Formal and technical document
aiLearning Challenge
Formal and technical dossier · Fall Edition 2026
This is the public edition of the dossier. The document is the same one that is delivered as a PDF: a single source of text, so that the two can never say different things.
Who it is for. For the whole education sector: public and private schools, universities, ministries and education-policy bodies, organizations that accompany education, allies and sponsors. It grew out of the questions raised by the Ministry of Education of the Dominican Republic —Department of Educational Informatics—, and those questions turned out to be the questions of anyone who has to decide whether this enters their school: what it is, what holds it up, how it runs, how it is assessed and what protects minors. That is why the document is published for everyone and not for one addressee.
Issued by the Learning Innovation Lab of aiLearning.global, an initiative of CEMI.ai dedicated to the adoption of artificial intelligence in learning.
Scope of this document. It describes what the aiLearning Challenge is, what it is founded on, how its complete process runs, how it is assessed, what safeguards accompany it and how it connects with the formative areas of the Secondary Level. It is written for technical teams: it says what the design claims, it says what holds up each claim, and it says —with the same care— what the design does not claim. Where a fact belongs to a specific edition and is not yet confirmed, the document declares it unconfirmed instead of omitting it.
Why now
At a moment when AI is transforming education —with the risk of narrowing students'
cognitive patterns and the development of their abilities—, the aiLearning Challenge sets out to
demonstrate that AI can be used without externalization or cognitive offloading, with creativity and
critical thinking, reaching remarkable results that were previously unimaginable while building at the
same time skills relevant to the future on solid foundations.
1. What it is
A Challenge exists to demonstrate one single thing, live and in front of witnesses: used with judgement, artificial intelligence strengthens learning instead of replacing it — it widens creativity, demands more thinking and not less, and develops cognitive competencies that stay with the person when the tool is switched off. We do not merely claim it: we put it to the test.
The aiLearning Challenge is the first event instance of the Challenges modality of CEMI's Smoother canon: a live competition in creativity, innovation and problem solving with artificial intelligence, for students from primary school through university. The modality is the model —what makes a challenge an instrument of learning and not a contest, how a challenge brief is written so that a generic AI answer cannot win it, how it is verified that the work belongs to whoever presents it, and how it is judged without collapsing learning into a grade—. The event is an instance of that model and never the model itself.
Two distinctions matter from the start. A Challenge is not a hackathon: a hackathon centres on the product of a single day, whereas the Challenge is a cycle —preparation, event, transfer— whose assessment weight sits in the process, and whose final question is not "what did you build?" but "what are you now capable of doing, even with the tool switched off?". And a Challenge is not a structural container: it is a modality delivered inside a Learning Program or a Learning Experience, with its Learning Units, its Learning Object and its rubric.
The principle that orders the whole design is that AI is the instrument, never the author. The modality does not prohibit AI —a prohibition is unenforceable and contradicts the premise— but builds the room so that the machines do not lift the weight for anyone.
2. Methodology, Process and Event
Almost everybody's first question —the Ministry of Education put it exactly this way, and school directorates and deaneries repeat it— is whether the aiLearning Challenge is a methodology, a process or an event. The answer is all three, and each one is held up by something different. This is Carlos Miranda Levy's answer, rendered without rewriting it into something more comfortable.
| What it is | What holds it up | |
|---|---|---|
| Methodology | "We have one, based on facts and with a sound and solid pedagogical foundation." | The modality canon: the Learning Object, the analytic rubric, the tool standard, the safeguard rules and the foundation published at ailearning.global/challenge/foundation/. It is a design position with a documented evidence base, never a "validated pedagogy" nor a "proven method". |
| Process | "It is not a one-day competition: we work with schools, educators, teams and students to prepare them and to learn to use AI without externalization or cognitive offloading and with creativity, critical thinking, reaching competencies and results significantly better than without it." | The preparation of the mentor educators with its own verification gate; the Challenge Gym sessions; selection and advancement; and the AI-free transfer task after the event. |
| Event | "A finely detailed and structured event on the day it happens." | The live day in phases: sealed brief, the first minutes without AI, the Divergence Gate, the Process Gate, the perturbation delivered simultaneously, the Mentor Marketplace and the Live Defense. |
The three readings are not three descriptions of the same object: they are three layers that need one another. The methodology without the process is a document; the process without the event has no live moment to authenticate authorship; and the event without the other two would be exactly the contest the modality was written not to be. The extended explanation, with its own public page, is at ailearning.global/challenge/metodologia-proceso-evento/.
3. Pedagogical foundation
3.1 What it rests on
The modality applies Smoother's assessment canon without variants: a Learning Object with its declared components, an analytic rubric built from that Object, a canonical four-level scale of achievement and a multi-component profile as output. Its architecture converges with evidence the canon already documents —desirable difficulties, deliberate practice, scaffolding in the Zone of Proximal Development, the documented failure of project work when the adult running it is unprepared, polish bias as a documented failure of judging—. The word is converges, and never "is derived from".
