at US$2,500 per school.

The harness lets a vendor's agent observe and act in the school through a REST API or an MCP server, using four calls: register, observe, act and report. Every action is logged, and the run report shows what the agent did and what followed.

Two versions of an agent can face the same engine draws, and where their actions differ, the cast replies afresh. The vendor sees how the product behaves under a school's real pressures before any real student meets it.

The policy lab runs the engine without the cast, so it is fast. The prototype ran a year's attendance for 300 synthetic schools in about 5 seconds. Results stay labelled illustrative until validation supports stronger claims.

The engine's schools also produce case files: stand-alone data exercises, such as an attendance dataset with a question to answer, sold to programmes and individual leaders.

What the research says

Published research supports how Synthetic School is built more firmly than any claim it could make about leaders or students. Training studies back structured debriefs and realistic task demands, and simulation science backs an engine that holds state, with branches that share random draws. No study yet links simulation-based leadership preparation to leaders' practice or students' outcomes, and most of the evidence is from the United States.

Leaders' measured effects shrink, and most start underprepared

In six US value-added studies, a principal one standard deviation (SD) more effective added 0.13 SD a year in maths (Grissom et al., 2021). A 2024 re-estimate from two US states and New York City finds about 0.03 SD in maths (Bartanen et al., 2024, abstract read through a reproduction). Across 51 studies, principal behaviours were linked to school organisational health at 0.72 to 0.81 SD and to achievement at 0.08 to 0.16 SD (Liebowitz and Porter, 2019).

Only 31% of countries regulate induction for new principals (UNESCO, 2024). In the UK, "most headteachers do not feel fully prepared when they first take on the role" (Greany et al., 2025). Three months into the job, the volume and the unpredictability of problems were each salient for 59% of 17 new Chicago principals (Spillane and Lee, 2013, working-paper draft).

After three or more years, schools in six US pipeline districts that received new principals were 6.22 percentile points ahead of comparison schools in reading (Gates et al., 2019). RAND found little evidence that any single component was uniquely linked to the gain. Reviews call research on problem-based learning for school leaders "immature", and find digital-simulation studies limited by "weak research designs and methods" (Hallinger and Bridges, 2017; Hallinger et al., 2025).

For Synthetic School, simulated test-score gains are a weak measure of a trainee, so debriefs centre on practice and on teacher outcomes such as retention. A synthetic school can generate the volume and unpredictability novices describe.

In New Zealand, new principals who had coaching or mentoring, or a postgraduate leadership qualification, were twice as likely to feel prepared (Education Review Office, 2023, summary report). The simulation should sit alongside a coach, and none of this evidence supports a claim about students' outcomes.

Simulation training works, with smaller gains against active teaching

Across 609 health-professions studies, effects of simulation against no intervention fell from 1.20 for knowledge to 0.50 for direct effects on patients (Cook et al., 2011). In emergency medicine, effects against other instruction were 0.33 or less and not significant (Ilgen et al., 2013).

Across 145 higher-education studies, only seven from teacher education, simulation gave g = 0.85 for complex skills, and role play gave 0.63 (Chernikova et al., 2020). Against no debrief, debriefs improved performance by about 25% (d = 0.67) across 46 samples (Tannenbaum and Cerasoli, 2013). High- and low-fidelity simulators differed by 1% to 2% across 24 health studies (Norman et al., 2012).

No study has tested replays from a save point. Rapid cycle deliberate practice, which pauses and rewinds a scenario for coaching, has not proved superior to traditional simulation (Ng et al., 2021). A study of 40 airline pilots, carried out for the US Federal Aviation Administration, warns of "a potential hazard of too much contrast" between self-paced and surprising events (Landman et al., 2022).

For Synthetic School, claims should be pitched at the smaller effects found against active teaching. Objective review media "consistently contributed" to debrief effectiveness in a 61-study meta-analysis (Keiser and Arthur, 2021, abstract read through a reproduction), so the decision log serves as the debrief's objective record. No study compares AI-led with human-led debriefs, so the AI debrief needs its own test.

Some events should land mid-week without a pause, since weekly save points make every decision self-paced. Learning should be judged on new scenarios, because replays let trainees anticipate events.

School simulations are old, and their effects on real schools are untested

In a US study of 1958–59, 232 elementary principals each spent a week running a simulated school, Whitman, as a 1989 Korean review records (Shin, 1989; Hemphill et al., 1962). A 1992 meta-analysis of 34 validity studies of the NASSP principal assessment centre found criterion-related validity "very low" (Pantili, 1992).

