The Infrastructure of Thinking: How Societies Build Independent Minds

Most countries leave them to heroic individuals. This is the engineering brief for doing it on purpose, with Czechia as the worked example.

ENSI research report — the second of three. Built on a library of more than 400 primary documents: 220 downloaded for this study across 26 angles, and 186 reused from sibling ENSI libraries. It draws on education trials and meta-analyses, developmental and social psychology, organisational research, cross-national value surveys and official Czech and OECD data. The first report, “The Decisive Skill”, sets out the evidence on what independent judgement is and why it matters. This report sets out how to build the environments that produce it. Quotations from Czech-language sources are given in English translation. Compiled 22 September 2026.

The argument, before the list

Countries do not leave their motorways, power grids or water systems to chance. They specify them, assign owners, fund maintenance and measure performance. The environments in which people learn to think for themselves are left almost everywhere to accident: a gifted teacher here, a debate club there, a parent who argues at the dinner table. The first report in this series showed why that is no longer acceptable. Answers are becoming nearly free and judgement scarce. The junior roles in which judgement used to be learned are the first to be automated. And the absence of independent judgement is an institutional hazard, not merely a personal failing.

Calling the response an infrastructure of thinking changes the question. The question stops being “how do we find more inspiring teachers?” and becomes: what are the components, what does each require to work, who owns it, how is it scaled without losing its effect, and how do we know it is working?

The evidence gives the components a clear logic. Independent judgement has two halves, and each is grown by a different kind of environment.

The thinking half is grown by friction with scaffolding.

• The combination works. The largest meta-analysis of critical-thinking instruction finds effects of g+ = 0.30 on average. Designs that combine authentic problems, dialogue and mentoring reach 0.57 (Abrami et al. 2015 · angle 01).

• Thinking programmes work and last. A meta-analysis of thinking-skills programmes finds effects of 0.62 on both cognitive and curricular outcomes (Higgins et al. 2005 · angle 19). Metacognitive strategy instruction produces effects that grow after it ends (de Boer et al. 2018 · angle 19).

• Unguided friction fails. Inquiry learning works when it is guided, and unguided discovery fails novices (Lazonder & Harmsen · angle 07; Kirschner et al. · XL).

The courage half is grown by allies and belonging.

• Two defiant peers. When two peers defied the experimenter in Milgram’s obedience paradigm, 36 of 40 participants defied too, against 14 of 40 without them (Milgram 1965 · angle 20).

• One truthful partner. A single truthful partner cut the pressure of an Asch majority to a quarter (angle 10).

• Belonging where it is real. Belonging interventions improve persistence where the setting offers real opportunities to belong (Walton et al. 2023 · angle 13).

The hard lesson is that these environments do not scale by themselves.

• Philosophy for Children. It produced about two months’ extra progress in an efficacy trial. In a larger effectiveness trial run at scale, it produced “no evidence” of impact (angle 19).

• Cognitive Acceleration. It raised national examination results years later in its original schools, while a later large replication found no effect on science attainment (angle 19).

The infrastructure therefore has to be engineered like a clinical service: trained facilitators, fidelity checks, sufficient dose, and outcome measurement built in from the start.

Around the two layers sit two further requirements. The first is an open talent funnel. Children from the richest US families are ten times as likely to become inventors as children from below-median families, and differences in early ability explain little of the gap (Bell et al. · angle 14). In Czechia, official statistics identify 0.3% of basic-school pupils as gifted, while the literature cited by the school inspectorate assumes 10–15% (angles 16 and 26). The second is institutions that can hear dissent — schools, universities, companies and public bodies in which the independent judgement produced upstream is not extinguished downstream. The Columbia accident board found that NASA’s culture “had as much to do with this accident as the foam” (angle 20).

This report specifies the design principles, eleven components, the conditions for scaling, a measurement framework, a worked national programme for Czechia, and the risks.

Summary of the main points

1. Build infrastructure, not heroes. Every component can be specified, owned, funded and measured.

2. Seven design principles govern every component: friction with scaffolding, knowledge-rich content, explicit reasoning rules, community first, voice protected, open doors, and measurement of unassisted performance.

3. Argument arenas carry the strongest evidence. Debate, tutorials and peer instruction measurably improve reasoning, most for weaker students.

4. Real-problem studios work only when guided. Authentic problems with fading scaffolds beat both lecturing and unguided discovery.

5. Explicit reasoning programmes work — if delivered with fidelity. Cognitive acceleration, philosophy for children and metacognitive instruction have strong trial effects and documented scale-up failures.

