Published research grounds Synthetic School's design choices

Published research supports how Synthetic School is built more firmly than anything it might claim about leaders or students. Training studies in medicine and aviation back structured debriefs and realistic task demands over visual polish. Social-simulation research backs holding outcomes in a rules engine and comparing branches that share random draws.

No study links simulation-based leadership preparation to changes in leaders' practice or students' outcomes, and reviews find weak research designs in leadership simulation studies. The engine's optimistic case, a five-point cut in daily absence, sits at the top of the published range, matched only by one non-randomised New York City estimate. Most of the evidence comes from the United States.

Each section names the system behind each finding and ends with the design choices the evidence supports or challenges. Findings read only through a summary or a draft are labelled in the text, including abstracts reproduced on third-party sites. A table after the six sections maps every design choice to its evidence.

Principal effects fall from 0.13 to 0.03 SD as research designs tighten

Estimates of leaders' effects depend on how they are measured

The most-cited synthesis rests on US administrative data. Grissom and colleagues pooled six value-added studies covering more than 22,000 principals in four states and two urban districts (Grissom et al., 2021). A principal one standard deviation (SD) more effective raised students' achievement by 0.13 SD in maths and 0.09 SD in reading each year.

For elementary principals, moving from a 25th- to a 75th-percentile principal adds about 2.9 months of maths learning and 2.7 months of reading a year (Grissom et al., 2021). The authors call this "nearly as large as the effect of having a similarly effective teacher". They also warn that the six studies come from "just a few states and districts, which may not be representative".

Texas data show the estimate moving with the design (Branch et al., 2013). Comparing schools gives a spread of 0.21 SD in principal effectiveness, which the authors call "perhaps unbelievably large". Within-school and leadership-transition designs reduce it to slightly over 0.10 SD and about 0.05 SD.

A 2024 re-estimate in the Journal of Public Economics draws on statewide data from Tennessee and Oregon and on New York City records. Its abstract, read here through a reproduction, puts persistent principal effects at roughly 0.03 SD in maths and 0.01 SD in reading (Bartanen et al., 2024). The 2023 working-paper version puts its estimates at 40 to 60 per cent below the Texas lower bounds (Bartanen et al., 2023).

The authors argue that standard models credit principals with school changes they inherit, such as teacher turnover (Bartanen et al., 2023). A placebo test that reassigned principals across schools barely changed the estimated variance. They judge attempts to identify good principals "through student outcomes" to be "fundamentally flawed", although 46 US states use school outcomes in principal evaluation.

Practice-focused meta-analyses report larger associations on a different scale. Robinson and colleagues reviewed 27 studies, 18 from the US. In the 12 that compared sets of leadership practices, leaders promoting and participating in teacher learning showed an average effect of 0.84, from 17 effect sizes in six studies (Robinson et al., 2008). Those effects came mainly from comparing otherwise similar high- and low-performing schools, and the authors urge "cautious interpretation" of the apparently strong difference between instructional and transformational leadership.

Liebowitz and Porter reviewed 51 studies, and their abstract does not state which systems the studies cover (Liebowitz & Porter, 2019). Principal behaviours were linked to achievement at 0.08 to 0.16 SD and to school organisational health at 0.72 to 0.81 SD. The authors describe the causal evidence base on leadership behaviours as "nonexistent".

Turnover studies give a more causal signal. In Missouri and Tennessee, a principal's departure lowered school achievement by 0.03 SD the following year, and an experienced successor largely offset the drop (Bartanen, Grissom & Rogers, 2019). The design compared affected schools with matched schools before and after the change.

Outside the US the evidence is associational. Across more than 1,800 schools in eight countries, the US among them, "higher management quality is strongly associated with better educational outcomes" (Bloom et al., 2015). This review found no causal estimate of school leader effects for England, Scotland, Australia or any Asian system.

For Synthetic School, these estimates argue against judging trainees mainly on simulated test-score gains. Real principals' persistent effects on scores are small and hard to recover from outcome data. Associations are largest for teacher and organisational outcomes, so practice quality and teacher-level outcomes such as retention make better scoring targets.

Most new school leaders start underprepared, and induction is rarely regulated

About half of principals in OECD and other richer countries start without specific preparation. According to South Africa's TALIS 2018 country note, 54 per cent of principals across the OECD had completed school administration or principal training before appointment (DBE, 2019). UNESCO's 2024/5 summary report says "almost half of principals in richer countries do not receive any training before appointment" (UNESCO, 2024).

Bush observes that a teaching qualification and teaching experience "are often considered to be sufficient attributes for new principals" (Bush, 2018). Only 31 per cent of countries regulate the induction of new principals (UNESCO, 2024). Across UNESCO's profiles of 211 education systems, law or policy mentions pre-service training in 60 per cent of countries (D'Addio & April, 2024).

Programme content is uneven too. Of 142 preparation programmes analysed for UNESCO, "Only 18% of all programmes covered all four dimensions" of leadership (D'Addio & April, 2024). A quarter of principals in richer countries lack adequate training in data use and financial management, among other areas (UNESCO, 2024).

US syntheses agree on programme features without isolating any of them. Darling-Hammond and colleagues reviewed 104 US studies and highlight authentic applied learning, expert mentoring, cohorts and district partnerships (Darling-Hammond et al., 2022). Only 46 per cent of US principals reported an internship that gave them real leadership responsibility.

In California, teachers' odds of staying were 89 per cent under principals with high-quality preparation and 78 per cent under low-quality preparation (Darling-Hammond et al., 2022). The authors state that few studies have adequate controls or comparison groups. They add that no single study can establish definitive causal claims.

The strongest quasi-experimental result concerns whole principal pipelines in six US districts (Gates et al., 2019). After three or more years, schools with newly placed pipeline principals were 6.22 percentile points ahead in reading and 2.87 in maths. RAND found "little evidence that individual components were uniquely correlated" with the effects.

Problem-based learning and simulation are the methods closest to Synthetic School, and their evidence is weakest. A review of 73 sources on problem-based learning for school leaders judged the research "immature" (Hallinger & Bridges, 2017). A 2025 scoping review of 41 articles on digital simulations and serious games found that "weak research designs and methods limited the strength of evidence" (Hallinger et al., 2025).

In the UK, "most headteachers do not feel fully prepared when they first take on the role" (Greany et al., 2025). England's NPQH "prioritises domain-specific knowledge", while Scotland's Into Headship "seeks to balance leadership knowledge with wider aspects, including self-awareness and professional values". Since August 2020, teachers taking a first permanent headship in a Scottish local authority or grant-aided school must hold the GTCS Standard for Headship, awarded on completion of Into Headship (Scottish Government, 2019).

In England, 63 per cent of secondary school leaders under 50 first appointed in 2015 were still school leaders five years later (DfE, 2022). School leaders who moved into trust executive roles count as not retained, which may depress that figure.

A 2024 report by Big Education and the Centre for Education and Youth, as reported by Schools Week, criticised the leadership NPQs (Whittaker, 2024). It said they lack the "full extent of skills and knowledge" that increasingly complex senior roles require. This review found no evaluation of NPQH or Into Headship against student outcomes.

New Zealand gives the clearest readiness figures. The Education Review Office surveyed 596 new principals, and its summary report finds only 27 per cent felt prepared or very prepared on starting (ERO, 2023). Those who had coaching or mentoring, or a postgraduate leadership qualification, were twice as likely to feel prepared.

Induction was patchy: 15 per cent of boards with a new principal gave none, and 42 per cent did not know if one happened (ERO, 2023). Confidence rose from 58 per cent in the first two years to 81 per cent by years four and five. New principals felt least prepared for administrative and legal work and for partnership with Māori.

In the US, 11 per cent of public school principals in 2020-21 had left the principalship a year later (NCES, 2023). An Australian Catholic University news release covers a 2025 survey of 2,000 Australian school leaders (ACU, 2026). It reports a 53.9-hour average week, with 54.4 per cent seriously considering leaving.

Some Asian systems require preparation before appointment. Singapore's six-month, full-time Leaders in Education Programme prepares educators selected by the ministry for school leadership before their appointment (NIE, 2016). According to NCEE, every Singapore principal also has a coach and one or more mentors (Keo, c. 2016).

Hong Kong has required aspiring principals of public-sector schools to hold a Certification for Principalship since 2004/05 (Education Bureau, 2024). Acting appointees without it must obtain it within two years.

International school leaders face governance pressure. Morrison's survey of 177 school leaders, with 10 interviews, found the most significant challenges were governance and board relations, finance, staffing and organisational politics, and that average tenure had risen to 4.8 years (Morrison, 2024). Tes reports that a 2025 Morrison publication found governance the most common reason school leaders left a previous post (Seith, 2025).

New school leaders describe a sudden jump to sole responsibility. In a 2013 working-paper draft, Spillane and Lee interviewed 17 new Chicago elementary principals, and all but two described that jump (Spillane & Lee, 2013). Three months in, the volume and the unpredictability of problems were each salient for 59 per cent, and 47 per cent reported early loneliness.

Budgets and building maintenance generated much of that load and pulled attention from instructional leadership (Spillane & Lee, 2013). Principals placed in unfamiliar schools reported more problems than those promoted from within. A 2025 US study, pairing a national survey with interviews of 63 principals, found that learning about school climate and strategic management was "less common and often informal"; interviewees described relying on on-the-job learning and trial and error (Grissom et al., 2026).

This evidence supports Synthetic School's aim of rehearsing school leadership in a whole school. The problems novices name, including volume, unpredictability, isolation and operational work, are problems a simulated school can generate on demand. A synthetic school is also unfamiliar by construction, which matches the harder entry route Spillane and Lee describe.

The pipeline and New Zealand findings point to placing the simulation alongside mentoring and coaching, with coach-led debriefs. The same evidence challenges any claim that Synthetic School will raise students' outcomes. Both Hallinger-led reviews call for more rigorous research designs, which the platform could help supply.

Simulation beats no training by a wide margin and other teaching by less

Pooled effects shrink against active comparisons

Health professions supply the largest evidence base. Cook and colleagues pooled 609 studies with 35,226 trainees and found large gains against no intervention, 1.20 for knowledge and 1.09 for process skills (Cook et al., 2011). Effects fell to 0.81 for behaviour with real patients and 0.50 for direct effects on patients.

Those headline figures carry bias. Trim-and-fill adjustment cut the knowledge effect from 1.20 to 0.86, and two-thirds of studies were single-group pre-post designs (Cook et al., 2011). The authors questioned the need for more studies comparing simulation with nothing and called for comparisons between simulation designs.

Against other instruction the advantage narrows. Cook's own summary of a companion meta-analysis calls simulation "non-inferior to other approaches", with "a small but statistically significant benefit" for knowledge and skills (Cook, 2013). In emergency medicine, effects against other instruction were 0.33 or less and not significant (Ilgen et al., 2013).

A 2023 meta-analysis of 59 randomised trials against didactic teaching found a larger standardised mean difference of 0.80 (Mitchell & Ivimey-Cook, 2023). It analysed only immediate post-tests and reported no separate estimate for patient outcomes. Its authors report "considerable uncertainty" throughout.

Outside medicine the evidence thins. Chernikova and colleagues found g = 0.85 for complex skills across 145 higher-education studies, of which 126 were medical and 7 from teacher education (Chernikova et al., 2020). Communication (0.44) and teamwork (0.50) gained least, and document-based simulations had the smallest effect of any type, g = 0.31.

Simulation-game evidence is weaker still. Sitzmann's meta-analysis of 65 samples found 11 to 14 per cent higher knowledge than comparison groups, with "strong evidence of publication bias" (Sitzmann, 2011; Sitzmann, 2010). This review read it only through a reproduced abstract and her own practitioner summary.

Aviation shows positive but uneven transfer. A US Army Research Institute report states that "Flight simulators nearly always produce some degree of positive transfer" (Stewart et al., 2008). Its summary of a classic 1992 meta-analysis reports negligible differences between simulator-plus-aircraft and aircraft-only helicopter training, and no positive transfer for helicopters "primarily because there were too few helicopter studies" that met the inclusion criteria.

In military ground-combat training, live training transferred better than either a desktop or a wearable simulator (Taylor & Barnett, 2013). The wearable interface, though physically more realistic, was harder to use and no more effective.

Debriefs, feedback and spaced practice improve outcomes

The evidence for debriefs is robust. Tannenbaum and Cerasoli pooled 46 samples and found that debriefs improved performance by about 25 per cent (d = 0.67) (Tannenbaum & Cerasoli, 2013). Alignment between debrief and measure mattered: team-focused debriefs measured at team level reached d = 1.20.

A later meta-analysis of 61 studies, read through a reproduced abstract, found an overall d of 0.79 for after-action reviews (Keiser & Arthur, 2021). Aligned reviews and objective review media "consistently contributed" to effectiveness. A practitioner research summary drawing mainly on the same meta-analysis says reviews are "particularly effective when the review includes objective records of the event" (Paul, 2021).