Its most demanding clause is the accountability one: if a piece of work can be produced without thinking and still earn the grade, what has been built is an assessment that rewards coasting. A Challenge is designed, in its entirety, against that sentence. The six mechanisms —the Initial Human Window, the Divergence Gate, the Process Gate, the perturbation, the mentor who asks and never builds, the Live Defense and the AI-free transfer task— are hypotheses about how to avoid that outcome, and the model treats and measures them as such.
3.2 What it does NOT claim
This section exists because a technical team has the right to know where the edge of what is claimed lies.
- It is not a validated pedagogy nor a proven method. It is a declared design position, with an explicit evidential label: original / design position. It is not derived from research and is not proven by it.
- It promises no results. No magnitude, no percentage of improvement, no before-and-after performance comparison is claimed on any Challenge surface.
- It claims no accreditation, endorsement, approval, validation or certification by anyone, for anyone — not of the Challenge, nor of a site, nor of a mentor, nor of a participant. Mapped to never means endorsed by. Where an allied institution holds an accreditation of its own, that is its claim, in its voice.
- It claims no detection. None of its safeguards is presented as technical verification of authorship; what runs is supervision, record-keeping and accountability (§7).
- It invokes no types of intelligence nor any taxonomy of capacities as a foundation.
- Observation of the rehearsals is observation of practice. The in-person rehearsal of August 2026 with teams of educators produced findings that were incorporated into the canon; its only honest record is "in the rehearsal we observed", never "the evidence demonstrates".
4. The complete process
A Challenge is a cycle, not a date. Its five phases map onto CEMI's Impact Arc —Engage → Enable → Inspire → Empower → Connect—, created by Carlos Miranda Levy; its label is also original / design position, never evidence-backed pedagogy.
| Phase | What happens | The failure mode it watches |
|---|---|---|
| Engage | Forming the field before a topic exists: sites, communities listened to and not merely summoned, verified mentors, registration with consent resolved. | A global framing lands unchanged at every site and the teams solve a problem nobody near them has. |
| Enable | Capacity before the clock: mentor educator preparation, mentor preparation, Challenge Gym, tool audit, measurement baselines. | The teams meet AI for the first time on the clock; the adults improvise; a tool that is free in the plan turns out to be paid on the day of the event. |
| Inspire | The live Challenge: reveal, creation under authentic constraint, perturbation, demonstration and defense. | Single-shot optimization; the slide-deck event. |
| Empower | Ownership transfers in full: absolute intellectual property of the participant, their own decision whether to publish, AI-free transfer task. | Trophies are handed out and nothing is measured; someone claims rights over what minors created. |
| Connect | Cross-category showcase, structured reflection, persistent community, honest report of the edition. | The edition ends at the award ceremony and the next one starts from zero. |
4.1 Preparing the educators
Mentor educator preparation is not a recommendation: it is a verified condition, with a gate of its own, and it is the first of all the gates to fire. It fires before a site runs its first Gym session —not before the edition—, because the Gym is the first thing a mentor educator actually does.
- Condition. Every adult who is going to run a Gym session, verify a gate or hold a follow-up meeting has completed the preparation programme and has signed the mentor boundary commitment.
- Verification. It is verified by the local lead, never by the person concerned. The evidence artifact is a preparation roster per site: one row per adult, with name, date of completion, signed commitment and passed comprehension check.
- Remediation, not exclusion. Anyone not on the roster may supervise for safety and nothing more: they do not run a Gym, do not verify a gate, do not hold a follow-up meeting. A failed comprehension check means running the boundary Learning Unit again, with no limit on repetitions.
- Required prerequisites almost nil, deliberately. Someone who has never opened a generative tool is admitted. A programme that presumes fluency prepares only those who already have it and leaves the least-resourced site without provision, which is the opposite of what it exists for.
- The backbone is doing, not reading. The educators run a real Challenge as a team: someone who has never been inside the experience cannot hold its gates. On top of that sit a short pass/retain comprehension check, restricted to the non-negotiables where a wrong answer causes harm; three role simulations; a supervised Gym session with a safeguard variation injected without stopping the clock; and a site conditions statement, in which the person writes, for their own site, what must be true beyond the free tools for the standard to produce equity, and which of those things is not true today and what they will do about it.
- Verification and veto are separated. The local lead verifies and reports; the central organizing team holds the veto over whether a site with insufficient provision opens, runs reduced or defers. The reason travels with the rule: whoever has invested months in opening a site is the worst placed to be asked to close it.
- There is never a young person present at a preparation session.
4.2 The Challenge Gym
A championship is a performance event and on its own does not supply the conditions of practice —repeated focused work, feedback and the opportunity to improve—. The Challenge Gym is the layer that does supply them: short challenges of 20 to 45 minutes, run at the sites before the edition, open and free of charge.
Two conditions define it. It is unscored and non-qualifying —no Gym result has any effect whatsoever on advancement or on eligibility— and its tasks are authored on a reduced set of AI-resistance properties: authentic context, competing values and demonstrable artifact. A short unscored practice task needs no more engineering than that.
