Current edition: Version 3.0 · Mark 3 · 26 September 2026. From an article plan to a federated education and civilisation system. This edition governs the interpretation and execution of the retained Mark 1 and Mark 2 registers below. Their numbered proposals remain traceable; their conflicting ownership assignments, numerical targets, feature promises and release rules are superseded by the controls here. Updating this paper does not declare the proposed services built or the whole federation audited.

The purpose: help people become able to do what matters

The World Education White Paper sets out how eduKate will build a connected education resource for learners, parents, teachers and tutors: a place to understand a subject, identify a learning difficulty, choose appropriate support, practise, check progress and use knowledge beyond the original lesson. Its ambition is not simply a larger education website. It is a coherent route from a human question to a capability the person can increasingly carry.

The federation connects curriculum knowledge, learning science, examination preparation, tutor guidance, recovery, research skills and practical life. A family searching for English tuition, a student struggling with Mathematics, a teacher planning a lesson and an adult returning to study should encounter different entrances into the same dependable architecture. They should not have to understand the estate, learn its internal terminology or read an entire textbook before finding the next useful step.

The longer horizon is education and civilisation: how people acquire knowledge, test claims, preserve useful understanding, work together and improve the systems on which ordinary life depends. The direction is Learning → Education → Knowledge → Inquiry → Civilisation. These are connected perspectives, not compulsory stages in a person’s life and not a ranking of cultures. School learning remains valuable in its own right; it also opens into work, art, science, relationships, public reasoning and responsible contribution.

The governing question: What should this person be able to understand, decide or do after using this route; what evidence would support that claim; and where should they go when the route is not enough?

Read by purpose: ownership and existing sites · learning routes · evidence and research · release gates · corrections to the article register · the 48 upgrade proposals · the ten-year build.

1. Keep the ambition; change the unit of progress

An article is a publishing unit. A completed learning route is a service to a person. The two are not interchangeable. A thousand explanations without a way to choose between them can leave a learner lost. A beautiful diagnostic page without a repair can leave a family worried. A practice bank without meaningful feedback can reproduce mistakes. A confident answer without a source can make uncertainty harder to see.

Mark 3 therefore makes the complete route the primary build unit. A route needs a recognisable entrance, a defined job, the right knowledge owner, an activity or decision aid where appropriate, an interpretable check and a next step. Not every route needs a quiz. A factual reference can finish with a clear answer and its limits; an admissions guide can finish with an official action and a cohort warning; a skill lesson normally needs an opportunity to attempt the skill.

Progress has three separate meanings. Coverage asks what we have made available. Usability asks whether people can find and use it. Learning or decision quality asks what changed after use. Page counts, search impressions and time on page can inform the first two. They cannot, by themselves, establish the third. Search is an instrument panel, not the architect.

2. The article numbers are a register, not a claim of completion

The retained Mark 1 register records 762 baseline planning entries and 218 additional proposals: 980 planning entries. Mark 2 expands the additional proposals to 282, giving an arithmetic planning total of 1,044, and separately records 48 upgrade actions. Those 48 actions are not 48 more articles. Mark 1’s 30-hub arrangement and Mark 2’s 36-hub arrangement are historical planning views, not two simultaneously authoritative inventories.

None of these sums is a newly verified count of unique live pages. Baseline entries may include planned material, existing owners, overlaps and cross-references. A proposal can be satisfied by an existing owner and a useful link rather than a new publication. Before reporting a live total, reconcile stable IDs, canonical URLs, content types, publication states and duplicate intents. Count each canonical owner once; report gateways, tools, editions and article proposals separately. Do not count a vocabulary index again simply because another atlas points to it.

Every retained row must acquire one disposition: reuse an existing owner; surgically improve an authorised owner; create a genuinely missing owner; create a distinct application or local edition; or hold pending evidence, competence or scope resolution. A title in a table is not an instruction to manufacture another page.

3. A federation with distinct responsibilities

The public front door and the intellectual centre are different jobs. eduKateSingapore helps people enter and navigate the library. eduKateSG carries system-level explanation, inquiry and the civilisation perspective. Specialist estates preserve disciplinary depth. Sengkang makes learner and tutoring decisions usable; Yishun develops recovery routes; Punggol places relevant decisions in family and local context. No domain receives ownership of every page containing one broad keyword.

Correction to the earlier map: eduKateSG already has a Civilisation Atlas: the Atlas of Human Capability. It must not be described as the site without an atlas, nor should this white paper commission a competing replacement. Its conceptual models remain identified as eduKate’s models rather than universal scientific classifications.

Protected tutoring architecture: The Tutor System connects parents, students and tutors. The Tutor Handbook remains the technical professional textbook. National tutor discovery belongs to its discovery owners; tutor classification and wider system theory connect to eduKateSG; English and Mathematics decisions descend to their subject specialists. N10 must help readers find this architecture, not publish a second generic handbook.

The ownership rule is one precise intent → one canonical owner → many purposeful entrances. Record the audience, task, depth, jurisdiction and edition before declaring two pages duplicates. A concept explanation, parent decision guide, technical tutor chapter and local implementation can legitimately coexist when each does different work. Editorial ownership is also different from a search-engine canonical tag: do not point distinct pages at one another with canonical markup merely because they share a subject.