The medium matters less than the content. Across four health simulation studies, video-assisted debriefing added a negligible 0.10 on time-based skills over debriefing without video (Cheng et al., 2014). Multimedia aids made little difference in Tannenbaum and Cerasoli's data (0.66 against 0.62).

Facilitated debriefs scored 0.75 against 0.25 without a facilitator, but only two unfacilitated samples existed (Tannenbaum & Cerasoli, 2013). No study reviewed here compares AI-led with human-led debriefs. A 2024 review of technology in after-action reviews found "little extant guidance" on its best use (Keiser, 2024).

Earlier syntheses rank feedback first. Issenberg and colleagues' 2005 review, read through the authors' 2007 summary, called feedback "the single most important feature" and judged the research "generally weak" (Issenberg & Scalese, 2007). Cook's summary of a 289-study meta-analysis lists feedback, repetition, range of difficulty and distributed practice among features that improved skills (Cook, 2013).

Within the main health meta-analysis, design features showed "no consistent statistically significant interactions" with outcomes (Cook et al., 2011). A 2025 narrative review found that most design features had only one to four studies measuring real clinical behaviour or outcomes; simulator modality, with 11, was the exception (Low et al., 2025). Error management training, distributed practice, dyad training and in situ training showed improved outcomes there.

Scaffolding interacts with experience. In Chernikova and colleagues' data, reflection phases helped postgraduate and in-service learners (g = 1.10 with, 0.86 without) and did not help undergraduate and graduate students (0.52 with, 0.81 without) (Chernikova et al., 2020). Prompts helped in unfamiliar contexts but lowered effects in familiar ones, from 0.96 without prompts to 0.33 with them.

Deliberate practice is central in theory and modest in measured effect. Ericsson's conditions include a well-defined goal, motivation, feedback and repeated chances to refine performance (Ericsson, 2008). A publisher's press release on a meta-analysis of 88 studies reports that deliberate practice explained under 1 per cent of performance variance in professions (Macnamara et al., 2014).

Functional fidelity outweighs physical realism

Fidelity reviews find physical resemblance adds little once the task's demands are represented. Across 24 health studies, high- and low-fidelity simulators differed by 1 to 2 per cent on average (Norman et al., 2012). Hamstra and colleagues recommend replacing "fidelity" with physical resemblance and functional task alignment (Hamstra et al., 2014).

Cook's summary states that "the appearance of the simulator is less important than a proper representation of the key steps in the task" (Cook, 2013). Chernikova and colleagues coded authenticity and found high authenticity linked to larger effects (g = 0.86) than low (0.58) (Chernikova et al., 2020). Role play without technology performed about as well as screen-based simulation.

Their 2024 extension of 214 studies found that the salience of relevant information explained variation in effectiveness (Chernikova et al., 2024). The trial-only health meta-analysis found larger effects with high-fidelity mannequins or simulation centres than with virtual reality (0.90 against 0.70, with overlapping intervals) (Mitchell & Ivimey-Cook, 2023). Such between-study comparisons cannot rule out confounding by cost or training intensity.

Aviation adds a caution about gains inside a simulator. In a meta-analysis of whole-body motion, true transfer designs gave d = 0.10, and expert pilots gained nothing (d = -0.01) (de Winter et al., 2012). Features that improve performance inside a simulator need not improve it outside.

Transfer weakens down the outcome chain

Transfer to real practice is documented and smaller. Against no intervention, health effects fall from about 1.1 to 1.2 for simulated skills to about 0.8 for behaviour and 0.5 for patient outcomes (Cook et al., 2011). The 2023 trial-only meta-analysis calls for "further high-quality studies" of effects on patient and public health outcomes (Mitchell & Ivimey-Cook, 2023).

In the NCSBN National Simulation Study in the US, ten prelicensure nursing programmes randomised 847 students (Hayden et al., 2014). Replacing up to half of clinical hours with simulation made no significant difference to end-of-programme clinical competency, comprehensive nursing knowledge or NCLEX pass rates, or to managers' ratings up to six months into practice. Trained study team members led a structured debrief after every scenario, using the Debriefing for Meaningful Learning method.

Error management training shows a pattern Synthetic School can design around. In a meta-analysis of 24 studies, a post-review preprint reports d = 0.80 for adaptive transfer (11 studies), while performance during training was slightly worse (d = -0.15, 4 studies, not significant) (Keith & Frese, 2008). None of those studies measured performance on the job.

The Army report finds that "True negative habit transfer is rare" in flight simulation (Stewart et al., 2008). A 2022 FAA study of 40 airline pilots compared stall training frozen at the moment of response with training that flew on into the stall (Landman et al., 2022). Its authors flag "a potential hazard of too much contrast between self-paced stall events in the simulator and externally-paced surprising events".

Replays and counterfactual reflection rest on adjacent evidence

No study reviewed here tested replaying a scenario with different choices at a save point. The nearest evidence is rapid cycle deliberate practice, which pauses a resuscitation scenario for coaching and then rewinds it. A scoping review of 23 studies found learners liked it, but concluded: "This review has not found that RCDP is superior to traditional simulation" (Ng et al., 2021).

In that review, checklist scores dropped more in the rewind groups by three to four months (Ng et al., 2021). The International Liaison Committee on Resuscitation found no clinical outcomes reported in eight studies and issued a "weak recommendation, very low quality of evidence" (Abelairas-Gómez et al., 2023).

Replay helped in the simulation-game meta-analysis. According to a reproduction of Sitzmann's abstract, games taught more when trainees could replay them as often as they wanted (Sitzmann, 2011).

Reviewing successes as well as failures improved later performance. On an Israel Defense Forces navigation course, soldiers debriefed daily on both improved more than those debriefed on failures alone (Ellis & Davidi, 2005). The study had no condition without any review.

Counterfactual thinking drives learning from close calls. Morris and Moore found that pilots learned from dangerous incidents by imagining how their own actions could have produced a better outcome (Morris & Moore, 2000). Accountability to organisational superiors suppressed that thinking in an incident-report study and in a flight-simulator experiment.

What this means for the design

Claims for Synthetic School should be pitched at the smaller effects found against active teaching and for non-procedural skills. Its closest analogues in the cross-domain meta-analysis are role play (g = 0.63) and communication or teamwork outcomes (0.44 to 0.50) (Chernikova et al., 2020).

The weekly cycle gives distributed practice, which improved outcomes in several health reviews. The evidence for debriefs is robust, and the decision log can serve as the objective record that strengthens it. Debrief scaffolding should adapt to experience, with worked examples for novices and reflection phases for experienced leaders.

Branch replays should prompt private, self-focused reflection on what the trainee could have done differently, because accountability suppresses it. Learning should be judged on new scenarios, since error management training pays off in adaptive transfer and replays let trainees anticipate events. Replays should not be assumed to beat a well-run debrief, given the mixed rewind evidence.

The same evidence challenges two assumptions. Weekly save points make every decision self-paced, which the stall study flags as a hazard when real events arrive unannounced. Some events should therefore land mid-week without a pause, and any AI-led debrief needs its own test against human facilitation.

School simulations date from 1958 and rarely test transfer to real schools

The Whitman simulation began as a research and selection instrument

Simulation entered school leadership as a research and selection tool. According to a 1989 Korean review, 232 elementary principals each spent a week in 1958-59 running a simulated school called Whitman (Shin, 1989). The review places Whitman within a US project to define criteria of success in school administration and improve principal selection.

ETS published the technical report in 1961, indexed under in-basket tests, and Columbia University published the book in 1962 (Hemphill et al., 1961; Hemphill et al., 1962). The report's abstract frames the study as a response to demand for better ways to educate and select principals and to improve school administration. The Korean review itself shows the tradition being taken up in South Korea by the late 1980s.

UCEA spread simulated materials into US administrator training, and the Korean review calls it the most important single force in that spread (Shin, 1989). UCEA's own history records that nine of nineteen development teams failed to produce their "Monroe City" simulations, while the other ten succeeded (Culbertson, 1995).

Selection research moved to principal assessment centres, and a 1992 meta-analysis of 34 NASSP Assessment Center validity studies found criterion-related validity "very low" (Pantili, 1992). Its twelve skill dimensions were hard to tell apart. One district centre in Dade County, Miami, did predict later job-performance ratings for 121 people (Gomez & Stephenson, 1987).

Problem-based learning carried the teaching tradition forward from the late 1980s (Bridges & Hallinger, 1992; Hallinger & Bridges, 2017). Hallinger's change-management simulations were built or adapted for Thai and Chinese school leaders (APCLC, n.d.; Tang, 2015). The Chinese adaptation followed an eight-stage research-and-development cycle, and Tang reports that Chinese school leaders face different contextual constraints, such as China's social structure, government policy, cultural norms, educational theories and teaching principles (Tang, 2015).

Coaching between mixed-reality sessions beat self-reflection in five trials

The strongest causal evidence comes from five randomised trials at one university in the south-eastern US (Cohen et al., 2021). Between mixed-reality sessions, in which trained actors voiced student avatars, candidates received five minutes of directive coaching or a self-reflection protocol. The working paper reports an average effect of 1.44 SD on blinded ratings of teaching quality.

Those gains were measured inside the simulator. The authors say more evidence is needed that better teaching in simulated classrooms corresponds to outcomes with real children (Cohen et al., 2021). An earlier trial with 105 prospective teachers used the simulator as both practice space and assessment platform (Cohen et al., 2020).

A 2026 scoping review of 82 TeachLivE and Mursion articles quotes earlier work warning that "one cannot claim that skills developed in MRS generalize to real classrooms" (Mikeska and colleagues, 2019, as cited in Anderson & Calandra, 2026). It found that 86 per cent of 72 research and practitioner pieces had US-based authors. Researchers it reviews called for better measures of whether gains generalise, including "moving beyond reliance on teacher self-reports".

A Dutch team's systematic review of 15 studies found classroom simulations mostly improved classroom management, with technical problems the main hindrance (Theelen et al., 2019). Positive experiences depended on how realistic and authentic the simulation felt.

At Syracuse University, Dotger's clinical simulations adapt medicine's standardised patients, using trained actors who follow exact verbal triggers (Dotger, 2015). The article cites a study of 141 candidates whose high self-ratings on work with families exceeded their assessed ability. Dotger presents simulations as "limited, bounded engagements" that "intentionally reduce the complexity of everyday school environments", and treats this focus on a single, discrete encounter as a design strength.

Simulations built for school leaders have thinner evidence. Virginia Commonwealth University's Project ALL runs "a year in the life of a school principal" through more than 400 story elements (Mann & Shakeshaft, 2013). Its developers claim more potent learning than discussion methods, but the abstract reports no outcome data (Mann et al., 2011).

SchoolSims branching simulations ran across recruitment, preparation, induction and in-service phases in one US university-district partnership (Azukas et al., 2025). Twenty-seven interns completed 15 simulations each, and some repeated scenarios voluntarily. The authors list three limits: fixed choices, no customisation to district policy or state law, and no extended practice in dialogue. The evidence comes from participants' written reflections and focus groups, and the authors note that "social desirability or peer influence" may have shaped some responses.

In a UCLA dissertation, 32 aspiring principals from three Southern California preparation programmes did either mixed-reality simulation or a non-simulation reciprocal learning partnership protocol (16 each), and self-awareness rose in both groups on two validated surveys (Barnett, 2025). An AERA 2026 paper on mixed-reality simulation for emerging school leaders reports gains in self-awareness on validated surveys and gives no sample size (AERA, 2026). A study from Bar-Ilan University used 50 simulation sessions with teachers in leadership roles to surface ethical tensions in principals' work (Shapira-Lishchinsky, 2015).

LLM role-play has spread faster than its evidence

Since 2023, large language models (LLMs) have played students and parents, among other roles, in training tools. Stanford's GPTeach let novice teaching assistants practise with GPT-simulated students and was later used with over 800 novice teachers (Markel et al., 2023). Its preprint abstract reports no figures.

Evelyn AI, from Florida State University and the University of Hawai'i at Mānoa, aims to replace the human actors used in systems such as TeachLivE (Bhowmik et al., n.d.). Raters' judgements matched the agent's intended prior-knowledge level only moderately (r = 0.47). The paper reports no learning outcomes for teachers.

The one measured learning gain found here comes from a small pilot. In a 2026 preprint with 15 prospective physics teachers, diagnostic scores on written vignettes rose significantly from pre- to post-test among the 11 participants with paired data (Wilcoxon p = 0.014) after work with a simulated class of five AI students (Tufino, 2026). Their transcripts used the target vocabulary in only 2 of 71 substantive turns, a gap between knowing and doing.

AI-driven simulations aimed at school leaders remain at pilot stage. At the University of Melbourne, aspiring school leaders hold conversations in virtual reality with AI-supported avatars representing teachers, parents and other stakeholders; the item does not say how the avatars' responses are produced, and feedback so far is "preliminary" (Cornes, 2026). Take-up rose when the activity was scheduled into class time.