Across five randomised trials at one US university, directive coaching between mixed-reality sessions gave an average 1.44 SD over self-reflection on blinded ratings of teaching quality (Cohen et al., 2021, working paper). Those gains were measured inside the simulator. A 2026 scoping review repeats an earlier warning that skills from mixed-reality simulation cannot be claimed to "generalize to real classrooms" (Mikeska and colleagues, 2019, as cited in Anderson and Calandra, 2026).

Project ALL, at Virginia Commonwealth University, stages "a year in the life of a school principal" through more than 400 story elements (Mann and Shakeshaft, 2013). The abstract of its study reports no outcome data (Mann et al., 2011).

Role play with AI now runs in pilots and projects in the US, Australia, Norway and Switzerland (Markel et al., 2023; Cornes, 2026; SLATE, n.d.; Dani, 2026). The tools range from language-model students to AI avatars and chatbots. GPTeach's A/B test compared the tool with a baseline in simulated office hours. The review found no randomised trial of any of them measuring outcomes in real classrooms or schools.

For Synthetic School, agents remove the live human actor, the cost bottleneck in mixed-reality training. One US study, of eight preservice teachers, found that a mixed-reality platform's simulations might have perpetuated overtly racist and sexist stereotypes (Bondurant and Reinholz, 2023), so personas need audits. Logs of what trainees do are a more honest record than their self-ratings.

Language-model agents reproduce averages better than individuals

Agents built from interviews with 1,052 US adults predicted their survey answers at 83% of each person's own two-week consistency (Park et al., 2024, revised 2026). Synthetic School's staff and families are invented, so its benchmark is the demographic or persona agent, at 71% to 74% on survey items.

Across four models, language-model answers flattened and misportrayed identity groups (Wang et al., 2025). In US survey data, synthetic samples had means near human averages but far less spread (Bisbee et al., 2024). Claude 1.3 wrongly admitted a mistake on 98% of challenged questions (Sharma et al., 2024).

AI-Agent School ran 50 teacher and student agents for five days, with no school leader or human player. It scored them against reference data largely generated by a language model and edited by experts (Jin et al., 2025).

On a text-world benchmark, GPT-4 predicted the full next state correctly in at most 59.9% of transitions that changed it (Wang et al., 2024). The Concordia framework keeps hard state, such as money and votes, in code (Vezhnevets et al., 2023), and no study has tested engine-decides against model-decides designs.

For Synthetic School, this is the strongest citable case for letting the engine decide. Model predictions of social-science experiments systematically overestimated effect sizes (Ashokkumar et al., 2026, read through the Stanford GSB listing), so agents should not judge whether an intervention worked. With the engine deciding, the documented failures fall on narration, which cannot change the school's state.

Narration still needs checks against engine state and audits by identity group, including students with additional support needs. A review of 35 generative social simulations found subjective tests such as believability still dominant, and asks for validity for the intended purpose (Larooij and Törnberg, 2026). The recognition panels therefore count as usability evidence.

Policy models stay suggestive

In a Chicago model, households that weighted test scores far above distance left new, better schools struggling to survive (Maroulis et al., 2010). The authors caution that their extended results rest on the assumption that new schools add more value than existing ones (Maroulis et al., 2014).

The closest precedent for a policy lab is a Washington State model of teacher sorting, in a CALDER working paper (Goldhaber et al., 2022). Calibrated on 564,296 teacher-year records, with every inequity pathway on, it reproduced observed exposure to novice teachers within 90% intervals.

Simulation science asks models to fit several observed patterns and to report the ones they miss (Grimm et al., 2020, building on Grimm et al., 2005). Giving each decision its own random stream let one agent-based study cut the runs needed by up to four orders of magnitude (Klein et al., 2024). Critics judge true prediction "infeasible for most social phenomena" (Edmonds and ní Aodha, 2019).

For Synthetic School, a single random generator would break the match between branches, so each decision needs its own stream. An identical-policy run, in which two branches with the same policy must give zero difference, tests that match.

Intervals from shared draws measure simulation noise only, so the lab also needs sensitivity analysis over behavioural parameters. Its results stay conditional on stated assumptions.