What is rehearsed: every signature mechanism at least once, so that none is met cold on the day of the event, and also the gate moments, because a team that arrives cold at the Process Gate is exactly the failure the Gym exists to prevent. The requirement is coverage, never one stage per session. The floor is functional: every mechanism rehearsed at least once, plus the gate moments, and never fewer than three sessions —below three there is no room left for the debrief, and the debrief is where rehearsal becomes learning—.
Inside the Gym, and only there, one exception to the mentor boundary applies: the educator may demonstrate by doing. Outside the Gym, no.
4.3 Selection and advancement
- The process gates are scaffolding. Registration and eligibility, compliance with the tool standard, Divergence, Process and evidence freeze: failing one costs time or requires remediation, never eligibility.
- The only gate that decides advancement is the Live Defense, because it is the only instrument that verifies that the work belongs to whoever presents it. Its outcomes are PASSED, HELD or NOT PASSED; a hold means unresolved doubt about authorship and is re-examined by a second panel before any non-pass is recorded. It is never decided by a single member of the jury. Passing is not a score: it is a threshold, and ordering happens afterwards, on the profile.
- Advancement is decided by nomination, not by score (§6.2).
- Hard rule of the cycle: the judged final is live, in every category and every stream. Reviewing a video authenticates nothing.
4.4 The AI-free transfer task
After the event, each participant solves a new task from the same capability family, with no AI available, paired with their earlier baseline where there is consent. It is not a punishment for using AI and it is not a gate: it runs after all advancement is decided, it can neither add to nor subtract from recognition, and making it consequential would turn a measurement into a threat and corrupt precisely the data it exists to produce. The individual result is returned to the participant and to nobody else.
4.5 Externalization or cognitive offloading — the distinction that orders the process
The distinction is Carlos Miranda Levy's and it is what gives the process a question rather than a slogan: the question is not how much AI is used, but what the participant externalizes and what stays.
To externalize is to hand an instrument an operation the participant already knows how to do or that is not the object of the learning —formatting, transcribing, translating a draft, generating twenty variants to choose among them, looking up a fact that is then verified—. To offload cognitively is to hand over the operation that was the learning: deciding what the problem is, choosing among approaches, judging whether something works, holding a reason. The first frees attention to think higher up; the second is exactly what the design exists to prevent.
The design does not resolve that distinction by exhortation. It resolves it by architecture: the Initial Human Window reserves the framing of the problem for the person before any instrument; the Divergence Gate forces the final direction to be chosen from a space of possibilities rather than accepted from the first plausible answer; the Process Gate forces the way of working to be declared and accounted for; the perturbation invalidates work that was not understood; the Decision Log records what was delegated, what was rejected and why; the Live Defense tests all of it in person; and the AI-free transfer task looks, with the instrument switched off, at what stayed.
5. The day of the event
The live day is a sequence of phases of approximately 45 minutes each, in the same order in which the mechanisms appear in every stream and every category.
- The sealed brief. The challenge is revealed when the clock starts; nobody has it before. It is authored centrally against the AI-resistance properties, locally reframed as a mandatory step —an anti-extractive guard: a framing that lands unchanged produces teams solving a problem nobody near them has— and reviewed before being published. It is deliberately open, with a low floor, a high ceiling and wide walls: you can enter with a drawing and a sentence, and you can leave with a working system.
- The first minutes without AI — the Initial Human Window. The first 10 to 15 minutes after the reveal are AI-free. Each participant records individually —written or drawn, before any group convergence— what they think the real problem is, what is ambiguous, what they assume, what occurs to them and what they need to find out. That record is the first entry of the Decision Log and the jury's human baseline. It protects the framing from the homogenizing pull of a tool's first answer, and it protects the voice of whoever is quietest in a team.
- Three distinct approaches before producing — the Divergence Gate. No team enters production without showing three substantially distinct approaches and articulating what each one optimizes and what it sacrifices. They are not three polished products: they are evidence that the final direction was chosen. On passing it, the team also declares the output format —it leaves the gate knowing what will exist at the end—; a difference of format alone never makes two approaches substantially distinct. It is verified by the mentor educators, not the jury: it is scaffolding and not a score, and failing to pass it costs time, never eligibility.
- The Process Gate. Immediately afterwards, each team declares the process framework it is going to work with: one of those taught, a combination, or its own way named in its own words —"our way: we talk, we build, we check, we argue" is a complete and valid declaration—, with the why: which alternatives were looked at and how the chosen one fits. There the accountability warning is delivered: at the defense the team will present its work in the terms of its own framework and will have to account for any stage it skipped. Changing framework later is adaptation with reasons, never an infraction. Which framework was chosen and with how much fidelity it was followed are never read. What the gate eliminates is not freedom: it is improvising in silence.
- The build, with the Mentor Marketplace open. Mentors from many disciplines are available to all teams through equal consultation tickets, so that access to expertise is distributed by rule and not by familiarity or contacts. No team has an assigned mentor. Choosing which expertise to consult is itself assessed problem solving, and giving up all consultations in order to test with users is legitimate.