6. Mentors are catalysts. Mentoring works inside dialogue and real problems, with questioning and “wise” feedback, and aims at the mentee’s divergence.

7. Teacher development is the multiplier. Autonomy-supportive teaching is learnable. Training on gifted pupils shifts stereotypes that otherwise persist.

8. Ally networks build the second capacity. Cohorts, mixed ages and visible backing matter more than individual grit.

9. Voice can be trained, but it decays. Bystander and voice programmes show real effects that fade without reinforcement.

10. The talent funnel must be universal. Screening everyone multiplies the talent found; referral finds the already advantaged.

11. Competitions should develop, not only sort. Team rounds, several ladders, feedback and no public low ranks.

12. Adults need the infrastructure too. Organisations must be built to hear dissent, or they waste the judgement schools produce.

13. Scale with engineering discipline. Train facilitators, monitor fidelity, protect dose and measure unassisted outcomes.

14. Czechia has the mandate and the gaps. National strategy already names critical thinking and talent support; the inspectorate documents the implementation gap.

How this report is organised

The report proceeds in six steps:

1. The seven design principles, with the evidence behind each.

2. Eleven components, each set out as an operating brief:

   • In short

   • The evidence

   • Design specification

   • Failure modes

   • What to measure

   • Who builds it

3. Scaling without losing the effect.

4. A measurement framework.

5. A worked national programme for a mid-sized EU state, using Czech data.

6. Equity, risks and first moves.

Library angles are cited as (angle NN).

The seven design principles

1. Friction with scaffolding. Independent judgement grows when a person must produce and defend a position against real resistance: an opposing argument, a problem that pushes back, a demanding question. Resistance without support produces failure, not learning.

• Evidence: active learning cut failure rates from 33.8% under lecturing to 21.8% across 225 studies (Freeman et al. · angle 04). Guidance is what makes inquiry learning effective (d = 0.50; Lazonder & Harmsen · angle 07). Students who replace their own work with AI-generated solutions cover more but understand less (Lehmann et al. · angle 25).

2. Knowledge-rich, always. Reasoning runs on domain knowledge. Generic thinking-skills training shows little far transfer (Sala & Gobet; Tricot & Sweller · angle 02). Every format should be anchored in substantive content: science, history, policy, technology, law.

3. Make the rules of reasoning explicit. Transfer happens when the rules are taught through content that carries them. Graduate training in psychology and medicine, both built on statistical and methodological rule systems, produced large gains in everyday statistical and methodological reasoning; chemistry training did not (Lehman, Lempert & Nisbett · angle 19). Metacognitive instruction, which teaches learners to plan, monitor and evaluate their own thinking, produces durable effects (de Boer et al. · angle 19).

4. Community first. Holding a judgement is socially supported. A visible, active ally transforms resistance to pressure (Asch; Milgram 1965 · angles 10 and 20). A lone, silent example may not: in Burger’s replication, a single refusing confederate who then sat silently did not reduce obedience (angle 20). Every participant should belong to a cohort and have allies who stand with them in public.

5. Protect voice. Criticism of the work must be distinguishable from pressure on the person. Feedback directed at the task helps, feedback directed at the self hurts, and more than a third of feedback interventions made performance worse (Kluger & DeNisi · angle 12). Teams learn when members feel safe to raise doubts (Edmondson · angle 12). Critical feedback lands when paired with high standards and explicit belief in the learner (Yeager et al. · angle 12).

6. Open doors. Access must not depend on who happens to notice a child. Universal screening raised the identification of disadvantaged gifted pupils by 180% without changing standards (Card & Giuliano · angle 14). Exposure — meeting people who do the work — causally raises the chance of entering a field (Bell et al. · angle 14).

7. Measure the mind, unassisted. The outcome is what a person can argue, solve and defend without help. With AI in the loop, assisted and unassisted performance diverge sharply. Students with unguarded AI support performed 17% worse than never-users once it was removed (Bastani et al. · EAA-11). Developers learning with AI understood less (Shen & Tamkin · angle 25).

Component 1 — Argument arenas

In short. Recurring, structured settings in which people must state a position, face a real opposing one, and defend or revise their view before peers:

• competitive debate leagues

• small-group tutorials

• peer-instruction rounds in lectures

• structured academic controversy

• moot hearings on real policy questions

The evidence.

• Reasoning evolved for argument. It performs best in groups whose members argue (Mercier & Sperber; Trouche, Sander & Mercier · angle 04).

• Explanation builds and corrects understanding. Self-explanation improves understanding (Chi et al. · angle 04), and explaining mechanisms punctures illusions of understanding (Rozenblit & Keil · angle 02; Fernbach et al. · angle 04).