Two European projects use AI parents to train student teachers. The University of Bergen's PRACTICE project in Norway has published no findings yet (SLATE, n.d.). Zurich's PHZH pairs an AI parent with an AI coach who leads a structured debrief, and an initial pilot reported higher confidence (Dani, 2026).

Methodological papers name the threats to these tools. Yuan and colleagues describe a "competence paradox", in which capable models playing partially knowledgeable learners make unrealistic errors (Yuan et al., 2026). Gonnermann-Müller and colleagues found that the observed behaviour of high- and moderate-intensity ADHD student personas drifted over unscripted dialogue, and that scripted interactions with recurring task-relevant prompts cut that drift by up to 97 per cent (Gonnermann-Müller et al., 2026).

Representation is a documented risk. A US study of eight preservice teachers found that a mixed-reality platform's simulations might have perpetuated overtly racist and sexist stereotypes, with negatively stereotyped avatars afforded lower-quality opportunities to participate (Bondurant & Reinholz, 2023). The 2026 scoping review reports a participant calling the avatars "five extreme stereotypes" and warns that generative-AI automation raises concerns about bias (Anderson & Calandra, 2026).

This review found no randomised trial of LLM-based simulation measuring real teaching or leadership outcomes. It found no AI-driven whole-school simulation lasting a year, and no simulation built around one school system's own procedures. No study documents trainees gaming a teacher or leader simulation, although any scored simulation invites that risk.

What this means for the design

Whitman and Project ALL show long-standing demand for whole-school simulation, and Monroe City shows the cost of authoring one by hand. LLM agents remove the live human actor, the cost bottleneck named in mixed-reality research. Generated content could also answer the SchoolSims limits of fixed choices and generic policy, though that remains an untested design hypothesis.

The record also sets requirements. Scores used for selection or certification need their own validity evidence, given the assessment-centre findings. Personas need audits for stereotyping, and logs of what trainees do offer a more honest assessment source than their self-ratings.

Cohen's trials suggest the coaching loop between sessions matters more than realism, so the debrief is where Synthetic School should concentrate design effort and evaluation. The Chinese adaptation of Hallinger's simulation supports building around one system's procedures before transferring to another.

Language-model agents simulate aggregates better than individuals or effects

Fidelity depends on grounding data that invented people lack

The best individual-level fidelity comes from agents grounded in rich data about real people. Park and colleagues built agents from two-hour interviews with 1,052 US adults (Park et al., 2026). In the current preprint version, those agents predicted General Social Survey answers at 83 per cent of each person's own two-week consistency.

Agents given only demographics reached 74 per cent (Park et al., 2026). The first version of the preprint reported 85 per cent for interview agents under a different title (Park et al., 2024). In the current version, grounding gave no clear benefit over demographic prompting in incentivised economic games.

The best-known generative-agent study, the 25-agent Smallville town, measured believability as judged by 100 human raters (Park et al., 2023). Believability is a separate property from accuracy against real people.

Aggregate patterns are easier to reproduce. GPT-3 conditioned on the backstories of real US election-survey respondents reproduced vote choice with correlations of 0.90 to 0.94, though pure independents were poorly matched, especially in 2020 (Argyle et al., 2023). In Park's study, agents' experimental effect sizes ran larger than humans' in a five-study analysis the authors call underpowered (Park et al., 2026).

A 2024 preprint, described on a Stanford project page, found GPT-4's predictions of 476 US survey-experiment effects correlated 0.85 with actual effects (Hewitt et al., 2024). The 2026 Nature version, read here through the Stanford GSB listing, reports that predictions systematically overestimated effect sizes (Ashokkumar et al., 2026). Correlations fell for an archive of 15 megastudies, to levels comparable with pooled expert forecasters.

A 2025 benchmark across 20 datasets scored the best models 40.80 out of 100 for simulation fidelity (Hu et al., 2025). Instruction tuning helped on questions where people agree and hurt where opinion is diverse. Models particularly struggled when simulating specific demographic groups.

Synthetic School's staff and families are invented, so no interview or survey can ground them. The relevant benchmark is therefore the demographic or persona agent, at 71 to 74 per cent on survey items. Two Stanford-led studies also report inflated LLM effect sizes, which argues against letting agents decide whether an intervention worked.

Documented failures cluster where a school needs accuracy

LLMs flatten and misportray identity groups. Across four models and 3,200 human participants, responses for non-binary people and people with impaired vision sat closer to outsiders' imitations than to in-group answers (Wang et al., 2025). GPT-3.5 and GPT-4 tended to use only three of the five Likert options across 100 responses per scenario.

Opinion alignment is uneven across groups. Santurkar and colleagues found "substantial misalignment" between models and 60 US demographic groups, which persisted when models were steered toward a group (Santurkar et al., 2023). People aged 65 and over and widowed people were poorly represented.

Synthetic samples compress variance. In US election-survey data, Bisbee and colleagues found synthetic group ratings with means near the human averages but far less spread (Bisbee et al., 2024). About 48 per cent of regression coefficients differed significantly from the human estimates.

The same prompt gave different distributions in April and July 2023, which Bisbee and colleagues link to undisclosed model changes (Bisbee et al., 2024). Five models set to be deterministic varied in accuracy by up to 15 per cent across ten runs (Atil et al., 2025). Meaning-preserving changes to prompt format moved one model's accuracy by up to 76 points (Sclar et al., 2024).

Education studies show the same pattern. EduAgent found that LLMs answer quiz items correctly with ease but struggle to reproduce students' specific wrong choices (Xu et al., 2024). Quiz-accuracy similarity with 310 real students reached only about 0.66 to 0.68.

Models also defer to users. Sharma and colleagues found five assistants sycophantic, and Claude 1.3 wrongly admitted a mistake on 98 per cent of challenged questions (Sharma et al., 2024). In Smallville, instruction tuning seemed to make agents overly formal and overly cooperative; one agent "rarely said no" to others' suggestions (Park et al., 2023).

Consistency decays over long runs. Instructed agents showed "a significant instruction drift within eight rounds of conversations" (Li et al., 2024). Smallville agents also embellished events and broke world rules, such as entering shops after closing time (Park et al., 2023).

No institutional simulation has been checked against real outcomes

Multi-agent schools and hospitals now exist. AI-Agent School ran ten teacher agents and 40 student agents for five simulated days and scored them against expert-refined references (Jin et al., 2025). Its reference data were largely generated by a language model and then edited by experts; the paper says only that the classroom teaching action data "came from real classrooms".

SimClass compared its LLM classrooms with human-taught ones using the Flanders interaction analysis system (Zhang et al., 2025). Its student-initiation ratios were about 0.9, against about 0.1 for expert and novice human teachers. On that measure the LLM classrooms behaved unlike human ones.

Agent Hospital evaluated its doctor agents on a medical exam benchmark and reported no expert judgement of the hospital's realism (Li et al., 2025). Its reported accuracy also changed between versions. Version 1's abstract gave 93.06 per cent with GPT-4 on a 72-question respiratory subset of MedQA; version 3 reports 92.22 per cent with GPT-4o on the whole MedQA set (96.15 per cent with o1-preview) and gives no figure in its abstract.

A critical review of 35 generative social simulations found subjective validation, such as believability judgements, still dominant (Larooij & Törnberg, 2026). Most studies relied on a single simulation run, and robustness checks and sensitivity analyses were widely missing. The authors conclude that LLMs "may exacerbate rather than alleviate the challenge of validating ABMs".

Hybrid designs have strong arguments and no head-to-head test

LLMs are unreliable at tracking state. On a text-world benchmark, GPT-4 predicted the full next state correctly in at most 59.9 per cent of transitions that changed it, its best result, obtained with human-written rules. Its best result for changes the environment drives was 49.7 per cent (Wang et al., 2024). Errors concentrated on properties needing arithmetic, common-sense or scientific knowledge.

Leading frameworks already keep hard state in code. Concordia's Game Master tracks grounded variables such as money and votes "in Python" and checks that agents' actions on them are valid (Vezhnevets et al., 2023). Its authors warn that concurrent agent actions "often leads to inconsistencies".

Ghaffarzadegan and colleagues couple a mechanistic model of interaction with an LLM that makes individual decisions (Ghaffarzadegan et al., 2024). A 2025 IEEE review advocates "hybrid ABM-GABM approaches and systematic validation" (Adornetto et al., 2025). The mixed subjects design treats LLM output as cheap, possibly inaccurate observations anchored to human data (Broska et al., 2025).

Validation guidance in this literature is consistent. Anthis and colleagues recommend testing on data a model has not seen in training, such as data past its training cutoff, and encourage publicly shared preregistrations of simulation predictions, with verification that they predate data collection (Anthis et al., 2025). Larooij and Törnberg ask for "operational validity", meaning accuracy for the model's intended purpose (Larooij & Törnberg, 2026).

This review found no trial comparing an engine-decides design with an LLM-decides design. It also found no study measuring how faithfully LLM narration matches state held outside the model.

What this means for the design

This literature gives Synthetic School its strongest citable rationale. Once the engine decides outcomes, compressed variance, inflated effects, sycophancy, drift and non-determinism fall on a layer that cannot change them. A seeded engine can be replayed exactly, and only the narrative surface varies.

The hybrid moves the validity burden without removing it. The engine needs the calibration and replication standards set out in the next section. Narration needs checks against engine state and audits by identity group, including students with additional support needs.

Sycophantic staff could still flatter a trainee's choices even when the underlying state is correct. Believability testing with staff should count as usability evidence, since Larooij and Törnberg call for operational validity instead.

Education policy models stay suggestive, and few report out-of-sample tests

Education agent-based models keep finding perverse system effects

Education agent-based models (ABMs) form a small literature with a recurring finding. 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 conclude that a rule one might expect to aid district improvement "can also limit district-level performance in certain conditions".

The 2014 extension, which draws incoming students from 17,131 Chicago ninth graders (2000 to 2003), found that over a range of conditions a stronger emphasis on achievement could lower district mean achievement (Maroulis et al., 2014). The authors stress that all their results are contingent on new schools having a higher value-added than incumbent schools, and that limitations must be addressed before such models support "more precise statements about school district planning". They report no formal validation against observed district outcomes.

School-choice models find that choice can raise segregation. A US conjoint study with simulations found that "unmitigated school-choice policies increase segregation", even when parents ignore race (Ukanwa et al., 2022). A model of tracking and accountability found that the two policies together mostly intensify school socio-economic segregation (Pensiero & Brede, 2024).

In the Netherlands, Dignum and colleagues estimated a choice model on 2019 household register data for Amsterdam and Almere and found distance the main driver (Dignum et al., 2025). Group-level and school-level statistics gave "very different results", and the authors judge that "the actual empirical estimates are more difficult to trust".

In England, Millington and colleagues compared model regressions with data from Barking and Dagenham but reported no formal statistical validation (Millington et al., 2014). Parents with above-average but not very high aspiration most often missed their top-ranked school.

The closest precedent for a policy lab is a Washington State model of teacher sorting, reported in a CALDER working paper (Goldhaber et al., 2022). It drew Bayesian-calibrated probabilities from 564,296 teacher-year records. With all inequity pathways switched on, it reproduced observed exposure to novice teachers within 90 per cent intervals.

In that model, equitable hiring brought the gap for students of colour close to zero by about year eight (Goldhaber et al., 2022). The Brookings commentary says the model covers "hundreds of different schools" and calls it "far from the final word" (Kasman, 2023).

In England, NFER's 2026 teacher supply model uses a synthetic population calibrated to the 2023 School Workforce Census (Worth et al., 2026). NFER says its cost estimates "should be treated as indicative". Its test runs led it to conclude that at least 500 iterations were needed for good confidence in results.

In France, Manzo calibrated a model of educational choices to a 2003 national intergenerational table and found network effects necessary to reproduce stratification (Manzo, 2013). Manzo calls it the study's "main limitation" that the parameter values of each micro-level mechanism "only come from arbitrarily chosen numerical values", although the model's aggregate outputs were matched to survey data.

No ABM reviewed here treats attendance as the policy outcome. The closest is a 2025 conference abstract that used group model building and a simulation to plan a school pilot on chronic absence among children with asthma, with no simulation or intervention results reported yet (Masotya et al., 2025). Pandemic models also simulated schools, such as the UK reopening scenarios run in Covasim under ten random seeds each (Panovska-Griffiths et al., 2020).

Microsimulation has modelled education for decades, mostly qualifications and finance. A survey of more than 60 dynamic models names education roles, mostly education finance, student support and higher-education policy, in models from Australia, Canada, France, Germany, Sweden and the UK (Li & O'Donoghue, 2013). New Zealand's MELC simulates a synthetic birth cohort of 10,000 children to age 13 and was validated against national reading norms (Milne et al., 2015).