- The perturbation, delivered simultaneously. Mid-race, a new condition delivered by the organizer, identical for every team in the category and division, announced live. It is authored with the brief and not on the day. It must be survivable —it stresses the design, it does not destroy the ability to finish— and legible to the rubric. Per-team perturbations and perturbations delivered by mentors are rejected: they reintroduce exactly the inequity the equal-ticket design exists to avoid. Perturbation families available to the author: most of the budget is cut; the most affected person cannot read; evidence appears that contradicts a fact the brief asserted; the solution must work with no network; a stakeholder objects for reasons not considered; a specified disability must be accommodated.
- Mentors who ask and never touch. The boundary is explicit and it is signed. They may: ask, explain a concept, critique, point out a relevant idea, question an assumption, point to a source, and say "I don't think you've asked the right question yet". They may not: operate the team's device, write its prompts, write its code, produce an asset, determine the final concept or edit the submission. Over-assistance is detected with the same instrument as AI over-assistance —the Decision Log and the Live Defense—: a team that cannot account for its own decisions reveals it, whether it came from a person or from a machine, and no separate surveillance is needed. Whoever crosses the boundary does not mentor again.
- Evidence freeze, demonstration and Live Defense, without AI. The defense is the authentication gate and the only one that decides advancement. It is held without AI, in person, testing what is presented against the Decision Log and against the terms of the framework the team itself declared. Its form is calibrated by category —show-and-tell and a friendly conversation in the first, structured defense with published question categories, Socratic questioning, a rigorous viva at university level— and in all of them it verifies the same thing: that the work belongs to whoever presents it.
- Structured reflection, and after the event the AI-free transfer task (§4.4).
6. Assessment
This is the section a technical team will test hardest, so it is stated precisely.
6.1 The output is a profile, never a grade
A Challenge declares its own Learning Object, stable across editions and across categories. The capability in formation is constant —AI-assisted creative problem solving, under human direction—; the edition's topic is the vehicle that varies, and the differences between categories are handled by calibrating the levels of achievement, never by forking the Object.
From there the rubric follows: rows = assessable components of the Object; columns = the canonical four-level scale —Not achieved · Developing · Achieved · Mastered—; output = multi-component profile. Mastered means Mastered for that Category, never in absolute terms.
6.2 The nine rows, and process weighing more than product
| Row | Cluster | Emphasis |
|---|---|---|
| Problem framing and human gaze | Process | 12 |
| Divergence and creative originality | Process | 15 |
| Direction of instruments and collaborators (orchestration and agency) | Process | 15 |
| Iteration, failure literacy and adaptation | Process | 12 |
| Collaboration and transparency of contributions | Process | 6 |
| Demonstrated value and usefulness of the artifact | Product | 15 |
| Fit to Setting — local relevance and feasibility | Product | 10 |
| Quality of communication and defense | Product | 10 |
| Ethical judgement in consequential decisions | Product | 5 |
The process cluster carries 60 of the emphasis and the product cluster 40, and that asymmetry is the event's central claim: a beautiful artifact can hide an empty learning process, and polish bias is a documented failure of judging. A spectacular artifact with no visible human direction loses against a bold, well-directed attempt.
What those numbers are, and what they are not. The weights express emphasis: jury attention, evidence budget and reading order when two profiles have to be compared. They are not coefficients of a total, and no document, record or interface may present a weighted sum. There is no composite number per submission, anywhere in the instrument.
Contents are covered and verified by checklist, and they produce no level of achievement.
The creativity row is judged with the Consensual Assessment Technique —a product is creative to the extent that appropriate observers agree that it is—, applied only to that row, with a jury holding genuine familiarity in the domain of the submission and calibrated by category.
The instrument-direction row additionally reads, as a sub-anchor and not as a new row, the deliberation of the act of adopting the process framework: that a way of working was consciously chosen, with alternatives explored and fit reasoned. High = alternatives explored and the choice reasoned against the problem, whatever the framework, including one of the team's own. Low = the first framework to hand was grabbed without consideration, or none was chosen and the work wandered. Which framework, how much fidelity, and how eloquent the narration was are never read. Honest label: scoring the deliberation of adoption is a design position, not a research result, and its residual risk is said out loud —a feigned exploration is possible—, with the guard of the contrast between the defense and the Decision Log and the bounded character of the act: a declaration, not a continuous narrative that rewards embellishment.
6.3 Aggregation is by vote, never by score
Here the model structurally protects what the canon requires, instead of exhorting it.
- The jury nominates submissions by named award category; the submissions with the most nominations advance.
- Aggregation happens among jury members within an award category, never across categories.
- No submission ever receives a composite number.
- Aggregation from site to global panel is by profile, never by average: the profile's pattern travels with the nomination, and the panel receives the profile, the nominations by category and the defense gate's record — never a rank number.
- No ordering across categories is ever produced.