• Adolescents improve through structured argument. Three years of structured peer argument developed young adolescents’ counterargument-aware reasoning (Kuhn & Crowell · angle 04).

• Peer instruction works. It roughly doubled normalised conceptual gains in university physics (Crouch & Mazur · angle 04).

• Debate helps the weakest most. Competitive policy debate in Boston raised English scores by 0.13 SD, “68% of a full year” of ninth-grade learning. Gains were concentrated in analytical subskills and largest for previously low-achieving students, with positive effects on graduation and college-going (Schueler & Larned · angle 16).

• The tutorial is the historic exemplar. The Oxford tutorial asks a student to write a position and defend it aloud against a tutor’s questions every week (Palfreyman ed. · angle 16).

Design specification.

1. Weekly argument rounds in groups of 12–16 on content-rich motions. Sides are assigned and switched halfway through, and each round ends with a written reflection on what changed one’s mind.

2. A steelman rule: no rebuttal counts until the speaker has restated the opponent’s strongest argument to the opponent’s satisfaction.

3. Tutorial pairs: a short position paper defended for thirty minutes against a questioner, at least monthly.

4. Peer instruction in every large lecture: answer individually, convince a neighbour, answer again.

5. Evidence packs for every motion, so that argument runs on knowledge rather than rhetoric.

Failure modes.

• Point-scoring rhetoric. Counter it with side-switching, steelmanning and judging on the use of evidence.

• Sacred-value topics on which explanation does not moderate views (Sloman & Vives · angle 04). Use them sparingly and deliberately.

• Slow change in classroom talk. A large English trial of dialogic teaching found about two months’ progress, with slow change in teacher practice (EEF · angle 04). Plan for a multi-year build.

What to measure.

• Argument quality on a blind-scored performance task (claim, evidence, counterargument, rebuttal).

• Actively open-minded thinking on a validated scale.

• Revisions for a stated reason: the share of reflections that record a changed view and why.

Who builds it.

• Schools: peer instruction and argument rounds.

• Universities: tutorials and peer instruction in large courses.

• Civil society: regional debate leagues, following the Boston Debate League model.

• The state: funds coaches and the competition infrastructure.

Component 2 — Real-problem studios

In short. Teams working for weeks on authentic, open problems with a real client: a laboratory question, a municipal data problem, a cyber-security scenario, a product for a real user. Explicit teaching arrives just in time, and the scaffolding fades.

The evidence.

• Authentic problems are a core ingredient. They are one of the three most effective elements of critical-thinking instruction (Abrami et al. · angle 01).

• PBL builds durable capability. Problem-based learning wins on long-term retention, skill development and application, while conventional teaching wins on short-term exam knowledge (Strobel & van Barneveld · angle 07).

• Guidance matters. Guided inquiry outperforms unguided inquiry by d = 0.50 on learning outcomes (Lazonder & Harmsen · angle 07).

• A randomised trial. Project-based Advanced Placement courses raised pass rates by 7.6 percentage points in year one (Saavedra et al. · angle 07).

• Authentic research redirects ambition. 68% of undergraduate researchers reported increased interest in a STEM career (Russell et al. · angle 07). Faculty describe gains in “thinking and working like a scientist” (Hunter, Laursen & Seymour · angle 06).

• Dose predicts accomplishment. Among equally able young people, those who later earned doctorates, patents and tenure had a denser “dose” of enrichment: competitions, advanced courses, mentored research (Wai et al. · angle 14).

Design specification.

1. A problem bank co-owned with universities, employers and public bodies. Every problem has a named client who will critique the result.

2. Six- to ten-week sprints: framing → evidence → prototype or position → client critique → revision → public defence.

3. Just-in-time mini-lessons on the concepts the team needs.

4. Scaffolding that fades across successive sprints.

5. An incubation pause — device-free time — between framing and solving (Sio & Ormerod · angle 09).

Failure modes.

• Unguided discovery. It fails novices (Kirschner et al. · XL). Inquiry teaching is not automatically better for knowledge (Jerrim, Oliver & Sims · angle 07).

• Output over learning. Treating the product as the goal; the hackathon literature is almost entirely about outputs (Medina Angarita & Nolte · angle 07).

• Narrowness. An analysis of 15,963 capture-the-flag cyber-security challenge write-ups found a concentration on cryptography and networking and neglect of human aspects (Švábenský et al. · angle 05).

What to measure.

• Unassisted transfer to a new problem of the same structure.

• Quality of the public defence, judged by the client and an independent expert.

• Intentions to pursue research or advanced study, at twelve months.