MELC's authors caution that "it is unclear whether the effects tested in the model can be considered causal" (Milne et al., 2015). A 2025 review of 20 studies found education digital twins in three main areas: engineering and vocational training, campus facility management and individual learners (Prasetyo et al., 2025). This review found no implemented twin of a school system.

Validation standards exist, and education models seldom meet them

Simulation science has a mature validation toolkit. Pattern-oriented modelling, proposed as a general strategy for designing, testing and analysing agent-based models (Grimm et al., 2005), bases a model's design, evaluation and parameterisation on several patterns observed in the real system (Grimm et al., 2020). The 2020 ODD update asks modellers to link their code and to report patterns the model fails to reproduce, since reporting only successes "would resemble 'HARKing'".

Sargent separates verification, which checks the program, from validation, which checks accuracy for a stated purpose (Sargent, 2005). His historical-data validation builds a model on part of the data and tests it on the rest. He warns that "A model may be valid for one set of experimental conditions and invalid in another."

Docking compares independent implementations of the same model. Axtell and colleagues defined distributional equivalence and traced their main mismatch to the order in which agents act (Axtell et al., 1996). Wilensky and Rand's replication also traced differences to event and activation order (Wilensky & Rand, 2007).

Reporting lags behind these standards. The STRESS authors cite a review of 114 simulation studies of mobile ad hoc networks in which "no studies detailed the pseudo random number generator" (Kurkowski et al., 2005, cited in Monks et al., 2019). Most education models reviewed here calibrate to aggregates, and few report out-of-sample tests or independent replication.

Calibration to published statistics carries an identification risk. Windrum and colleagues warn that an over-parameterised model "can generate any result" (Windrum et al., 2007). Microsimulation aligns to external totals almost universally, and Li and O'Donoghue warn that alignment can hide behavioural or structural change (Li & O'Donoghue, 2013).

Shared random draws cut noise only while the streams stay matched

Branches that share every random draw apply common random numbers (CRN), the standard variance-reduction method for comparing alternatives. CRN shrinks the variance of a difference when outputs across scenarios correlate positively (Chen, 2001). Chen also warns that CRN may backfire in procedures that select the best of three or more alternatives.

In ABMs, a single global generator breaks the match. If one branch draws a number the other does not, "all subsequent stochastic realizations could be different" (Klein et al., 2024). Klein and colleagues assign an independent stream to every decision and report up to four orders of magnitude fewer simulations needed.

Generators with many independent streams and substreams exist for this purpose (L'Ecuyer et al., 2002). Travel-demand microsimulation reports up to 99 per cent fewer runs, roughly a hundredfold saving, from reusing draws across scenarios (Zhou et al., 2023). None of the education models reviewed here reports using CRN.

For models that use common random numbers, STRESS asks authors to describe "how streams or seeds are distributed across components within the model" (Monks et al., 2019). Lee and colleagues warn that "A large sample size can easily classify a minute difference as being significant", and note that many therefore argue for more attention to effect sizes (Lee et al., 2015).

Documented failures argue for possibilistic claims

Critiques of model-based policy advice are specific. Edmonds and ní Aodha list ten pitfalls, from institutionalised assumptions to downplayed uncertainty, and judge true prediction "infeasible for most social phenomena" (Edmonds & ní Aodha, 2019). They recommend mapping the plausible ways a policy could fail.

Saltelli and 21 co-authors warn that "Excessive regard for producing numbers" can push a discipline toward being "precisely wrong" (Saltelli et al., 2020). They add that thought experiments "should not be treated as predictions".

Model outputs can also drop out of debate. A re-run of Imperial College's CovidSim confirmed a "little-reported forecast" about UK school closures (Rice et al., 2020). Combined with isolation of younger people, closures were predicted to increase the final number of deaths, albeit postponed to second and subsequent waves.

What this means for the design

The policy lab has a clear precedent in the Washington State teacher model, which spans hundreds of simulated schools. The shared-draw design has a strong rationale in CRN, provided each stochastic decision draws from its own stream. An A-versus-A run, in which identical policies must give zero difference, would test that the streams stay matched.

Paired CRN intervals measure simulation noise and say nothing about uncertainty in behavioural parameters. The lab therefore needs sensitivity analysis over those parameters and a stated fit criterion for calibration. Results should be framed as conditional on stated assumptions, and unfavourable runs should be published.

Calibration to the two school systems' statistics should target several patterns at once, such as attendance distributions and deprivation gaps, and report the patterns the engine misses. The digital twin review defines a twin as a replica linked to its physical system through real-time data, a description Synthetic School does not meet.

Attendance trials cut a few days a year, and five points tops the range

Persistent absence rose everywhere after the pandemic

Most English-speaking systems draw the persistent-absence line at 10 per cent of sessions or days. In England, persistent absence was 18.14 per cent in 2024/25, against 10.86 per cent in 2018/19 (DfE, 2026a). Severe absence, missing half of sessions or more, rose from 0.85 to 2.39 per cent over the same period.

Students in England eligible for free school meals were "over two and a half times as likely" to be persistently absent in 2024/25 (DfE, 2026a). In autumn and spring 2025/26, students with an education, health and care plan were "over seven times as likely to be severe absentees" as students with no identified SEN (DfE, 2026b). Severe absence kept rising as persistent absence eased.

In Scotland, persistent absence was 28.5 per cent in 2024/25, down from 31.4 per cent the year before (Scottish Government, 2025). Between 2010/11 and 2020/21 it ranged from 19.1 to 21.8 per cent. Secondary persistent absence reached 37.4 per cent.

Attendance was 87.6 per cent in SIMD quintile 1 and 94.1 per cent in quintile 5 (Scottish Government, 2025). Students with additional support needs attended 88.2 per cent, against 92.9 per cent for other students. England and Scotland share the 10 per cent definition yet differ by about ten points in persistent absence.

Other systems show the same rise and partial recovery. US chronic absenteeism was 15 per cent in 2019 and 28 per cent in 2022, with 2023-24 estimates of 19 to 25 per cent (Malkus, 2024). In Australia, the share of full-time Year 1–10 students attending at least 90 per cent of Semester 1 fell from 73.1 per cent in 2019 to 62.1 per cent in 2025 (ACARA, n.d.).

New Zealand counts "regular attendance" as more than 90 per cent of a term and reserves "chronic absence" for attending 70 per cent or less (Ministry of Education, n.d.). Regular attendance was 64.3 per cent in Term 2 2026, against a government target of 80 per cent (DPMC, 2026).

The 90 per cent line is a reporting convention. Australia's education evidence body calls the 10 per cent threshold "a common though non-scientifically supported cut-off point" (Melvin et al., 2025). Western Australian research on more than 415,000 students concluded that "There is no 'safe' threshold" (Hancock et al., 2013).

Each missed day costs a little attainment

Quasi-experimental studies put the cost of a missed day at roughly 0.003 to 0.05 SD, on this review's conversion of their figures. In a Massachusetts NBER working paper, each weather-induced absence lowered maths achievement by 0.05 SD, while snow-day closures had no effect (Goodman, 2014). A US study in the Journal of Public Economics, using within-student comparisons, found that missing ten classes cut test scores by 3 to 4 per cent of an SD (Liu et al., 2021).

In Scotland, Klein and colleagues linked school and SQA records for two cohorts through the Scottish Longitudinal Study (Klein et al., 2022). Each extra percentage point of absence was associated with 0.03 SD lower attainment at S4. Attainment was measured in standardised UCAS tariff points.

Excused and unexcused absences "were not significantly differently associated" with attainment in Scotland (Klein et al., 2022). Western Australian data found unauthorised absence more strongly linked to achievement than authorised absence (Hancock et al., 2013). Whether the reason for absence matters therefore remains unsettled.

Long-run links are smaller and partly confounded. In the 1970 British Cohort Study, five days absent at age 10 were associated with a 0.66-point higher chance of having no qualifications by 42 (Dräger et al., 2024). The authors caution that the estimates may not be causal.

Raising attendance has not reliably raised test scores. In a Chicago trial reported in a working paper, Check & Connect mentoring cut absences in grades 5 to 7 but had no significant effect on grades or test scores (Guryan et al., 2020). Text alerts to US parents cut course failures by 27 per cent with no effect on state tests (Bergman & Chan, 2021).

Illness, attitude and the calendar drive absence

Illness is the largest recorded reason for absence in England and Scotland, and it remains above pre-pandemic levels in both (DfE, 2026b; Scottish Government, 2025). A meta-analysis covering 43 absenteeism studies found negative attitude to school the strongest correlate (r = 0.50), ahead of internalising problems and weak student-teacher relationships (Gubbels et al., 2019).

Parents' beliefs matter. In a pilot survey in one US district, parents of high-absence students estimated 9.6 absences against an actual 17.8 (Rogers & Feller, 2018). Focus-group research with parents, read here through Public First's published summary, reports term-time holidays are now "entirely socially acceptable across all socioeconomic groups" (Burtonshaw, 2023).

Kearney's functional model sorts school refusal into four functions, from avoiding distress at school to seeking rewards outside it (Gonzálvez et al., 2016). School-level frameworks use tiers of support, from universal promotion of attendance to intensive individual work (Kearney & Graczyk, 2014). Attendance Works places students missing 10 to 19.99 per cent in its tier for moderate chronic absence (Attendance Works, 2018).

Calendar patterns are documented in detail for England. Absence at the end of the week runs 1.5 points, or 20 per cent, above the rest of the week, and the unauthorised end-of-week effect is larger in deprived areas (Clifton-Sprigg & James, 2025). Weekly absence in 2024/25 peaked at 9.6 per cent in the final week of spring term, which coincided with Eid al-Fitr (DfE, 2025).

Five points matches only a non-randomised estimate

The engine's optimistic case cuts daily absence by five percentage points for supported students just under 90 per cent attendance. Over a 190-day year that is about 9.5 days per student, roughly a 45 per cent relative cut in absence. These conversions, and those in the table below, are this review's arithmetic on the published figures.

Randomised parent mailings in the US cut absence by up to about 6 to 8 per cent. The most effective mailing to 28,080 households of high-risk students cut absences by 1.1 days between the first mailing in October and the end of the school year, about 6 per cent (Rogers & Feller, 2018). In ten California districts, early-grade mailings cut 0.53 days, a 7.7 per cent reduction (Robinson et al., 2018).

Intensive mentoring does better with older students. In a Chicago randomised trial, Check & Connect cut absences by 4.22 days (22.9 per cent) in grades 5 to 7 (Guryan et al., 2020). It had close to no effect in grades 1 to 4 and cost about $1,700 per student a year.

Only New York City's Success Mentors match five points. In a non-randomised comparison, previously chronically absent students with mentors had attendance "5 percentage points higher on average", about nine more days a year (Balfanz & Byrnes, 2013). The peer-reviewed version found that mentees were not "any more likely to exit chronic absenteeism" (Balfanz & Byrnes, 2018).

Meta-analyses of controlled designs find small average effects. Eklund and colleagues found g = 0.25 for between-group designs and concluded that most attendance practices "are either understudied, lead to small effects, or both" (Eklund et al., 2022). In a rapid review of 72 studies, all but three from the US, the EEF found a "very small" effect of 0.07 for parental communication (seven studies pooled) and judged the overall quality of evidence "weak" (EEF, 2022).

A Campbell review of 28 studies from the US, Canada, the UK and Australia found interventions for chronic truants raised attendance by about 4.7 days (Maynard et al., 2012). Its randomised trials averaged g = 0.57, against 0.95 for single-group pre-post designs. Eklund's within-group designs similarly gave 1.04, against 0.25 between groups.

Points assume a 180-day US year, or 170 to 181 days for Check & Connect, and a 190-day year for the engine. Maynard's figure cannot be converted because programme durations vary.

Five points is about twice the best randomised mentoring result and 8 to 17 times the parent nudges. Its implied 45 per cent relative cut exceeds every randomised relative effect found, the largest being Check & Connect's 22.9 per cent. Five points is defensible only as an optimistic upper bound for intensive, well-run mentoring.

Students selected for a recent dip below 90 per cent can be expected to recover partly without help. The larger effects in uncontrolled designs fit that pattern, so the engine's counterfactual should include natural recovery. Effects also vary by age and by baseline absence, so one uniform effect for every supported student overstates precision.

On this review's reading, light-touch parent communication supports about 0.3 to 0.6 points, and intensive mentoring about 2 to 2.5 points as a central case. Five points would then serve as a labelled upper bound, consistent with the non-randomised New York estimate. No source documents how effects build week by week, so the engine's phase-in is a modelling assumption.

No UK randomised trial of attendance mentoring was found. Persistence after programmes end is largely unmeasured.