Attendance support cuts a few days a year

Persistent absence rose after the pandemic in every system the review covers. US chronic absenteeism went from 15% in 2019 to 28% in 2022 (Malkus, 2024). Across England's state-funded schools, persistent absence was 18.14% in 2024/25, against 10.86% in 2018/19 (DfE, 2026).

Randomised parent mailings in the US cut absence by up to about 6% to 8% (Rogers and Feller, 2018; Robinson et al., 2018). In a Chicago trial reported in a working paper, Check & Connect mentoring cut absences by 4.22 days, or 22.9%, in grades 5 to 7 (Guryan et al., 2020). Only New York City's Success Mentors reach five points, in a non-randomised comparison (Balfanz and Byrnes, 2013).

On the review's arithmetic, parent communication supports about 0.3 to 0.6 points of daily absence, and intensive mentoring about 2 to 2.5 points. Five points serves only as a labelled upper bound. Synthetic School's interactive demonstration now defaults to 2.5 points, with 0.5 and 5 points as cautious and optimistic cases.

The Chicago mentoring had no significant effect on grades or test scores. Text alerts to US parents cut course failures by 27% with no effect on state tests (Bergman and Chan, 2021). The engine should therefore not convert attendance gains into attainment automatically.

The engine's counterfactual should include the partial recovery expected of students picked for a recent dip. Absence also varies by day. In England, absence at the end of the week runs 20% above the rest of the week (Clifton-Sprigg and James, 2025), so the engine should not draw each day independently at a fixed rate.

On the same definition, persistent absence differs by about ten points between two neighbouring national systems (Scottish Government, 2025; DfE, 2026). Calibration to one system should not be assumed to transfer to another.

Evidence for each design choice

Ratings are for a school-leadership setting, where most of the evidence is indirect.

The full review, with every reference checked against its source on 18 October 2026, is in Literature review.

Evidence so far

A working prototype of the engine is calibrated to published attendance figures in two school systems, System A and System B. In each, it was fitted on a pool of 80 synthetic schools and then tested on 200 fresh ones, new seeds of the fitted model. The fit holds within 0.2 points of every published figure.

Published figures for System A: Scottish Government, School attendance, absence and exclusions statistics 2024–25.

System B's state-funded secondary schools in 2024–25 are the second test. Two of the four targets are published. The other two split attendance by eligibility for free school meals (FSM), worked out from the ministry's all-schools ratio because no secondary-only figure could be read.

Published figures for System B: DfE, Pupil absence in schools in England, 2024/25 (26 March 2026). FSM share of 25.85% in state-funded secondary schools: DfE, Schools, pupils and their characteristics, 2024/25. Ratio of FSM to other students' absence, all schools: DfE, Pupil attendance in schools, end of 2024/25 (7 August 2025).

The fit did not aim at two further figures. In the engine, students eligible for FSM are 2.73 times as likely as others to be persistently absent, against "over two and a half times" in the ministry's figures for all schools.

Across synthetic schools, persistent absence runs from 17.7% to 28.9% between the 10th and 90th percentiles. System B's 152 local areas run from 19.5% to 27.2% on the same percentiles, and single schools should vary more than area averages.

A calibrated fit shows the engine can reproduce published figures. Whether it predicts what a decision will do is tested at the later gates.

The prototype also shows the properties the product depends on:

• Replay: the same seed and pack give identical snapshots on every run.

• Speed: a 1,000-student school's attendance for a whole year runs in about 0.02 seconds, and 300 schools in about 5 seconds.

• Spread: across 300 synthetic System A schools, persistent absence runs from 32% to 43% between the 10th and 90th percentiles. System A's local areas ranged from 25.4% to 45.8% in 2024–25.

• Clean branching: two branches share every random draw until a decision differs. Measured that way, the effect of supporting 40 students varied a quarter to a third as much as when each branch drew its own luck.

The prototype covers attendance only. The full engine adds the cast, the leader's chair and the debrief, and each stage passes a gate before the next begins (see Roadmap).

Validation runs nine checks, and every gate uses at least one. Every pass mark is a starting figure, to be revised after the first panel round.

Calibration checks spread as well as averages, because synthetic survey samples have matched human means while varying far less (Bisbee et al., 2024). Blind mixes follow the Social Science Turing Test, one of four criteria for algorithmic fidelity (Argyle et al., 2023). Sensitivity analysis follows the call for models to report how far their results rest on their assumptions (Saltelli et al., 2020).