Who builds it.

• Universities and research institutes: problems and mentors.

• Employers: problems and clients.

• Schools: time in the timetable.

• The state: funds the brokering.

Component 3 — Explicit reasoning programmes, delivered with fidelity

In short. Structured programmes that teach learners to reflect on and regulate their own reasoning:

• cognitive acceleration in science and mathematics

• philosophical inquiry with children

• metacognitive and self-regulation strategy instruction

• statistical and methodological reasoning courses

Each comes with the training and quality assurance needed to preserve its effects.

The evidence.

• Thinking-skills programmes. Effect sizes of 0.62 on cognitive and curricular outcomes (Higgins et al. · angle 19).

• Cognitive Acceleration through Science Education. In its original schools, the share of pupils achieving grade C or above in science at 16 rose from a national average of 44% to 63%, with similar relative gains in mathematics and smaller but substantial gains in English (Shayer 1999 · angle 19).

• Philosophy for/with Children. A 2025 meta-analysis of 30 studies reports an overall effect of 0.65, and 0.89 on critical thinking (Kilby · angle 19). The first English efficacy trial found about two additional months’ progress in reading and maths, largest for disadvantaged pupils (EEF 2015 · angle 19).

• Metacognitive strategy instruction. Effects grew from g = 0.50 at post-test to 0.63 at follow-up, with low-SES students benefiting most (de Boer et al. · angle 19).

• Instrumental Enrichment. Mixed results (Shiell · angle 19).

Failure modes — the scale-up record.

• P4C at scale. The large EEF effectiveness trial of Philosophy for Children, run from 2017 to 2019, found “no evidence that P4C had an impact on reading outcomes” or on mathematics, for disadvantaged pupils or the whole cohort (Lord et al. · angle 19).

• CASE replicated. The EEF trial of Let’s Think Secondary Science found “no evidence” of improved science attainment. Pupils did worse than controls on English and maths, though the authors caution this could be chance (Hanley et al. · angle 19).

• Why programmes fade at scale. The recurrent reasons are thin facilitator training, reduced dose, loss of the developer’s quality control, and time squeezed out by other priorities.

Design specification.

1. Choose programmes with long-run evidence, and adopt them in full rather than as borrowed activities.

2. Train facilitators to a certified standard, with in-class coaching in the first year.

3. Protect dose: a fixed weekly slot for at least two school years.

4. Monitor fidelity through lesson observation against the programme’s core practices.

5. Scale in waves, evaluating each wave, and stop or redesign when effects vanish.

What to measure.

• Reasoning tasks from the programme’s validated instruments, plus general reasoning measures.

• Curricular attainment two to three years later.

• Fidelity scores.

Who builds it.

• The ministry and the national curriculum institute: adoption, training standards and time in the timetable.

• Universities: facilitator training and evaluation.

• The inspectorate: fidelity and outcomes.

Component 4 — Mentor and tutor corps

In short. A trained body of mentors — university students, researchers, practitioners, alumni — whose task is to teach people how to ask rather than what to think, to give critical feedback that lands, and to push the mentee towards a question of their own.

The evidence.

• A catalyst, not a stand-alone. Mentoring alone “did not generate especially strong results” in the critical-thinking meta-analysis, but it lifted the effect of combined authentic problems and dialogue from 0.32 to 0.57. It works “in a catalytic capacity” (Abrami et al. · angle 01).

• Average effects are small. Youth and workplace mentoring show real but modest effects, larger when mentors provide intentional teaching and advocacy (Eby et al.; DuBois · angle 06).

• Harm is possible. A randomised mentoring programme followed for thirty years left participants worse off than controls on several outcomes (McCord · angle 06).

• Questioning is teachable. Teaching students to generate questions produced median effects of 0.36 on standardised tests and 0.86 on experimenter-made tests (Rosenshine et al. · angle 06). Achievement tracks the quality of students’ questions, and one-to-one settings multiply how many questions students ask (Graesser & Person · angle 06).

• Wise feedback transforms uptake. Critical feedback prefaced with “I have very high expectations and I know that you can reach them” led an estimated 71% of students to revise, against 17% with a neutral note (Yeager et al. · angle 12).

• Divergence is the goal. Scientists’ protégés become stars when they do not simply continue their mentors’ topics (Ma, Mukherjee & Uzzi · angle 08). Moderate divergence works best (Xing et al. · angle 08).

Design specification.

1. Twelve hours of mentor training before any match: question stems, wise feedback, autonomy-supportive language, safeguarding, and how to hand a mentee on.