Seven award categories, each traceable to a row of the rubric, so that judging and awarding share a single instrument: General (integrative, reads the whole profile) · Greatest creativity and divergence · Best human direction of AI · Best iteration and failure literacy · Greatest local relevance · Best collaboration · Best ethical judgement. "Most innovative" and "most creative" are deliberately collapsed into one: they would split the same votes.
The guard against polish bias is structural. Visual finish and interactivity are recognized as mentions outside the advancement path —never judged categories, never nominated, never counted—. An award for visual beauty would institutionalize the bias; an award for interactivity would privilege one medium.
6.4 Operating parameters of judging
- A minimum of three independent raters per submission, who record before discussing.
- A discrepancy of two levels or more on any row triggers a documented reconciliation.
- A minimum of four submissions for a site to run its own round; below that, the site groups regionally instead of being excluded.
- No inter-rater correlation threshold is adopted: that instrument belongs to scoring systems, and borrowing a figure from the literature would be a reliability claim we have not measured on our own instrument. In its place operates the disagreement trigger: where nominations do not overlap, a senior jury member reviews. The agreement reached may be computed and published per edition as descriptive information about the process, and low agreement is a finding to act on, never a result to bury.
- Independence. Nobody judges a team they mentored, taught, hosted or is related to; whoever gave consultations does not judge that round.
6.5 Recognition, and the accreditation firewall
Three layers, and nothing is ever summed: (1) a record of participation for everyone who completes the cycle —recognition, never certification—; (2) distinctions by dimension, taken from the profile, which is the mechanism by which the multi-component profile survives a competitive format instead of collapsing into a ranking; (3) principal recognition by category, decided by jury deliberation over profile patterns in a declared order of priority, with the deliberation recorded, never by arithmetic.
Certification is separate and optional: whoever completes the entire cycle with sufficient rubric evidence may be certified against the Challenge's Learning Object, which produces a profile and never a grade. The three layers of recognition are not certifications and never imply one. The accreditation firewall of §3.2 governs all of the above.
7. The Natural Intelligence mode
Any participant or team may compete without AI, on equal terms with those who use it. It is not a concession to whoever distrusts the tool: it is the control condition that makes the claim of §1 observable. If the Challenge claims that AI used with judgement strengthens learning, then someone has to be able to compete without it and do well. It applies from this Fall Edition onward.
The name and what it names. "Natural Intelligence" was chosen so as not to name the option by a negation. It names a declared mode of participation — never a faculty, never a type of intelligence, never a measured construct, and never an insinuation that whoever uses AI is using something unnatural.
Three modes, with person-level and team-level grain.
| Mode | What it is |
|---|---|
| AI-assisted | A person or team working with AI instruments within the tool standard. |
| Natural Intelligence | A person or team competing without AI. |
| Mixed | A team one or more of whose members declare Natural Intelligence while the rest work with AI. |
What makes the mixed team legible is a single Decision Log per team that says which contribution came from which member and in which mode.
Equal terms, literally. The rubric produces a profile in the same way in all three modes. The row that stood in the way —human-AI direction, with the greatest weight of the process cluster— was rewritten as direction of instruments and collaborators: what the team delegated (to a model, to a source, to a mentor, to a tool, to a teammate), what it rejected and why, read from the same Log. That rewrite is provisional and will be reviewed after the edition that runs it for the first time; the previous state is recorded in full.
Declaration. The mode is declared with the Process Gate's framework declaration and binds during the event. A participant or team that declared Natural Intelligence and turns to AI is reclassified, never disqualified: it moves to AI-assisted, or to Mixed if it was only part of the team; the change is noted in the Log with its moment and its reason, it is recorded by a named person at the site, validated by the local lead, and read as adaptation in the iteration row and as evidence of direction. Disqualification would turn the mode into a trap; silent tolerance would empty it.
Honest safeguards: supervision and accountability, never detection. No artifact of this Challenge claims that non-use of AI is "guaranteed" or "verified", because it is not attainable and a document must not assert what it cannot hold up. What runs, and what is said: supervised room; declared devices; the Decision Log; the live defense; the AI-free transfer task that every participant already takes; and a named person who records any reclassification, with its moment. Those conditions are observed, and that is exactly what is claimed.
Shared awards; a visible mention; never an advancement category. The same award categories for all modes. Natural Intelligence entries that reach the final receive a visible mention, in the sense of the record of participation — never an award row, never a nomination slot in the aggregation by vote.
The choice is the participant's, never the school's nor the sponsor's. No site, school or sponsor assigns the mode. Steering towards Natural Intelligence whoever is poorly connected, without a licence or without equipment, because it is cheaper, would turn a free choice into a sorting mechanism. The provision for those cases is the tool standard of §8 —loaner devices, a low-bandwidth path, local backup—, and the rule that nobody is penalized for an organizer's failure reaches whoever was pushed out of AI by circumstance.
8. Tool equity
The Tool Neutrality Standard is the rolling rule that keeps the field level.
- Only freely and publicly available tools.
- No paid capability in competition work, whatever account anyone personally holds.