The pilot trial tests the first link between the product and practice. A two-stage randomised trial compares 12 weekly sessions of Synthetic School with case-based learning given the same time and the same coach-led debrief. Each session gives 60 minutes to one school week, or to a written case, then 30 minutes to the debrief.

Stage 1 checks feasibility with about 80 trainees, and Stage 2 tests efficacy with about 150 across at least two school systems. The primary outcome is practice on held-out scenarios, scored by raters who do not know each trainee's group. The full design is in Synthetic School: pilot evaluation design.

What the research says summarises the published evidence behind the design. No study yet links simulation-based leadership preparation to leaders' practice or students' outcomes. The trial does not measure students' outcomes either, because no trial of its size could detect them.

Safety, privacy and ethics

No real student, family, member of staff or school is in Synthetic School. The only personal data in the system belongs to the adults who play it.

Synthetic only:

• No import path accepts real student or staff records.

• No persona, storyline, case file or school is modelled on an identifiable person or school.

• Free text is screened for real names before it reaches a model or the log. A name that matches no one in the synthetic school is flagged, and the trainee is asked to remove it.

Formative only:

• Nothing leaves the system as a score or a ranking.

• The trainee owns the play record. A coach sees it only with the trainee's consent, and an employer only when the trainee shares it.

• Sessions played by a board or a leadership team show only the group's joint decisions.

• Individual measures exist only for research. They are collected with consent and shown to no trainee, coach or buyer.

• Every contract bars any use of play records to select, appraise, promote, discipline or rank a school leader.

Child protection content:

• Every child protection storyline comes from a reviewed library, and agents cannot start a new one.

• Harm reaches the school leader only as a report, the way a member of staff would record it, without graphic detail.

• No sexual content involving students appears in any form. Self-harm and suicide appear only as concerns and procedure.

• Facilitators choose which kinds of storyline a term may contain, and trainees see that list before they start. Self-serve trainees get child protection storylines only after opting in, and anyone can pause or leave a scene.

Data protection:

• Play records are stored in a UK or EU region, encrypted at rest, and kept for 12 months after a cohort ends unless the buyer's agreement sets a shorter period or the trainee keeps them. They are deleted on request, and automatically when retention ends.

• Model calls go to Anthropic in the US by default. A buyer who needs data kept in a region can use a regional cloud endpoint, at 10% more. The cost figures in this paper leave that premium out.

• Each market gets a data protection impact assessment before its first cohort, under its own data protection law. The privacy notice names the laws that apply.

Regulation of AI in education:

• The EU AI Act treats four education uses as high-risk: deciding admission, evaluating learning outcomes, assessing the level of education someone receives and detecting prohibited behaviour in tests (Annex III). The Digital Omnibus on AI, published in the EU's Official Journal on 24 July 2026, moved those obligations to 2 December 2027 (Garrigues).

• Synthetic School never infers a trainee's emotions, a use the Act has banned in education settings since February 2025 (Article 5).

• Whether a programme's use of the debrief counts as evaluating learning outcomes is a question for legal advice before any EU sale. The product is designed to stay outside that category, because it never scores or ranks a trainee.

• These rules give the harness part of its market. High-risk systems, such as the four education uses above, must be tested against defined metrics before they reach the market (Article 9). England's product safety standards, published as guidance, ask for testing "with a diverse and realistic range of potential users and use cases" (DfE).

• Abu Dhabi's ADEK now requires private schools' AI learning tools to meet its requirements (Abu Dhabi Media Office). Dubai's KHDA is developing AI guidelines for schools, expected to come into effect in 2027 (The National). China's 2025 guide asks education departments to keep a whitelist of generative AI tools allowed in schools (eol.cn).

• The US Department of Education left safety rules for AI in schools to states and districts in its 2026 grant priority (Federal Register). Ohio, for one, required every school district to adopt an AI policy by 1 July 2026 (Ohio Revised Code 3301.24).

The market

The first buyers are international school groups and leadership programmes. Public systems and AI companies follow, on the same engine.

Groups matter most at first, because one academy deal reaches dozens of schools. Several already run their own leadership programmes, among them Nord Anglia and Cognita with UCL.

Why now:

• The technology is ready. Agent simulations of people can now be tested against real people's answers. In the base case, 12 played weeks cost a seat about US$50 in model calls on Claude Sonnet, or US$81 on Claude Opus.