2. Mentoring embedded in argument arenas and studios, never as a free-floating relationship.

3. A “question first” protocol: the mentee writes a question and a best guess before receiving help.

4. A divergence milestone: within six months the mentee defines a question the mentor does not already work on.

5. Several mentors over time, to enable synthesis (Liénard et al. · angle 08).

6. Supervision for mentors, with a harm-reporting channel.

Failure modes.

• Friendly but aimless relationships. Mentoring without an agenda.

• Rigid, curriculum-driven mentoring.

• Concentrating at-risk young people together, a suspected mechanism of harm in the thirty-year follow-up.

• Mentor cloning: the mentee becomes a junior copy.

What to measure.

• Question quality, on a rubric.

• Revision after critique.

• Independence milestones.

• Harm indicators.

Who builds it.

• Universities: recruit and train, and credit mentoring in doctoral and faculty careers.

• Employers: release practitioners.

• Civil society: coordinates.

Component 5 — Teacher development for thinking

In short. A sustained programme that equips teachers to run argument, inquiry and reasoning activities, to support pupils’ autonomy, and to recognise and stretch able pupils. It includes a network for the teachers who already do this against the grain.

The evidence.

• Teachers matter for decades. Individual teachers affect college-going and earnings (Chetty, Friedman & Rockoff · angle 08).

• Tests miss much of it. Teacher effects on complex reasoning and mindset are only weakly related to test-score effects (Kraft; Jackson · angle 08).

• Autonomy support is learnable. Across 51 interventions it produced broad benefits (Reeve & Cheon · angle 08).

• Stereotypes about gifted pupils are strong. Teachers-in-training hold a “disharmony” stereotype of gifted children — clever but maladjusted — that a counter-stereotypical cue did not dislodge (Weyns et al.; Matheis et al. · angle 22).

• Training shifts them. Teachers’ attitudes are significantly more positive where schools engage in gifted education and where teachers have in-service training (Lassig · angle 22).

• Czechia has a training gap. Only 4% of basic-school teachers and 3% of secondary teachers had attended any training on gifted pupils in the previous two years (ČŠI 2022 · angle 26).

• The practice gap persists. The inspectorate reports that the “traditional model of teaching persists” despite its recommendations (ČŠI 2024/25 · angle 16).

• Autonomy is not the obstacle. Czech teachers’ formal classroom autonomy is close to the OECD average (TALIS 2024 · angle 16).

Design specification.

1. A two-year teaching-thinking programme covering peer instruction, structured argument, question generation, metacognitive routines and guided inquiry. Delivery combines in-class coaching with lesson study.

2. Gifted-education training for every school, with at least one certified coordinator per school.

3. A fellowship for “the one teacher”: found through pupil nominations (”who made you think?”), given time, a community of practice and resources.

4. Recognition for teachers whose pupils surpass them. This makes being outgrown a professional honour rather than a threat.

Failure modes.

• Workload. Programmes added on top of overloaded timetables.

• One-off training days without follow-up.

• Treating the gifted as a niche rather than a school-wide responsibility.

What to measure.

• Pupil-reported encouragement of original answers, using the PISA item for national comparability.

• Lesson-observation indicators of cognitive activation.

• Teacher attitudes towards able pupils.

• Fellowship retention.

Who builds it.

• The national teacher-development institute and universities of education.

• The inspectorate: indicators.

• School leaders: protected time.

Component 6 — Ally networks and belonging

In short. Deliberately built communities — stable cohorts, teams, near-peer mentors and alumni circles — so that nobody holds an unpopular judgement alone, and ambition is admired rather than punished.

The evidence.

• Active allies free people. Two defiant peers freed 36 of 40 participants from destructive obedience (Milgram 1965 · angle 20). One truthful partner cut Asch-style conformity to a quarter (angle 10).

• A silent refuser did nothing. A lone confederate who refused and then sat silently had no significant effect (Burger 2009 · angle 20).

• Support buffers stress, and resilience is built from ordinary relationships (Cohen & Wills; Masten · angle 13).

• Belonging changes outcomes. A one-hour belonging intervention halved a minority achievement gap over three years (Walton & Cohen 2011). At scale, it improved first-year completion only where the setting afforded belonging (Walton et al. 2023 · angle 13).

• The local norm decides what is punished. In one school culture students hid effort; in another they hid low ability, because there ability was admired (Bursztyn, Egorov & Jensen · angle 11).

• An older member helps. Adding one slightly older person to a group of young adults brought risky decisions back to the level of individuals deciding alone (Silva et al. · angle 10).

• Benign envy motivates. Envy that drives improvement arises when success looks attainable (van de Ven et al. · angle 22).