- No licensed educational tier as a requirement. It is permitted solely as an optional privacy accommodation that confers no capability advantage whatsoever: requiring it would make participation depend on an institution holding a licence.
- Audit two weeks before the event and on the day of the event, because free tiers change. An audit failure removes the tool from the universe and substitutes for all submissions at once; a submission that used in good faith a tool later removed is not penalized.
- Switching plan across several providers and offline and low-bandwidth backup provided per site.
- Loaner devices, so that operating equipment is never the barrier — and without stigma: the loan and the low-bandwidth path are ordinary site provision, not a concession named over anyone.
- The standard is taught as criteria, not as brands —free, age-eligible, accessible, with sufficient quota—: a list of tools expires; a standard does not.
And what the standard does not achieve on its own is said. Free tools do not produce equity automatically: connectivity, devices, prepared adult time and accompaniment are also needed. That is exactly what the site conditions statement (§4.1) obliges each site to put in writing about itself.
9. Collaboration with Microsoft in the aiLearning Challenge
On 9 November 2026, Carlos Miranda Levy ruled on the operating terms of the open collaboration with Microsoft for the Fall Edition of the aiLearning Challenge (5 December 2026 at Microsoft Headquarters, Blue Mall, Santo Domingo), accompanied by free provision of tools and learning paths available from 15 November 2026.
The collaboration starts from a non-negotiable principle of the Smoother canon: technology is designed around the dynamics and processes of learning, never the other way round. Technological tools serve to deepen the capability, agency and reflection of those who learn, never to replace human judgement nor to dictate the pedagogical sequence. As the Challenge's maxim establishes: "AI is the instrument. The judgement, the creativity and the decisions remain the participants'".
Five-phase pedagogical sequence
Tool integration is structured into five defined moments, ensuring that technology amplifies human reasoning instead of replacing it:
| Phase | Approach | Role of technology and safeguard |
|---|---|---|
| Phase 0: Analyze, Understand and Formulate | Strictly WITHOUT AI | Initial Human Window. Analysis, debate, formulation and structuring 100% human. AI does not intervene. The starting point is set down in the Decision Log. |
| Phase 1: Review, Validate and Improve | AI-assisted | Socratic questions, questioning of assumptions and exploration of alternatives. Leads into the Divergence Gate (three substantially distinct approaches). |
| Phase 2: Create the Work Plan | AI-assisted | Organization of flows, dependencies, roles and deliverables. Connects with the Process Gate (explicit declaration of the working framework). |
| Phase 3: Build the Solution | AI-assisted / spatial | Prototyping, iterative development, spatial construction and live response to the perturbation. Each adaptation is recorded in the Decision Log. |
| Phase 4: Communicate with Knowledge | Rehearsal with biofeedback / live WITHOUT AI | Rehearsal with Speaker Coach in PowerPoint. Culminates in the Live Defense: 3 minutes of oral presentation and jury questions with no AI assistance. |
The six tools provided at no cost (from 15 November)
All confirmed teams and educators have, from 15 November 2026, free access to six tools provided for their preparation and development:
- Educational Copilot Chat + Study and Learn agent: A protected environment for consultation and ideation under educational data safeguards. The Study and Learn agent acts as an interactive learning guide, helping to put hypotheses to the test and review key concepts without supplying fabricated answers.
- Microsoft 365 Copilot (for adult participants and educators): Contextual integration in Word, Excel and PowerPoint under strict privacy isolation for educators, mentors and university participants of legal age.
- Agents in Copilot (process guides and building your own): Three preconfigured accompaniment agents: Challenge guide (rules, phases and schedules), Assessment criteria (analytic rubric and weighting of human direction) and Jury questions (simulation of Socratic defense questions). Participants also have the official Microsoft Learn workshop for developing agents of their own.
- Speaker Coach (in PowerPoint): An objective individual rehearsal tool for the 3-minute Live Defense. It offers private feedback on speaking pace, detection of verbal fillers, eye contact and reduction of dependence on reading slides.
- Minecraft Education (spatial and environmental prototyping): A three-dimensional immersive environment for challenges in sustainability, physical infrastructure and land-use planning.
- Azure for Students Starter (advanced technical teams): Cloud services for hosting prototypes, lightweight databases and repositories. Its use is completely optional and confers no grading advantage in the rubric over conceptual or multimedia solutions.
Tool neutrality and sponsorship safeguard: The aiLearning Challenge remains fully open to any technology and platform (Google Workspace, OpenAI, Anthropic, open-source tools or Natural Intelligence). Microsoft's tools are completely optional, the briefs are written with absolute neutrality, no sponsor obtains intellectual property rights over students' work, and no telemetry or participant data is shared.
Full documentation and explanatory video at ailearning.global/challenge/microsoft/.
10. Minecraft Education Extension
The Minecraft Education Extension is formally incorporated into the Fall Edition at the direct request of participating teams, educators and schools, who were asking for a three-dimensional, spatial medium to model architectural, ecological and physical-infrastructure interventions.