• Demand is forming. Rules for AI in schools ask for testing before use, from England's product safety standards to EU obligations for high-risk education uses from 2 December 2027. International school groups, meanwhile, are building leadership pipelines of their own.

The funding climate favours this shape of company. Edtech venture funding fell to US$2.4bn in 2024, its lowest level in a decade (HolonIQ, 2025).

Money has moved to AI that simulates people and to AI role-play training. Simile raised more than US$200m at a US$2bn valuation in July 2026. Yoodli and Second Nature raised Series B rounds of US$40m and US$22m in late 2025.

Business model

Synthetic School sells seats and licences, and a hard model budget keeps model spend at or below 27% of the seat price.

These prices are proposals to test with design partners. A US$450 seat sits below every leadership programme it would join.

Those programmes run from £650 plus VAT for HMC's Aspiring Heads' Programme to £2,840 for an International NPQH. Seven days at the Principals' Training Center cost US$2,325, or US$2,095 for member schools. The nearest product, SchoolSims, starts at US$149 a month for up to 10 participants.

Model calls are the main running cost. In the cost model's base case they take 11% of the seat price on Claude Sonnet and 18% on Claude Opus.

Heavy play can cost more, so each seat carries a US$120 model budget that the cost meter enforces. The target average is nearer US$60, to be confirmed in the Gate 1 cost tests. Near the budget, the meter trims background scenes and moves calls to cheaper models before it stops a run.

Pack tailoring is priced at cost for now, to win the first groups. A group pack layers the group's own policies and standards over its regulator's pack.

Competition

We found no venture-funded company that simulates a whole school with a person as its leader. The nearest products rehearse single conversations or branching scenarios, and the companies that simulate people with agents do it for market research.

Synthetic School competes on persistence and fog: one school in which consequences carry across terms, seen only through its own channels. That needs an engine underneath, which is slow to add to scripted content.

The main competitive risk is an incumbent adding language-model agents to an existing scenario library. Early design partners, published gate results and a growing set of calibrated packs are the answer.

Roadmap

The build runs in four stages, and each ends in something a buyer can use. Without funding, the full build is 200 to 310 hours of the founder's time with coding agents.

Each stage passes a gate before the next begins, on published figures, panels of serving school leaders and the build's own replay and cost tests. The results of each gate are published as a short note, so buyers and peers can check them.

Seed funding speeds the plan up and brings in work one founder cannot do alone.

Founder and ask

Synthetic School is built by one founder with AI coding agents. Seed funding adds the people its gates depend on, and runs packs and validation alongside the build.

The founder is Jamie Scobie, a secondary school senior leader who leads on attainment and inclusion and builds the data platforms his colleagues use every day. He follows the research and policy debate on AI in education and writes about it. He holds two master's degrees with distinction, one of them from the University of Cambridge.

JS is employed in public education. The position on intellectual property and outside work will be confirmed in writing before anything is sold, which is the first gate's condition.

People the plan needs, each for a bounded piece of time:

The ask: £250,000 in a SEIS-eligible equity round, with HMRC advance assurance, for about 18 months of runway to Gate 3. Use of funds: engineering and model costs 40%, packs and validation panels 35%, legal, insurance and operations 15%, and 10% held in reserve. A sales lead is hired later, after about ten paying programmes, from revenue or the next round.

Contact: hello@synthetic.school.

References

Each source below was opened between 16 and 18 October 2026. On 18 October an independent check opened every cited source again, compared it with each sentence that cites it, and the text was corrected to match. Sources read only through a summary, a reproduction or a draft are labelled.

The problem and the training market:

• Education Review Office (2023), Everything was new: preparing and supporting new principals

• Education Review Office (2023), Everything was new: summary report

• Darling-Hammond, Wechsler, Levin, Leung-Gagné and Tozer (2022), Developing effective principals: what kind of learning matters?, Learning Policy Institute and Wallace Foundation

• NCES (2023), Roughly One in Ten Public School Principals Left Profession in 2021-22 School Year

• Department for Education (2025), School leadership retention

• Australian Catholic University (2026), Violence against principals nearly doubles in 15 years as workload soars, stress and burnout deepens, ACU study (news release, 30 March)

• Grissom, Egalite and Lindsay (2021), How principals affect students and schools: a systematic synthesis of two decades of research, Wallace Foundation

• Wallace Foundation (2025), Why measuring principal impact is harder than you think

• Morrison (2025), Survival at the top: factors impacting tenure and reasons for departure of the international school head, Journal of Research in International Education 24(2)