What this means for the design

The engine should not convert attendance gains into attainment gains automatically, given the trials that raised attendance without raising test scores. On this review's conversion, a per-day cost of about 0.004 to 0.016 SD, with 0.05 SD as a high-end case, spans the published designs. The 90 per cent line should act as a reporting threshold, with no change in student behaviour at that point.

Calibration should reproduce the gradients by deprivation and additional support needs and the long tail of severe absence. Calendar effects such as end-of-week absence and term-end peaks are documented for one of the two calibration systems, and this review found no comparable analysis for the other. Because the two differ by about ten points on the same definition, calibration to one should not be assumed to transfer to the other.

Interventions in the engine should vary with the function an absence serves, because parent nudges target misbeliefs rather than anxiety. Absence bunches at the end of the week and around breaks, so drawing each day independently at a fixed rate would understate that clustering.

Each design choice rests on evidence of different strength

The table pairs each design choice with the evidence that supports it and the evidence that challenges it or is missing. The final column rates the evidence for a school-leadership setting, where most of it is indirect.

Conclusion

Synthetic School's split between engine and narrator turns the main validity problem in LLM simulation into a modelling problem with established tests. Failures that would corrupt outcomes become risks to narration, and the engine inherits simulation science's standards for calibration and replication. The white paper's strongest claim is therefore architectural, and claims about effects on leaders should wait for trials.

The review also identifies evidence Synthetic School could generate for the field. A randomised comparison of debrief designs, with transfer measured on new scenarios and in logged decisions, would answer questions both Hallinger-led reviews left open. An A-versus-A run returning zero difference, and attendance effects bounded by randomised trials, would show the engine meets the standards it cites.

References

Entries read only through a summary, a reproduction or a draft say so. Where a page gave no title, a description appears in square brackets. Initials are given only where the source showed them.

Abelairas-Gómez, C., Sawyer, T., Donoghue, A., Cortegiani, A., & Greif, R., on behalf of the ILCOR Education, Implementation and Teams Task Force. (2023, 1 December). Rapid cycle deliberate practice in resuscitation training (EIT6414) TF SR. Consensus on Science with Treatment Recommendations. International Liaison Committee on Resuscitation. https://costr.ilcor.org/document/rapid-cycle-deliberate-practice-in-resuscitation-training-eit6414-tf-sr

ACARA (Australian Curriculum, Assessment and Reporting Authority). (n.d.; page showing 2025 data, accessed 18 October 2026). National Report on Schooling in Australia: Student attendance. https://acara.edu.au/reporting/national-report-on-schooling-in-australia/student-attendance

ACU (Australian Catholic University). (2026, 30 March). Violence against principals nearly doubles in 15 years as workload soars, stress and burnout deepens, ACU study [News release]. https://www.acu.edu.au/about-acu/news/2026/march/violence-against-principals-nearly-doubles-in-15-years-as-workload-soars-stress-and-burnout-deepens

Adornetto, C., Mora, A., Hu, K., Garcia, L. I., Atchade-Adelomou, P., Greco, G., Pastor, L. A. A., & Larson, K. (2025). Generative agents in agent-based modeling: Overview, validation, and emerging challenges. IEEE Transactions on Artificial Intelligence, 6(12), 3165-3183. https://doi.org/10.1109/TAI.2025.3566362 (record: https://www.media.mit.edu/publications/generative-agents-in-agent-based-modeling-overview-validation-and-emerging-challenges/)

AERA. (2026). Improving efficacy: Using mixed-reality simulation to develop self-awareness in emerging school leaders [paper, AERA 2026 Annual Meeting; authors not listed on the programme page]. https://convention2.allacademic.com/one/aera/aera26/online_program_direct_link/view_paper/2274888/

Anderson, E., & Calandra, B. (2026). Mixed-reality simulations in teacher education: A scoping review. Contemporary Issues in Technology and Teacher Education, 26(2). https://citejournal.org/volume-26/issue-2-26/current-practice/mixed-reality-simulations-in-teacher-education-a-scoping-review/

Anthis, Liu, Richardson, Kozlowski, Koch, Brynjolfsson, Evans, & Bernstein. (2025). Position: LLM social simulations are a promising research method. Proceedings of the 42nd International Conference on Machine Learning, PMLR 267, 81005-81034. https://proceedings.mlr.press/v267/anthis25a.html (preprint arXiv:2504.02234: https://arxiv.org/abs/2504.02234)

APCLC (Asia Pacific Centre for Leadership and Change), Education University of Hong Kong. (n.d.). Problem-based leadership development. https://www.eduhk.hk/apclc/tools-PBL.html

Argyle, L. P., Busby, E. C., Fulda, N., Gubler, J. R., Rytting, C., & Wingate, D. (2023). Out of one, many: Using language models to simulate human samples. Political Analysis, 31(3), 337-351. https://doi.org/10.1017/pan.2023.2 (arXiv: https://arxiv.org/abs/2209.06899)

Ashokkumar, A., Hewitt, L., Ghezae, I., & Willer, R. (2026). Large language models can predict the results of social science experiments. Nature, 656(8126), 115-122. https://doi.org/10.1038/s41586-026-10742-x. Read through the Stanford GSB publication listing, which gives these details: https://www.gsb.stanford.edu/faculty-research/publications/large-language-models-can-predict-results-social-science-experiments

Atil et al. (2025). Non-determinism of "deterministic" LLM settings. arXiv:2408.04667 (v5). https://arxiv.org/abs/2408.04667

Attendance Works. (2018). Fill out your pyramid [tiered intervention worksheet, revised 3 December 2018]. https://www.attendanceworks.org/wp-content/uploads/2018/12/Pyramid-worksheet-revised-12-3-18.pdf

Axtell, R., Axelrod, R., Epstein, J. M., & Cohen, M. D. (1996). Aligning simulation models: A case study and results. Computational and Mathematical Organization Theory, 1(2), 123-141. https://doi.org/10.1007/BF01299065 (PDF: https://public.websites.umich.edu/~axe/research/Aligning_Sim.pdf)

Azukas, M. E., Dexter, S., & Gibson, D. (2025). An exploratory study of simulations for leadership development in the principal pipeline. Education Sciences, 15(6), 770. https://doi.org/10.3390/educsci15060770

Balfanz, R., & Byrnes, V. (2013). Meeting the challenge of combating chronic absenteeism: Impact of the NYC Mayor's Interagency Task Force on Chronic Absenteeism and School Attendance and its implications for other cities. Everyone Graduates Center, Johns Hopkins University School of Education. [Undated report; released November 2013.] https://www.nyc.gov/html/truancy/downloads/pdf/meeting_the_challenge_of_combating_chronic_absenteeism.pdf

Balfanz, R., & Byrnes, V. (2018). Using data and the human touch: Evaluating the NYC inter-agency campaign to reduce chronic absenteeism. Journal of Education for Students Placed at Risk, 23(1-2), 107-121. https://doi.org/10.1080/10824669.2018.1435283

Barnett, D. (2025). Coaching for consciousness: Using mixed-reality simulation and reciprocal learning partnerships to develop self-awareness in emerging school leaders [Doctoral dissertation, University of California, Los Angeles]. https://escholarship.org/uc/item/4f8489zt

Bartanen, B., Grissom, J. A., & Rogers, L. K. (2019). The impacts of principal turnover. Educational Evaluation and Policy Analysis, 41(3), 350-374. https://journals.sagepub.com/doi/full/10.3102/0162373719855044

Bartanen, B., Husain, A. N., & Liebowitz, D. D. (2023). Rethinking principal effects on student outcomes [working-paper version, November 2023] (EdWorkingPaper No. 22-621). Annenberg Institute at Brown University. https://doi.org/10.26300/r5sf-3918 (PDF: https://www.edworkingpapers.com/sites/default/files/Principal_Effects_Nov2023.pdf)

Bartanen, B., Husain, A. N., & Liebowitz, D. D. (2024). Rethinking principal effects on student outcomes. Journal of Public Economics, 234, 105115. https://doi.org/10.1016/j.jpubeco.2024.105115 (abstract read through a reproduction: https://academicnewsletter.sufe.edu.cn/info/273109)

Bergman, P., & Chan, E. W. (2021). Leveraging parents through low-cost technology: The impact of high-frequency information on student achievement. Journal of Human Resources, 56(1), 125-158. https://doi.org/10.3368/jhr.56.1.1118-9837R1

Bhowmik, S., West, L., Barrett, A., Zhang, N., Dai, C.-P., Sokolikj, Z., Southerland, S., Yuan, X., & Ke, F. (n.d.; venue and year not shown). Evaluation of an LLM-powered student agent for teacher training. NSF Public Access Repository. https://par.nsf.gov/servlets/purl/10659916

Bisbee, J., Clinton, J. D., Dorff, C., Kenkel, B., & Larson, J. M. (2024). Synthetic replacements for human survey data? The perils of large language models. Political Analysis, 32(4), 401-416. https://doi.org/10.1017/pan.2024.5

Bloom, N., Lemos, R., Sadun, R., & Van Reenen, J. (2015). Does management matter in schools? The Economic Journal, 125(584), 647-674. https://doi.org/10.1111/ecoj.12267 (earlier version: NBER Working Paper 20667, 2014, https://www.nber.org/papers/w20667)

Bondurant, L., & Reinholz, D. (2023). "Rahul is a math nerd" and "Mia can be a drama queen": How mixed-reality simulations can perpetuate racist and sexist stereotypes. Mathematics Teacher Educator, 11(3), 189-209. https://doi.org/10.5951/mte.2021-0041 (record: https://ecrhub.org/publications/rahul-math-nerd-and-mia-can-be-drama-queen-how-mixed-reality-simulations-can)

Branch, G. F., Hanushek, E. A., & Rivkin, S. G. (2013). School leaders matter. Education Next, Winter 2013. https://www.educationnext.org/wp-content/uploads/2022/08/ednext_20131_branch.pdf

Bridges, E. M., & Hallinger, P. (1992). Problem based learning for administrators. ERIC Clearinghouse on Educational Management, University of Oregon. https://scholarsbank.uoregon.edu/items/48c61602-5340-4af1-97dd-48f7989fa0b5/full

Broska, D., Howes, M., & van Loon, A. (2025). The mixed subjects design: Treating large language models as potentially informative observations. Sociological Methods & Research, 54(3), 1074-1109. https://doi.org/10.1177/00491241251326865

Burtonshaw, S. (2023, 21 September). Public First research finds parental support for fulltime schooling has collapsed [summary of focus-group research; full report not opened]. Public First. https://pf.global/insights/public-first-research-finds-parental-support-for-fulltime-schooling-has-collapsed

Bush, T. (2018). Preparation and induction for school principals: Global perspectives. Management in Education, 32(2), 66-71. https://doi.org/10.1177/0892020618761805

Chen, E. J. (2001). Using common random numbers for indifference-zone selection. In Proceedings of the 2001 Winter Simulation Conference (pp. 408-416). https://informs-sim.org/wsc01papers/053.PDF

Cheng, A., Eppich, W., Grant, V., Sherbino, J., Zendejas, B., & Cook, D. A. (2014). Debriefing for technology-enhanced simulation: A systematic review and meta-analysis. Medical Education, 48(7), 657-666. https://doi.org/10.1111/medu.12432

Chernikova, O., Heitzmann, N., Stadler, M., Holzberger, D., Seidel, T., & Fischer, F. (2020). Simulation-based learning in higher education: A meta-analysis. Review of Educational Research, 90(4), 499-541. https://doi.org/10.3102/0034654320933544

Chernikova, O., Holzberger, D., Heitzmann, N., Stadler, M., Seidel, T., & Fischer, F. (2024). Where salience goes beyond authenticity: A meta-analysis on simulation-based learning in higher education. Zeitschrift für Pädagogische Psychologie, 38(1-2), 15-25. https://doi.org/10.1024/1010-0652/a000357

Clifton-Sprigg, J., & James, J. (2025). The "Friday effect": School attendance over the week. British Educational Research Journal, 51(1), 73-92. https://doi.org/10.1002/berj.4066

Cohen, J., Krishnamachari, A., & Wong, V. C. (2021). Experimental evidence on the robustness of coaching supports in teacher education (EdWorkingPaper 21-468). Annenberg Institute at Brown University. https://doi.org/10.26300/dgf9-ca95

Cohen, J., Wong, V., Krishnamachari, A., & Berlin, R. (2020). Teacher coaching in a simulated environment. Educational Evaluation and Policy Analysis, 42(2), 208-231. https://doi.org/10.3102/0162373720906217

Cook, D. A. (2013, 1 March). The literature on health care simulation education: What does it show? [Perspective]. AHRQ PSNet, Agency for Healthcare Research and Quality. https://psnet.ahrq.gov/perspective/literature-health-care-simulation-education-what-does-it-show

Cook, D. A., Hatala, R., Brydges, R., Zendejas, B., Szostek, J. H., Wang, A. T., Erwin, P. J., & Hamstra, S. J. (2011). Technology-enhanced simulation for health professions education: A systematic review and meta-analysis. JAMA, 306(9), 978-988. https://doi.org/10.1001/jama.2011.1234