Minecraft Education is an optional mode of expression: it does not replace or diminish the other expressive formats (analytical documents, code, physical models, audiovisual productions or presentations). Its pedagogical value lies in landing conceptual ideation in spatial systems with constraints of scale, ergonomics and physical feasibility.
Areas of application in real challenges
- Sustainability and ecological transition: Renewable energy networks, water basin management, waste treatment and biological corridors.
- Urban mobility and road safety: Traffic flows, safe pedestrian layouts, public transport corridors and cycle lanes.
- Coastal and riverbank resilience: Storm-surge barriers, mangrove restoration and river defenses adapted to Caribbean geography.
- Preservation and historic heritage: Planning and conservation in historic urban fabrics, with particular attention to the Colonial Zone of Santo Domingo.
- School buildings and community schools: Bioclimatic classroom design, universal accessibility and energy efficiency in school facilities.
AI-supported flow for recreating real environments (4-stage pipeline)
So that projects are built on geographic reality and not on generic fantasy landscapes, teams follow a methodical flow supported by open data and AI:
- Stage 1 — Open data collection: Satellite maps, elevation profiles, GIS layers and municipal records of the chosen locality (the school itself, its community, a stretch of coast or the Colonial Zone).
- Stage 2 — AI-assisted processing: Writing prompts in Educational Copilot Chat to convert metric elevations to block scale (1 block = 1 metre) and to generate procedural scripts (in MakeCode or Python) that automate complex topographic curves.
- Stage 3 — Generation of the
.mcworldfile: Conversion of the terrain to Minecraft Education's universal format, ensuring compatibility across Windows, macOS, iPad and Chromebooks without requiring dedicated graphics cards. - Stage 4 — Verification of the prior base canvas: The resulting file contains only the existing condition (topography, road layout and baseline buildings), completely free of the proposed intervention.
Live integrity rule (Fair Play)
Given the level of detail Minecraft allows, the extension imposes a strict integrity rule to guarantee competitive equity:
> Teams MAY bring base worlds prepared in advance (real topographic terrain, road layout or buildings of their school surroundings) to serve as a starting canvas.
> Teams MAY NOT bring pre-built solutions. The specific problem to be solved remains sealed until the morning of the event, and the complete solution must be conceptualized and built live during the day.
Authenticity is verified at five control points: submission and snapshot of the base .mcworld file on arrival (8:30–9:00 AM); Phase 0 with no screens and no AI touched; declaration of medium at the Divergence Gate; recording of every block and script in the Decision Log; and demonstration at the 3-minute Live Defense, where the team operates its world and explains its decisions in response to the jury's questions.
Official preparation paths: Preparation of educators and students is recommended through Microsoft's official programmes: Minecraft AI Foundations and Literacy (education.minecraft.net/discover/ai) and the Hour of AI: The First Night and Generation AI modules.
Full documentation and explanatory video at ailearning.global/challenge/minecraft/.
11. Safeguards and child protection
The safeguard rules are binding constraints, not guidance: where they collide with a design preference, an ally's request, a tool opportunity or an operational convenience, the rule wins, and a site that cannot comply with them does not run the event. They are set as a global floor applied everywhere, plus the stricter local law where it exists, because an organizer that knows the age of its participants carries more responsibility than a platform that infers it from a declaration.
The binding set:
- No participant uses an account of their own on an AI service. In the youngest category the account and the operation belong to the mediating adult, and that is mediation and not authorship: the concept, the decisions and the making remain the participants'. Age eligibility follows each provider's terms in force, is not waivable, is re-verified at every tool audit, and an age is never falsified by anyone or "on behalf of" anyone.
- Zero personal data in AI tools. No person's name and no address of anything goes into a prompt. The briefs are designed so that it is never necessary, and what comes back from local contact with stakeholders is turned into a de-identified finding before entering any tool.
- Written consent from the legal representative for every participant under 18, informed, specific, granular and severable, with assent from the participant themselves in language appropriate to their age. Research consent is severable, refusable and revocable, and invisible to the jury.
- Nothing leaves the room or is published without written family permission, piece by piece. Open publication is encouraged and never required: it is never a condition of entry, of advancement, of judging or of an award. Whoever does not want their piece shown says so and it is not shown. The organizers hold only a narrow and explicitly consented licence to show the work for the purposes of judging, showcasing and reporting on the edition.
- Absolute intellectual property of the participant. They retain full and unencumbered ownership of everything they create. Neither the host organization nor any sponsor may claim any right —no licence as a condition of entry, no assignment, no option, no right of first refusal—; a sponsorship agreement containing such a clause is rejected.
- Only tools from the audited list. Nothing outside it enters competition work, and the list is re-verified at every audit (§8).
- Two-adult rule whenever a participant who is a minor is present; adult-minor communication only through event or school channels; background checks; a signed code of conduct; and a named and published reporting route, including an anonymous one. The supervision ratios by category are read as prepared adults : participants and are all floors. Whoever supervises for safety counts towards the two-adult rule and never towards the supervision ratio.