• Bartanen, Husain and Liebowitz (2024), Rethinking principal effects on student outcomes, Journal of Public Economics 234, 105115, abstract read through a reproduction

• Liebowitz and Porter (2019), The effect of principal behaviors on student, teacher, and school outcomes, Review of Educational Research 89(5)

• UNESCO (2024), Leadership in education: lead for learning, Global Education Monitoring Report 2024/5 summary

• Greany et al. (2025), Signature pedagogies for leadership development, London Review of Education 23(1)

• Spillane and Lee (2013), Novice school principals' sense of ultimate responsibility, IPR Working Paper WP-13-05, January draft

• Gates, Baird, Master and Chavez-Herrerias (2019), Principal pipelines: a feasible, affordable, and effective way for districts to improve schools, RAND RR-2666

• Hallinger and Bridges (2017), A systematic review of research on the use of problem-based learning in the preparation and development of school leaders, Educational Administration Quarterly 53(2)

• Hallinger, Banoğlu and Gümüş (2025), Scoping review of research on digital simulations and serious games in educational leadership and management, Leadership and Policy in Schools 24(1)

• Mann and Shakeshaft (2013), Virtual school, real experience, Journal of Staff Development 34(6)

• Mann et al. (2011), Immersive, interactive, web-enabled computer simulation as a trigger for learning, Journal of Research on Leadership Education 6(5)

Calibration data:

• Scottish Government (2025), School attendance, absence and exclusions statistics 2024–25

• DfE, Pupil absence in schools in England, academic year 2024/25 (26 March 2026)

• DfE, Schools, pupils and their characteristics, academic year 2024/25

• DfE, Pupil attendance in schools, end of the 2024/25 academic year (7 August 2025)

Research:

• Cook et al. (2011), Technology-enhanced simulation for health professions education, JAMA 306(9)

• Ilgen, Sherbino and Cook (2013), Technology-enhanced simulation in emergency medicine, Academic Emergency Medicine 20(2)

• Chernikova et al. (2020), Simulation-based learning in higher education: a meta-analysis, Review of Educational Research 90(4)

• Tannenbaum and Cerasoli (2013), Do team and individual debriefs enhance performance? A meta-analysis, Human Factors 55(1)

• Norman, Dore and Grierson (2012), The minimal relationship between simulation fidelity and transfer of learning, Medical Education 46(7)

• Ng, Primiani and Orchanian-Cheff (2021), Rapid cycle deliberate practice in healthcare simulation: a scoping review, Medical Science Educator 31

• Landman et al. (2022), Negative transfer of training: simulator study into effects of going beyond alarms during stall recovery training, DOT/FAA/TC-22/10

• Keiser and Arthur (2021), A meta-analysis of the effectiveness of the after-action review (or debrief) and factors that influence its effectiveness, Journal of Applied Psychology 106(7), abstract read through a reproduction

• Shin (1989), Simulation game in organizational development of school, Gyoyuk Nonchong (교육논총) 8, Kookmin University

• Hemphill, Griffiths and Frederiksen (1962), Administrative performance and personality: a study of the principal in a simulated elementary school, Teachers College, Columbia University

• Pantili (1992), A meta-analytic validity study of the National Association of Secondary School Principals' Assessment Center process, PhD dissertation, Virginia Tech

• Cohen, Krishnamachari and Wong (2021), Experimental evidence on the robustness of coaching supports in teacher education, EdWorkingPaper 21-468

• Anderson and Calandra (2026), Mixed-reality simulations in teacher education: a scoping review, Contemporary Issues in Technology and Teacher Education 26(2)

• Markel et al. (2023), GPTeach: interactive TA training with GPT-based students, EdArXiv preprint; peer-reviewed version in Learning @ Scale 2023

• Cornes (2026), Transforming the future of leadership development through virtual reality and artificial intelligence, Faculty of Education, University of Melbourne (news item, 29 May)

• SLATE, University of Bergen (n.d.), PRACTICE project page

• Dani (2026), Practice parent-teacher meetings with AI avatars, DIZH (news item, 9 March)

• Bondurant and Reinholz (2023), How mixed-reality simulations can perpetuate racist and sexist stereotypes, Mathematics Teacher Educator 11(3), 189–209

• Park et al. (2024, revised 2026), LLM agents grounded in self-reports enable general-purpose simulation of individuals, arXiv 2411.10109