Cornes, J. (2026, 29 May). Transforming the future of leadership development through virtual reality and artificial intelligence [News item]. Faculty of Education, University of Melbourne. https://education.unimelb.edu.au/news-and-events/news/2026/transforming-the-future-of-leadership-development-through-virtual-reality-and-artificial-intelligence

Culbertson, J. (1995). Building bridges: UCEA's first two decades. University Council for Educational Administration. https://libraopen.lib.virginia.edu/downloads/5x21tf405

D'Addio, A. C., & April, D., with contributors. (2024, 4 November). New PEER profiles support policy dialogue on school leadership. World Education Blog, UNESCO GEM Report. https://world-education-blog.org/2024/11/04/new-peer-profiles-support-policy-dialogue-on-school-leadership/

Dani, S. (2026, 9 March). Practice parent-teacher meetings with AI avatars. DIZH. https://www.dizh.uzh.ch/en/2026/03/09/practice-parent-teacher-meetings-with-ai-avatars/

Darling-Hammond, L., Wechsler, M. E., Levin, S., Leung-Gagné, M., & Tozer, S. (2022). Developing effective principals: What kind of learning matters? Learning Policy Institute and The Wallace Foundation. https://doi.org/10.54300/641.201

DBE (Department of Basic Education, South Africa). (2019). TALIS 2018 South Africa country note, Volume I (content © OECD 2019). https://www.education.gov.za/Portals/0/Documents/Reports/2018%20Talis%20South%20Africa%20Country%20Note%20Final.pdf

de Winter, J. C. F., Dodou, D., & Mulder, M. (2012). Training effectiveness of whole body flight simulator motion: A comprehensive meta-analysis. International Journal of Aviation Psychology, 22(2), 164-183. https://doi.org/10.1080/10508414.2012.663247

DfE (Department for Education). (2022). School leadership in England 2010 to 2020: Characteristics and trends. https://assets.publishing.service.gov.uk/government/uploads/system/uploads/attachment_data/file/1071794/School_leadership_in_England_2010_to_2020_characteristics_and_trends_-_report.pdf

DfE. (2025, 7 August). Pupil attendance in schools: Week 29 2025 (end of 2024/25 academic year). Explore Education Statistics. https://explore-education-statistics.service.gov.uk/find-statistics/pupil-attendance-in-schools/2025-week-29-end-of-24-25-ay

DfE. (2026a). Pupil absence in schools in England: Academic year 2024/25 (published 26 March 2026, updated 21 May 2026). Explore Education Statistics. https://explore-education-statistics.service.gov.uk/find-statistics/pupil-absence-in-schools-in-england/2024-25

DfE. (2026b). Pupil absence in schools in England: Autumn and spring terms 2025/26 (published 15 October 2026). Explore Education Statistics. https://explore-education-statistics.service.gov.uk/find-statistics/pupil-absence-in-schools-in-england/2025-26-autumn-and-spring-term

Dignum, E., Boterman, W., Flache, A., & Lees, M. (2025). Empirically estimating an agent-based model of school choice on household-level register data. Journal of Artificial Societies and Social Simulation, 28(4), 8. https://doi.org/10.18564/jasss.5798

Dotger, B. H. (2015). Core pedagogy: Individual uncertainty, shared practice, formative ethos. Journal of Teacher Education, 66(3), 215-226. https://doi.org/10.1177/0022487115570093

DPMC (Department of the Prime Minister and Cabinet, New Zealand). (2026). Government targets factsheet: Increased student attendance (June 2026 quarter). https://www.dpmc.govt.nz/sites/default/files/2026-09/gt-factsheet-target-6-jun26.pdf

Dräger, J., Klein, M., & Sosu, E. (2024). The long-term consequences of early school absences for educational attainment and labour market outcomes. British Educational Research Journal, 50(4), 1636-1654. https://doi.org/10.1002/berj.3992

Edmonds, B., & ní Aodha, L. (2019). Using agent-based modelling to inform policy – What could possibly go wrong? In P. Davidsson & H. Verhagen (Eds.), Multi-Agent-Based Simulation XIX (MABS 2018), Lecture Notes in Computer Science 11463 (pp. 1-16). Springer. https://doi.org/10.1007/978-3-030-22270-3_1

Education Bureau, Hong Kong. (2024, 26 February). Education Bureau Circular No. 1/2024. https://applications.edb.gov.hk/circular/upload/EDBC/EDBC24001E.pdf

EEF (Education Endowment Foundation), Patel, R., Kay, J., Blausten, H., Madgwick, H., & Stevenson, J. (2022). Attendance interventions: Rapid evidence assessment. https://educationendowmentfoundation.org.uk/education-evidence/evidence-reviews/attendance-interventions-rapid-evidence-assessment (report PDF: https://d2tic4wvo1iusb.cloudfront.net/production/documents/pages/Attendance-REA-report.pdf)

Eklund, K., Burns, M. K., Oyen, K., DeMarchena, S., & McCollom, E. M. (2022). Addressing chronic absenteeism in schools: A meta-analysis of evidence-based interventions. School Psychology Review, 51(1), 95-111. https://doi.org/10.1080/2372966X.2020.1789436

Ellis, S., & Davidi, I. (2005). After-event reviews: Drawing lessons from successful and failed experience. Journal of Applied Psychology, 90(5), 857-871. https://doi.org/10.1037/0021-9010.90.5.857

Ericsson, K. A. (2008). Deliberate practice and acquisition of expert performance: A general overview. Academic Emergency Medicine, 15(11), 988-994. https://doi.org/10.1111/j.1553-2712.2008.00227.x

ERO (Education Review Office, New Zealand). (2023). 'Everything was new': Preparing and supporting new principals [summary report]. https://www.evidence.ero.govt.nz/media/g1nnyzax/new-principals-report-series-summary.pdf

Gates, S. M., Baird, M. D., Master, B. K., & Chavez-Herrerias, E. R. (2019). Principal pipelines: A feasible, affordable, and effective way for districts to improve schools (RR-2666). RAND Corporation. https://wallacefoundation.org/sites/default/files/2023-07/Principal-Pipelines-A-Feasible-Affordable-and-Effective-Way-for-Districts-to-Improve-Schools.pdf

Ghaffarzadegan, N., Majumdar, A., Williams, R., & Hosseinichimeh, N. (2024). Generative agent-based modeling: An introduction and tutorial. System Dynamics Review, 40(1), e1761. https://doi.org/10.1002/sdr.1761 (arXiv: https://arxiv.org/abs/2309.11456)

Goldhaber, D., Kasman, M., Quince, V., Theobald, R., & Wolff, M. (2022). How did it get this way? Disentangling the sources of teacher quality gaps through agent-based modeling (CALDER Working Paper 259-0122). https://www.edworkingpapers.com/sites/default/files/CALDER%20WP%20259-0122.pdf (peer-reviewed version: Social Science Research, 116, 102941, 2023, https://doi.org/10.1016/j.ssresearch.2023.102941)

Gomez, J. J., & Stephenson, R. S. (1987). Validity of an assessment center for the selection of school-level administrators. Educational Evaluation and Policy Analysis, 9(1), 1-7. https://doi.org/10.3102/01623737009001001

Gonnermann-Müller, J., Haase, J., Leins, N., Kosch, T., & Pokutta, S. (2026). LLM-based educational simulation: Evaluating temporal student persona stability across ADHD profiles. arXiv:2605.06307. https://arxiv.org/abs/2605.06307

Gonzálvez, C., Inglés, C. J., Kearney, C. A., Vicent, M., Sanmartín, R., & García-Fernández, J. M. (2016). School Refusal Assessment Scale-Revised: Factorial invariance and latent means differences across gender and age in Spanish children. Frontiers in Psychology, 7, 2011. https://doi.org/10.3389/fpsyg.2016.02011

Goodman, J. (2014). Flaking out: Student absences and snow days as disruptions of instructional time (NBER Working Paper 20221). https://doi.org/10.3386/w20221

Greany, T., Thomson, P., Perry, T., & Collins, M. (2025). Signature pedagogies for leadership development: A comparison of headteacher preparation programmes in England and Scotland. London Review of Education, 23(1), 18. https://doi.org/10.14324/LRE.23.1.18

Grimm, V., Railsback, S. F., Vincenot, C. E., Berger, U., Gallagher, C., DeAngelis, D. L., Edmonds, B., Ge, J., Giske, J., Groeneveld, J., Johnston, A. S. A., Milles, A., Nabe-Nielsen, J., Polhill, J. G., Radchuk, V., Rohwäder, M.-S., Stillman, R. A., Thiele, J. C., & Ayllón, D. (2020). The ODD protocol for describing agent-based and other simulation models: A second update to improve clarity, replication, and structural realism. Journal of Artificial Societies and Social Simulation, 23(2), 7. https://doi.org/10.18564/jasss.4259

Grimm, V., Revilla, E., Berger, U., Jeltsch, F., Mooij, W. M., Railsback, S. F., Thulke, H.-H., Weiner, J., Wiegand, T., & DeAngelis, D. L. (2005). Pattern-oriented modeling of agent-based complex systems: Lessons from ecology. Science, 310(5750), 987-991. https://doi.org/10.1126/science.1116681 (abstract read on the USGS record: https://pubs.usgs.gov/publication/70161783)

Grissom, J. A., Donaldson, M. L., Rigby, J. G., Doughty, M., & Forman, S. R. (2026). Understanding the landscape of professional learning for school principals. Vanderbilt University; The Wallace Foundation. https://doi.org/10.59656/EL-LS3669.001

Grissom, J. A., Egalite, A. J., & Lindsay, C. A. (2021). How principals affect students and schools: A systematic synthesis of two decades of research. The Wallace Foundation. https://www.wallacefoundation.org/knowledge-center/Documents/How-Principals-Affect-Students-and-Schools.pdf

Gubbels, J., van der Put, C. E., & Assink, M. (2019). Risk factors for school absenteeism and dropout: A meta-analytic review. Journal of Youth and Adolescence, 48(9), 1637-1667. https://doi.org/10.1007/s10964-019-01072-5

Guryan, J., Christenson, S., Cureton, A., Lai, I., Ludwig, J., Schwarz, C., Shirey, E., & Turner, M. C. (2020). The effect of mentoring on school attendance and academic outcomes: A randomized evaluation of the Check & Connect program (IPR Working Paper WP-20-40). Northwestern University. https://www.ipr.northwestern.edu/documents/working-papers/2020/wp-20-40.pdf (peer-reviewed version: Journal of Policy Analysis and Management, 40(3), 841-882, https://doi.org/10.1002/pam.22264)

Hallinger, P., Banoğlu, K., & Gümüş, S. (2025). Scoping review of research on digital simulations and serious games in educational leadership and management. Leadership and Policy in Schools, 24(1), 178-199. https://doi.org/10.1080/15700763.2024.2447315

Hallinger, P., & Bridges, E. M. (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), 255-288. https://doi.org/10.1177/0013161X16659347

Hamstra, S. J., Brydges, R., Hatala, R., Zendejas, B., & Cook, D. A. (2014). Reconsidering fidelity in simulation-based training. Academic Medicine, 89(3), 387-392. https://doi.org/10.1097/ACM.0000000000000130

Hancock, K. J., Shepherd, C. C. J., Lawrence, D., & Zubrick, S. R. (2013). Student attendance and educational outcomes: Every day counts. Report for the Department of Education, Employment and Workplace Relations, Canberra. Telethon Institute for Child Health Research. https://thekids.org.au/globalassets/media/images/pagessections/news--events/2013/march/final_report_2013.pdf

Hayden, J. K., Smiley, R. A., Alexander, M., Kardong-Edgren, S., & Jeffries, P. R. (2014). The NCSBN National Simulation Study: A longitudinal, randomized, controlled study replacing clinical hours with simulation in prelicensure nursing education. Journal of Nursing Regulation, 5(2 Suppl), S3-S40. https://doi.org/10.1016/S2155-8256(15)30062-4 (PDF: https://www.ncsbn.org/public-files/JNR_Simulation_Supplement.pdf)

Hemphill, J. K., Griffiths, D., Frederiksen, N. O., Stice, G., Iannaccone, L., Coffield, W. E., & Carlton, S. T. (1961). Dimensions of administrative performance (ETS Research Bulletin RB-61-10). Educational Testing Service. https://doi.org/10.1002/j.2333-8504.1961.tb00108.x

Hemphill, J. K., Griffiths, D. E., & Frederiksen, N. (1962). Administrative performance and personality: A study of the principal in a simulated elementary school. Bureau of Publications, Teachers College, Columbia University. https://archive.org/details/administrativepe00colu (ETS catalogue record: https://www.ets.org/research/policy_research_reports/publications/book/1962/bhey.html)