- The responsible adult acts if the safeguard line is crossed. It is a named role, with authority to stop whatever has to be stopped at the moment; violations are handled proportionately and transparently, with the response recorded.
- Risk disclosure, not prohibition. It is said in full, before a tool is chosen, that on many free tiers what is sent may be used to train the provider's models, that this is outside the organizers' control, and that it may affect what the person later considers publishing or commercializing; and that the copyright status of AI-assisted work is unsettled law in many jurisdictions, on which the event makes no representation whatsoever.
- Compliance with applicable legislation on the protection of children and adolescents and on personal data protection, in the Dominican Republic and in any other jurisdiction where a site runs.
12. Curricular areas
A school needs to be able to explain, to its own directorate and to families, which part of the curriculum what its teams are going to do connects with. The Ministry asked for that map and it serves any school equally. The eight formative areas of the Secondary Level of the Dominican Republic, in the Ministry's own order:
- Spanish Language
- Foreign Languages
- Mathematics
- Social Sciences
- Natural Sciences
- Arts Education
- Physical Education
- Human, Integral and Religious Formation
The Second Cycle adds modalities —Academic, Technical-Professional and Arts—, so the eight areas are not the complete picture.
The link has two faces and it is worth not confusing them. The methodology, the process and the competencies are transversal: the framing of a problem, divergence, the direction of instruments, iteration, collaboration, communication and ethical judgement do not belong to one area and are exercised in any. The challenge topics, by contrast, do anchor: a challenge is written on a real context, and that context falls within one or several formative areas. The complete map of both faces —competencies by area and topics by area— is published, with its detail, at ailearning.global/challenge/areas-curriculares/.
A possibility open to discussion, not a programme that exists: offering one challenge for each major curricular area. Carlos Miranda Levy raised it at the meeting and it is brought here as a proposal for conversation with the Ministry, not as something committed.
What this document does not say: any weekly teaching load per area. The only figure available comes from a presentation superseded by the curricular adaptation in force, which we have not read; therefore it is not stated.
13. The Fall Edition
| Date | Saturday 5 December 2026 |
| Hours | 9:00–16:00 |
| Venue | Microsoft Headquarters, Blue Mall, Santo Domingo, Dominican Republic |
| Levels | Primary · Middle · Secondary · University |
| Categories and localities | To be confirmed. They are stated as unconfirmed instead of being omitted. |
On the size and the nearness of the date, said head-on. The Fall Edition is the first at this scale and its size is deliberate: the format is being put to the test in public instead of being asserted, and a small edition is the one that allows the complete instrument to be run —the gates, the verified preparation, the supervision ratios, the tool audit— without softening any part to make it fit. The date is near, and that too is said: what makes it sustainable is that the preparation is open and free for any school, whether or not it takes part with a team this December — the Gym material, the school kit and the published programmes are available independently of registration. We do not promise here a larger edition with a date nobody has set.
14. Conditions of participation
- One educator per team.
- Two parents or companions per student —per student, not per team—.
- Participation as individuals or as teams of up to five people. The individual division exists in every category and is judged separately, never mixed with teams: comparing one person with a team makes both fairness and the collaboration dimension incoherent at once.
15. What the organizer provides
- Preparation of the mentor educators, with its programme, its material and its verification gate (§4.1).
- Accompaniment and coaching of the teams before, during and after the event, with a Lab counterpart for each school's coordination.
- Organization of the event: venue, logistics, jury, mentors, observation instruments and timing log.
- Support during the day, including technical backup and the provision of the tool standard (§8): loaner devices and a low-bandwidth path.
- Food and drink on the day of the event, for the participants and those accompanying them.
Transport is NOT included. It is said this way, expressly, so that no school and no family takes it for granted.
16. Documentation and transparency
- Extensive public documentation at aiLearning.global, in four languages: Spanish, English, French and Portuguese. The pedagogical foundation, the cycle, the tool standard, the consent and safeguard instruments and the school kit are published for whoever wishes to examine them.
- The honest report of the edition, which expressly includes the results that did not come out. Low agreement among raters, a mechanism that did not work, a site that could not meet a condition: they are published as findings to act on, they are not buried.
- Materials reusable by any programme. The framework, the templates, the challenge generator, the organizing guide and the Gym material remain open and free, for the Ministry and for any other initiative, with or without a relationship to the Challenge.
- The record of decisions, published. Every ruling, its date and what it superseded is at
ailearning.global/challenge/evolution/. It is the instrument with which a third party can audit not only what we claim today, but how we changed our minds and when. - What is unresolved is declared unresolved. The modality canon keeps its own register of open questions, and this dossier declares in §3.2, §6.2 and §7 the pieces that are a design position, provisional or unverifiable. An unverifiable competition recorded as such is better than a false instrument.
17. Contact
business@ailearning.global
Carlos Miranda Levy
Coordinator of Impact, Innovation and Intelligence
CEMI.ai · aiLearning.global · Learning Innovation Lab
All the event information, with videos, is at ailearning.global/challenge/.