• Wang, Morgenstern and Dickerson (2025), Large language models that replace human participants can harmfully misportray and flatten identity groups, Nature Machine Intelligence 7(3), preprint arXiv 2402.01908

• Bisbee et al. (2024), Synthetic replacements for human survey data? The perils of large language models, Political Analysis 32(4)

• Sharma et al. (2024), Towards understanding sycophancy in language models, ICLR 2024

• Jin et al. (2025), Evolution in simulation: AI-Agent School with dual memory for high-fidelity educational dynamics, Findings of EMNLP 2025

• Wang et al. (2024), Can language models serve as text-based world simulators?, ACL 2024

• Vezhnevets et al. (2023), Generative agent-based modeling with actions grounded in physical, social, or digital space using Concordia, arXiv

• Ashokkumar et al. (2026), Large language models can predict the results of social science experiments, Nature 656, read through the Stanford GSB listing

• Larooij and Törnberg (2026), Validation is the central challenge for generative social simulation: a critical review of LLMs in agent-based modeling, Artificial Intelligence Review 59(1), preprint arXiv 2504.03274

• Argyle et al. (2023), Out of one, many: using language models to simulate human samples, Political Analysis 31(3)

• Saltelli et al. (2020), Five ways to ensure that models serve society: a manifesto, Nature 582

• Maroulis et al. (2010), Complex systems view of educational policy research, Science 330

• Maroulis et al. (2014), Modeling the transition to public school choice, Journal of Artificial Societies and Social Simulation 17(2)

• Goldhaber et al. (2022), How did it get this way? Disentangling the sources of teacher quality gaps through agent-based modeling, CALDER Working Paper 259-0122

• Grimm et al. (2005), Pattern-oriented modeling of agent-based complex systems, Science 310, abstract read on the USGS record

• Grimm et al. (2020), The ODD protocol for describing agent-based and other simulation models: a second update, Journal of Artificial Societies and Social Simulation 23(2)

• Klein et al. (2024), Noise-free comparison of stochastic agent-based simulations using common random numbers, arXiv version 2

• Edmonds and ní Aodha (2019), Using agent-based modelling to inform policy: what could possibly go wrong?, Multi-Agent-Based Simulation XIX, Lecture Notes in Computer Science 11463

• Malkus (2024), The latest chronic absenteeism numbers, American Enterprise Institute

• Rogers and Feller (2018), Reducing student absences at scale by targeting parents' misbeliefs, Nature Human Behaviour 2, details read in an earlier author version titled "Reducing Student Absences at Scale By Involving Families"

• Robinson, Lee, Dearing and Rogers (2018), Reducing student absenteeism in the early grades by targeting parental beliefs, American Educational Research Journal 55(6), read in the author copy

• Guryan et al. (2020), The effect of mentoring on school attendance and academic outcomes, IPR Working Paper WP-20-40

• Balfanz and Byrnes (2013), Meeting the challenge of combating chronic absenteeism, Everyone Graduates Center, Johns Hopkins University, an undated report released in November 2013

• Bergman and Chan (2021), Leveraging parents through low-cost technology, Journal of Human Resources 56(1)

• Clifton-Sprigg and James (2025), The "Friday effect": school attendance over the week, British Educational Research Journal 51(1)

Regulation:

• EU AI Act, Annex III, Article 5 and Article 9

• Garrigues, on the Digital Omnibus regulation

• DfE, Generative AI: product safety standards

• Abu Dhabi Media Office (17 September 2026), on ADEK's AI policy for private schools, The National (1 October 2026), on KHDA's AI guidelines and eol.cn, the 2025 guide to generative AI in schools

• US Department of Education (13 April 2026), Final priority and definitions: Secretary's supplemental priority and definitions on advancing artificial intelligence in education, Federal Register 2026-07087 and Ohio Revised Code, section 3301.24, Artificial intelligence policy, effective 30 September 2025

Market and competitors:

• ISC Research 2026 snapshot, via School Management Plus and ISC Research ten-year figures

• NCES Digest, table 212.30

• DfE, school workforce data and NPQ starts 2025/26

• Learning Forward, federal funding update on ESSA Title II-A

• HolonIQ (2025, 15 January), 2024 edtech venture capital note

• Prices: HMC, International NPQH, Principals' Training Center, SchoolSims

• Funding rounds and companies are linked where they appear in Market and Competition.