Hewitt, Ashokkumar, Ghezae, & Willer. (2024). Predicting results of social science experiments using large language models [preprint, read through the Stanford project page]. https://ai4pb.stanford.edu/projects/predicting-results-of-social-science-experiments-using-large-language-models

Hu, T., Baumann, J., Lupo, L., Collier, N., Hovy, D., & Röttger, P. (2025). SimBench: Benchmarking the ability of large language models to simulate human behaviors. arXiv:2510.17516. https://arxiv.org/abs/2510.17516v1

Ilgen, J. S., Sherbino, J., & Cook, D. A. (2013). Technology-enhanced simulation in emergency medicine: A systematic review and meta-analysis. Academic Emergency Medicine, 20(2), 117-127. https://doi.org/10.1111/acem.12076

Issenberg, S. B., & Scalese, R. J. (2007). Best evidence on high-fidelity simulation: What clinical teachers need to know. The Clinical Teacher, 4(2), 73-77. https://doi.org/10.1111/j.1743-498x.2007.00161.x (PDF: https://laerdal.com/cdn-4adf0d/globalassets/documents/Issenberg_Scalese_Clinical_Teacher_BEME_Review.pdf; the authors' summary of Issenberg, McGaghie, Petrusa, Lee Gordon and Scalese's 2005 BEME review in Medical Teacher, 27(1), 10-28, which was not opened)

Jin, S., Wang, H., Gao, Z., Yang, Y., Chunjia, B., & Wang, C. (2025). Evolution in simulation: AI-Agent School with dual memory for high-fidelity educational dynamics. In Findings of the Association for Computational Linguistics: EMNLP 2025 (pp. 5843-5857). https://doi.org/10.18653/v1/2025.findings-emnlp.312

Kasman, M. (2023, 28 November). How can we address racial inequalities in access to experienced teachers? Brookings Institution. https://www.brookings.edu/articles/how-can-we-address-racial-inequalities-in-access-to-experienced-teachers

Kearney, C. A., & Graczyk, P. A. (2014). A response to intervention model to promote school attendance and decrease school absenteeism. Child and Youth Care Forum, 43(1), 1-25. https://doi.org/10.1007/s10566-013-9222-1

Keiser, N. L. (2024). A systematic review of technology in the after-action review (or debrief). Organizational Psychology Review, 14(3), 449-477. https://doi.org/10.1177/20413866241245314

Keiser, N. L., & Arthur, W., Jr. (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), 1007-1032. https://doi.org/10.1037/apl0000821 (abstract read through a reproduction: https://academicnewsletter.sufe.edu.cn/info/369459)

Keith, N., & Frese, M. (2008). Effectiveness of error management training: A meta-analysis. Journal of Applied Psychology, 93(1), 59-69. https://doi.org/10.1037/0021-9010.93.1.59 (read as a post-review preprint: https://www.humanw.tu-darmstadt.de/media/orgpsy/preversioneffectivenessoferrormanagement.pdf)

Keo, S.-D. (c. 2016; no date on the page). Shaping strong principals in Singapore: Success by design. National Center on Education and the Economy. https://ncee.org/2016/02/shaping-strong-principals-in-singapore-success-by-design

Klein, D. J., Abeysuriya, R. G., Stuart, R. M., & Kerr, C. C. (2024). Noise-free comparison of stochastic agent-based simulations using common random numbers. arXiv:2409.02086 (v2). https://arxiv.org/html/2409.02086v2

Klein, M., Sosu, E. M., & Dare, S. (2022). School absenteeism and academic achievement: Does the reason for absence matter? AERA Open, 8(1), 1-14. https://doi.org/10.1177/23328584211071115

Landman, A., Mol, D., van Emmerik, M. L., & Groen, E. L. (2022). Negative transfer of training: Simulator study into effects of going beyond alarms during stall recovery training (DOT/FAA/TC-22/10; TNO-2022-R10427). https://doi.org/10.21949/1524474

Larooij, M., & Törnberg, P. (2026). Validation is the central challenge for generative social simulation: A critical review of LLMs in agent-based modeling. Artificial Intelligence Review, 59(1), Article 15. https://doi.org/10.1007/s10462-025-11412-6 (online 18 November 2025; preprint arXiv:2504.03274: https://arxiv.org/abs/2504.03274)

L'Ecuyer, P., Simard, R., Chen, E. J., & Kelton, W. D. (2002). An object-oriented random-number package with many long streams and substreams. Operations Research, 50(6), 1073-1075. https://doi.org/10.1287/opre.50.6.1073.358

Lee, J.-S., Filatova, T., Ligmann-Zielinska, A., Hassani-Mahmooei, B., Stonedahl, F., Lorscheid, I., Voinov, A., Polhill, G., Sun, Z., & Parker, D. C. (2015). The complexities of agent-based modeling output analysis. Journal of Artificial Societies and Social Simulation, 18(4), 4. https://doi.org/10.18564/jasss.2897

Li, J., & O'Donoghue, C. (2013). A survey of dynamic microsimulation models: Uses, model structure and methodology. International Journal of Microsimulation, 6(2), 3-55. https://doi.org/10.34196/ijm.00082

Li, J., Lai, Y., Li, W., Ren, J., Zhang, M., Kang, X., Wang, S., Li, P., Zhang, Y.-Q., Ma, W., & Liu, Y. (2025). Agent Hospital: A simulacrum of hospital with evolvable medical agents. arXiv:2405.02957 (v3, 17 January 2025; v1, 5 May 2024). https://arxiv.org/abs/2405.02957

Li, Liu, Bashkansky, Bau, Viégas, Pfister, & Wattenberg. (2024). Measuring and controlling instruction (in)stability in language model dialogs. COLM 2024, per the arXiv record. arXiv:2402.10962. https://arxiv.org/abs/2402.10962

Liebowitz, D. D., & Porter, L. (2019). The effect of principal behaviors on student, teacher, and school outcomes: A systematic review and meta-analysis of the empirical literature. Review of Educational Research, 89(5), 785-827. https://doi.org/10.3102/0034654319866133

Liu, J., Lee, M., & Gershenson, S. (2021). The short- and long-run impacts of secondary school absences. Journal of Public Economics, 199, 104441. https://doi.org/10.1016/j.jpubeco.2021.104441 (open-access version: EdWorkingPaper No. 19-125, Annenberg Institute at Brown University, https://doi.org/10.26300/xg6s-z169)

Low, M. J. W., Yeo, H. T. J., Samarasekera, D. D., Chan, G. W. H., & Lee, S. S. (2025). Simulation instructional design features with differences in clinical outcomes: A narrative review. The Asia Pacific Scholar, 10(4), 5-25. https://doi.org/10.29060/TAPS.2025-10-4/RA3572

Macnamara, B. N., Hambrick, D. Z., & Oswald, F. L. (2014). Deliberate practice and performance in music, games, sports, education, and professions: A meta-analysis. Psychological Science, 25(8), 1608-1618. https://doi.org/10.1177/0956797614535810 (read through the Association for Psychological Science press release: https://www.psychologicalscience.org/index.php/news/releases/becoming-an-expert-takes-more-than-practice.html)

Malkus, N. (2024, 6 September; modified 20 June 2025). The latest chronic absenteeism numbers. American Enterprise Institute. https://cosm.aei.org/the-latest-chronic-absenteeism-numbers/

Mann, D., Reardon, R. M., Becker, J. D., Shakeshaft, C., & Bacon, N. (2011). Immersive, interactive, web-enabled computer simulation as a trigger for learning: The next generation of problem-based learning in educational leadership. Journal of Research on Leadership Education, 6(5), 272-287. https://doi.org/10.1177/194277511100600511 (record: https://scholarscompass.vcu.edu/edlp_pubs/5)

Mann, D., & Shakeshaft, C. (2013). Virtual school, real experience: Simulations replicate the world of practice for aspiring principals. Journal of Staff Development (now The Learning Professional), 34(6), 43-45. Learning Forward. https://learningforward.org/journal/december-2013-vol-34-no-6/virtual-school-real-experience/

Manzo, G. (2013). Educational choices and social interactions: A formal model and a computational test. Comparative Social Research, 30, 47-100. https://doi.org/10.1108/S0195-6310(2013)0000030007

Markel, J. M., Opferman, S. G., Landay, J. A., & Piech, C. (2023). GPTeach: Interactive TA training with GPT-based students [EdArXiv preprint]. https://doi.org/10.35542/osf.io/r23bu (record: https://scale.stanford.edu/ai/repository/gpteach-interactive-ta-training-gpt-based-students)

Maroulis, S., Bakshy, E., Gomez, L., & Wilensky, U. (2014). Modeling the transition to public school choice. Journal of Artificial Societies and Social Simulation, 17(2), 3. https://www.jasss.org/17/2/3/3.pdf

Maroulis, S., Guimerà, R., Petry, H., Stringer, M. J., Gomez, L. M., Amaral, L. A. N., & Wilensky, U. (2010). Complex systems view of educational policy research. Science, 330, 38-39. https://doi.org/10.1126/science.1195153

Masotya, M., Hovmand, P., Snyder, B., & Ronis, S. (2025). Chronic school absenteeism among children with asthma [Abstract 1341]. 2025 International System Dynamics Conference, Boston, 4-8 August 2025. https://proceedings.systemdynamics.org/2025/abstracts/1341.html

Maynard, B. R., McCrea, K. T., Pigott, T. D., & Kelly, M. S. (2012). Indicated truancy interventions: Effects on school attendance among chronic truant students. Campbell Systematic Reviews, 8(1), 1-84. https://doi.org/10.4073/csr.2012.10

Melvin, G., McKay-Brown, L., Heyne, D., & Cameron, L. (2025). Interventions to promote school attendance and address student absence: Rapid literature review. Australian Education Research Organisation. https://edresearch.edu.au/sites/default/files/2024-12/interventions-promote-school-attendance-address-student-absence-aa.pdf

Millington, J., Butler, T., & Hamnett, C. (2014). Aspiration, attainment and success: An agent-based model of distance-based school allocation. Journal of Artificial Societies and Social Simulation, 17(1), 10. https://doi.org/10.18564/jasss.2332

Milne, B. J., Lay-Yee, R., McLay, J. M., Pearson, J., von Randow, M., & Davis, P. (2015). Modelling the Early life-Course (MELC): A microsimulation model of child development in New Zealand. International Journal of Microsimulation, 8(2), 28-60. https://doi.org/10.34196/ijm.00116

Ministry of Education, New Zealand. (n.d.; accessed 18 October 2026). Education Counts: Attendance. https://www.educationcounts.govt.nz/statistics/attendance

Mitchell, A. A., & Ivimey-Cook, E. R. (2023). Technology-enhanced simulation for healthcare professionals: A meta-analysis. Frontiers in Medicine, 10, 1149048. https://doi.org/10.3389/fmed.2023.1149048

Monks, T., Currie, C. S. M., Onggo, B. S., Robinson, S., Kunc, M., & Taylor, S. J. E. (2019). Strengthening the reporting of empirical simulation studies: Introducing the STRESS guidelines. Journal of Simulation, 13(1), 55-67. https://doi.org/10.1080/17477778.2018.1442155

Morris, M. W., & Moore, P. C. (2000). The lessons we (don't) learn: Counterfactual thinking and organizational accountability after a close call. Administrative Science Quarterly, 45(4), 737-765. https://doi.org/10.2307/2667018 (record: https://business.columbia.edu/faculty/research/lessons-we-dont-learn-counterfactual-thinking-and-organizational-accountability)

Morrison, A. R. (2024). Survival at the top: Issues and challenges affecting the tenure of international school heads [EdD thesis, OISE, University of Toronto]. https://utoronto.scholaris.ca/server/api/core/bitstreams/2eb9f131-ae3e-41bf-9caf-37c4684d30a6/content

NCES (National Center for Education Statistics). (2023, 31 July). Roughly one in ten public school principals left profession in 2021-22 school year [Press release]. Institute of Education Sciences. https://ies.ed.gov/learn/press-release/roughly-one-ten-public-school-principals-left-profession-2021-22-school-year

Ng, C., Primiani, N., & Orchanian-Cheff, A. (2021). Rapid cycle deliberate practice in healthcare simulation: A scoping review. Medical Science Educator, 31(6), 2105-2120. https://doi.org/10.1007/s40670-021-01446-0

NIE (National Institute of Education, Singapore). (2016, 14 October). NIE celebrates Leaders in Education Programme graduation [Media release]. https://www.nas.gov.sg/archivesonline/data/pdfdoc/20161014009/NIE_Media%20Release_LEP%202016.pdf

Norman, G., Dore, K., & Grierson, L. (2012). The minimal relationship between simulation fidelity and transfer of learning. Medical Education, 46(7), 636-647. https://doi.org/10.1111/j.1365-2923.2012.04243.x

Panovska-Griffiths, J., Kerr, C. C., Stuart, R. M., Mistry, D., Klein, D. J., Viner, R. M., & Bonell, C. (2020). Determining the optimal strategy for reopening schools, the impact of test and trace interventions, and the risk of occurrence of a second COVID-19 epidemic wave in the UK: A modelling study. The Lancet Child & Adolescent Health, 4(11), 817-827. https://doi.org/10.1016/S2352-4642(20)30250-9

Pantili, L. D. (1992). A meta-analytic validity study of the National Association of Secondary School Principals' Assessment Center process [PhD dissertation, Virginia Tech]. https://vtechworks.lib.vt.edu/items/60f45e26-3f14-4ae6-b042-add48badc15b/full

Park, J. S., O'Brien, J. C., Cai, C. J., Morris, M. R., Liang, P., & Bernstein, M. S. (2023). Generative agents: Interactive simulacra of human behavior. In The 36th Annual ACM Symposium on User Interface Software and Technology (UIST '23) (Article 2, 22 pp.). https://doi.org/10.1145/3586183.3606763 (arXiv: https://arxiv.org/abs/2304.03442)

Park, J. S., Zou, C. Q., and seven co-authors. (2024). Generative agent simulations of 1,000 people (arXiv:2411.10109, version 1, 15 November 2024). https://arxiv.org/pdf/2411.10109v1

Park, Zou, Kamphorst, Egan, Shaw, Hill, Cai, Morris, Liang, Willer, & Bernstein. (2026). LLM agents grounded in self-reports enable general-purpose simulation of individuals (arXiv:2411.10109, version 3, 28 June 2026). https://arxiv.org/pdf/2411.10109v3

Paul, M. (2021, 5 May). Umbrella summary: After-action reviews [practitioner research summary drawing mainly on Keiser & Arthur, 2020 online / 2021 in print]. Quality Improvement Center for Workforce Development. https://www.qic-wd.org/sites/default/files/Umbrella%20Summary%20-%20After-Action%20Reviews%20050221.pdf

Pensiero, N., & Brede, M. (2024). An agent-based model of school tracking, accountability, and segregation. Journal of Simulation, 18(4), 524-538. https://doi.org/10.1080/17477778.2024.2326138

Prasetyo, M. M., Nurhidayah, & Muslaini, F. (2025). Utilization of digital twin technology in education ecosystem: A systematic literature review. Indonesian Journal of Educational Technology, 4(2), 13-26. https://journal.unm.ac.id/index.php/IJET/article/view/11418

Rice, W. K. M., Wynne, B., Martin, V. J., & Ackland, G. J. (2020). Effect of school closures on mortality from coronavirus disease 2019: Old and new predictions. BMJ, 371, m3588. https://doi.org/10.1136/bmj.m3588

Robinson, C. D., Lee, M. G., Dearing, E., & Rogers, T. (2018). Reducing student absenteeism in the early grades by targeting parental beliefs. American Educational Research Journal, 55(6), 1163-1192. https://doi.org/10.3102/0002831218772274 (read in the author copy: https://toddrogers.scholars.harvard.edu/resource/robinsonleedearingrogers2018reducingabsenteeismintheearlygradespdf)

Robinson, V. M. J., Lloyd, C. A., & Rowe, K. J. (2008). The impact of leadership on student outcomes: An analysis of the differential effects of leadership types. Educational Administration Quarterly, 44(5), 635-674. https://doi.org/10.1177/0013161X08321509

Rogers, T., & Feller, A. (2018). Reducing student absences at scale by targeting parents' misbeliefs. Nature Human Behaviour, 2(5), 335-342. https://doi.org/10.1038/s41562-018-0328-1 (details read in an earlier author version, titled "Reducing Student Absences at Scale By Involving Families": https://toddrogers.scholars.harvard.edu/resource/reducingstudentabsencesatscalepdf)

Saltelli, A., Bammer, G., Bruno, I., Charters, E., Di Fiore, M., Didier, E., Nelson Espeland, W., Kay, J., Lo Piano, S., Mayo, D., Pielke Jr, R., Portaluri, T., Porter, T. M., Puy, A., Rafols, I., Ravetz, J. R., Reinert, E., Sarewitz, D., Stark, P. B., Stirling, A., van der Sluijs, J., & Vineis, P. (2020). Five ways to ensure that models serve society: A manifesto. Nature, 582(7813), 482-484. https://doi.org/10.1038/d41586-020-01812-9

Santurkar, S., Durmus, E., Ladhak, F., Lee, C., Liang, P., & Hashimoto, T. (2023). Whose opinions do language models reflect? Proceedings of the 40th International Conference on Machine Learning, PMLR 202, 29971-30004. https://proceedings.mlr.press/v202/santurkar23a.html

Sargent, R. G. (2005). Verification and validation of simulation models. In Proceedings of the 2005 Winter Simulation Conference (pp. 130-143). https://www.informs-sim.org/wsc05papers/014.pdf

Sclar, Choi, Tsvetkov, & Suhr. (2024). Quantifying language models' sensitivity to spurious features in prompt design or: How I learned to start worrying about prompt formatting. ICLR 2024, per the paper's header. arXiv:2310.11324. https://arxiv.org/abs/2310.11324

Scottish Government. (2019). The Head Teachers Education and Training Standards (Scotland) Regulations 2019: Business and regulatory impact assessment. https://dera.ioe.ac.uk/id/eprint/33954/1/head-teachers-education-training-standards-scotland-regulations-2019-final-business-regulatory-impact-assessment.pdf

Scottish Government. (2025, 9 December). School attendance, absence and exclusions statistics 2024-25: Attendance and absence. https://www.gov.scot/publications/school-attendance-absence-and-exclusions-statistics-2024-25/pages/attendance-and-absence/

Seith, E. (2025, 7 October). You think international heads don't last long? That's just a myth. Tes Magazine. https://www.tes.com/magazine/analysis/general/keeping-international-school-heads-in-post-for-longer

Shapira-Lishchinsky, O. (2015). Simulation-based constructivist approach for education leaders. Educational Management Administration & Leadership, 43(6), 972-988. https://doi.org/10.1177/1741143214543203

Sharma et al. (19 authors). (2024). Towards understanding sycophancy in language models. ICLR 2024, per the paper's header. arXiv:2310.13548. https://arxiv.org/abs/2310.13548

Shin, J.-S. (申仲植). (1989). Simulation game in organizational development of school [學校組織發展을 위한 模擬狀況技法에 관한 硏究]. Gyoyuk Nonchong (교육논총), 8, 5-42. Institute of Education, Kookmin University. https://scholar.kyobobook.co.kr/article/detail/4010019858419

Sitzmann, T. (2010, October). Game on? The effectiveness of game use in the workplace depends on context and design. T+D, p. 20. American Society for Training and Development. https://media.moresteam.com/main/downloads/sitzmann-games2010.pdf

Sitzmann, T. (2011). A meta-analytic examination of the instructional effectiveness of computer-based simulation games. Personnel Psychology, 64(2), 489-528. https://doi.org/10.1111/j.1744-6570.2011.01190.x. Abstract read through a reproduction in Kapp, K. M., "Abstracts of three meta-analysis studies of serious games": https://karlkapp.com/?p=4333

SLATE (Centre for the Science of Learning & Technology), University of Bergen. (n.d.). PRACTICE: Parent Role AI Chatbot Training for Interactive Communication Exercises [project page]. https://slate.uib.no/projects/practice

Spillane, J. P., & Lee, L. C. (2013, January draft). Novice school principals' sense of ultimate responsibility: Problems of practice in transitioning to the principal's office (IPR Working Paper WP-13-05). Northwestern University. https://www.ipr.northwestern.edu/documents/working-papers/2013/IPR-WP-13-05.pdf (later published in Educational Administration Quarterly, 50(3), 431-465, online 2013, print 2014: https://doi.org/10.1177/0013161X13505290)

Stewart, J. E., Johnson, D. M., & Howse, W. R. (2008). Fidelity requirements for Army aviation training devices: Issues and answers (Research Report 1887). US Army Research Institute for the Behavioral and Social Sciences. https://apps.dtic.mil/sti/pdfs/ADA493584.pdf

Tang, S. (2015). Learning to make change happen in Chinese schools: Adapting a problem-based computer simulation for training school leaders [EdD thesis, Hong Kong Institute of Education]. https://repository.eduhk.hk/en/publications/learning-to-make-change-happen-in-chinese-schools-adapting-a-prob-5/

Tannenbaum, S. I., & Cerasoli, C. P. (2013). Do team and individual debriefs enhance performance? A meta-analysis. Human Factors, 55(1), 231-245. https://doi.org/10.1177/0018720812448394

Taylor, G. S., & Barnett, J. S. (2013). Evaluation of wearable simulation interface for military training. Human Factors, 55(3), 672-690. https://doi.org/10.1177/0018720812466892

Theelen, H., van den Beemt, A., & den Brok, P. (2019). Classroom simulations in teacher education to support preservice teachers' interpersonal competence: A systematic literature review. Computers & Education, 129, 14-26. https://doi.org/10.1016/j.compedu.2018.10.015

Tufino, E. (2026). A multi-agent AI classroom based on dual-process reasoning hazards: A pilot with prospective physics teachers. arXiv:2608.05820. https://arxiv.org/abs/2608.05820

Ukanwa, K., Jones, A. C., & Turner, B. L., Jr. (2022). School choice increases racial segregation even when parents do not care about race. Proceedings of the National Academy of Sciences, 119(35), e2117979119. https://doi.org/10.1073/pnas.2117979119 (record: https://ideas.repec.org/a/nas/journl/v119y2022pe2117979119.html)

UNESCO. (2024). Leadership in education: Lead for learning. Global Education Monitoring Report 2024/5, summary report (ED/GEM/MRT/2024/TB/1). https://unesco.at/fileadmin/user_upload/short_20245_GEM_report_Leadership_in_education_lead_for_learning.pdf

Vezhnevets, Agapiou, Aharon, Ziv, Matyas, Duéñez-Guzmán, Cunningham, Osindero, Karmon, & Leibo. (2023). Generative agent-based modeling with actions grounded in physical, social, or digital space using Concordia. arXiv:2312.03664 (v2). https://arxiv.org/abs/2312.03664

Wang, A., Morgenstern, J., & Dickerson, J. P. (2025). Large language models that replace human participants can harmfully misportray and flatten identity groups. Nature Machine Intelligence, 7(3), 400-411. https://doi.org/10.1038/s42256-025-00986-z (preprint arXiv:2402.01908: https://arxiv.org/abs/2402.01908)

Wang, R., Todd, G., Xiao, Z., Yuan, X., Côté, M.-A., Clark, P., & Jansen, P. (2024). Can language models serve as text-based world simulators? In Proceedings of ACL 2024 (Vol. 2, Short Papers, pp. 1-17). https://doi.org/10.18653/v1/2024.acl-short.1

Whittaker, F. (2024, 9 February). NPQs not up to scratch for increasingly 'complex' headship. Schools Week. Reports a 2024 report by Big Education and the Centre for Education and Youth, which was not opened. https://schoolsweek.co.uk/Npqs-Not-Up-To-Scratch-For-Increasingly-Complex-Headship

Wilensky, U., & Rand, W. (2007). Making models match: Replicating an agent-based model. Journal of Artificial Societies and Social Simulation, 10(4), 2. https://www.jasss.org/10/4/2.html

Windrum, P., Fagiolo, G., & Moneta, A. (2007). Empirical validation of agent-based models: Alternatives and prospects. Journal of Artificial Societies and Social Simulation, 10(2), 8. https://www.jasss.org/10/2/8.html

Worth, J., Tang, S., & McLean, D. (2026). NFER teacher supply forecast and simulation model: Technical report. National Foundation for Educational Research. https://www.nfer.ac.uk/media/ahhkmymw/teacher-supply-forecast-and-simulation-model-technical-report.pdf

Xu, S., Zhang, X., & Qin, L. (2024). EduAgent: Generative student agents in learning. arXiv:2404.07963. https://arxiv.org/abs/2404.07963

Yuan, Xiao, Li, Xuan, Tong, Diab, & Mitchell. (2026). Towards valid student simulation with large language models. arXiv:2601.05473. https://arxiv.org/abs/2601.05473

Zhang, Z., Zhang-Li, D., Yu, J., Gong, L., Zhou, J., Hao, Z., Jiang, J., Cao, J., Liu, H., Liu, Z., Hou, L., & Li, J. (2025). Simulating classroom education with LLM-empowered agents. In Proceedings of NAACL 2025 (Vol. 1, Long Papers, pp. 10364-10379). https://doi.org/10.18653/v1/2025.naacl-long.520

Zhou, H., Dorsman, J. L., Mandjes, M., & Snelder, M. (2023). On the use of common random numbers in activity-based travel demand modeling for scenario comparison. Transportation Planning and Technology, 46(3), 359-379. https://doi.org/10.1080/03081060.2023.2182784