The latest explicit owner decision and verified existing owner record take precedence over an unverified proposal in an old table. Conflicts become review items. They are not resolved through automatic deletion, re-slugging or mass rewriting. An upgrade to this document grants no permission to alter another site’s homepage, theme, menu or protected owner.

4. One library, several useful maps

Seven wings can remain a convenient browsing view, but a single tree cannot express every educational need. The same mathematical relationship can belong to a school course, a recovery sequence, a budgeting task and an inquiry project. These are views of related knowledge, not reasons to create four near-identical explanations.

Maintain five connected views: knowledge by concept and discipline; learner need by current difficulty or next ambition; person and job by student, parent, teacher, tutor or adult learner; course and qualification by authority, subject level and cohort; and context by language, place, resources and practical constraints. School years are useful entrances, not permanent descriptions of a person’s capability.

The atlas explains the territory. An index helps locate an item. A hub helps choose a route. A canonical owner teaches or resolves its particular job. A governance ledger records ownership, status and change. They may link closely, but should not pretend to be the same object. Public routes should use ordinary language; the underlying registry can carry the technical precision.

Links must state a relationship: prerequisite, explanation, practice, diagnostic distinction, evidence, alternative representation, application, local edition, professional handoff or return. Begin by finding what already exists. Add a bridge only when it reduces a real gap. A deeper connection should return the learner to the original task with something usable. Hidden needs may deserve visible diagnostic bridges; hidden links for search engines, cloaking and indiscriminate federation-wide link blocks do not.

5. The minimum complete learning route

A route begins by making a small contract with the reader. State who it is for, what it will help them do, what it assumes, what evidence it needs and what lies outside its scope. Then let the reader take the shortest useful path. A student who needs one distinction should not be forced through a complete syllabus. A tutor who needs the underlying mechanism should be able to descend into a substantial reference treatment.

The learning loop is orient → attempt or inspect evidence → teach and model → practise → use feedback → check a fresh performance → return after time or in a changed context → maintain, advance, repair or seek appropriate help. This is a design pattern, not a rigid lesson script. A beginner may need explicit teaching before a productive attempt is possible. An advanced learner may need a difficult problem and expert feedback rather than another introductory explanation.

Teaching conditions and evidence conditions must be distinguishable. Notes, examples, AI, a parent, a peer and a tutor can all contribute to learning. Record their relevant contribution when interpreting an attempt. Correct work produced with extensive support is not worthless; it simply supports a different claim from fresh independent work. Independence means owning the intended thinking, not refusing legitimate tools, accessibility accommodations or human help.

Do not turn five questions into a universal mastery certificate. A reading task, proof, oral explanation, scientific model and personal statement require different evidence. Keep some tasks fresh for later checking; use answer explanations or rubrics that expose why a response works; separate ordinary practice from high-stakes assessment. When a learner fails a check, the route must say what to inspect next rather than merely displaying a lower score.

6. Every handoff needs a return

A useful handoff states four things: why this destination is relevant, what the reader should take there, what they should bring back, and what would count as enough. Without that return contract, cross-linking can become a tour of the estate rather than educational progress.

Worked route A: the Mathematics question that may also be a language question

A student says, “I can calculate, but I cannot do word problems.” Begin with a recent independent attempt, not a permanent label. Can the learner explain what the quantities represent? Can they express the relationship when the language is clarified? Can they calculate after the relationship is supplied? These observations separate plausible educational difficulties; they are not a clinical diagnosis.

If the relation itself is missing, the route descends into the relevant Mathematics owner. If sentence interpretation or vocabulary is blocking an otherwise available relationship, a language owner can help. Sengkang’s learner-state layer coordinates the distinction; BTT or the established Mathematics owner teaches the mathematics; SETC or the established language owner handles the language. The student then returns to a fresh problem where the original cue has changed. The receipt is a better independent interpretation and solution, not three page visits.

Worked route B: a family choosing and then reviewing tuition

A discovery guide helps a family identify plausible support by subject, stage and practical constraints. The Tutor System explains what parents, students and tutors each contribute. The Tutor Handbook supplies the professional decision: fit, first lesson, working hypothesis, support, review, handover or release. The subject owner then provides the disciplinary teaching.

The return is not another enrolment prompt. It is a decision supported by work: continue, change the route, adjust the support, seek another professional, reduce regular tuition or stop an intervention whose job is complete. A learner who becomes able to carry more is a successful outcome even when that means less dependence on the service.

Worked route C: a school concept becoming a civilisation question

Consider a proposed inquiry about water use in a school or household. The Science owner explains the relevant physical processes. Mathematics helps distinguish quantities, rates, uncertainty and costs. English helps the learner describe observations and build an evidence-based explanation. Research guidance helps separate a source, a measurement, an assumption and a claim. The civilisation layer asks how shared systems are maintained, how knowledge is passed on and what responsibilities accompany a proposed change.

The learner returns with a modest, inspectable product: a clearly scoped question, an appropriately gathered dataset, a justified calculation, a source-aware explanation and a proposal that states its limitations. They do not need to solve a city to learn how a system works. This is an illustrative build pattern, not a claim that a particular project pack, experiment or measured outcome has already been published.

7. Depth and access must coexist

Major canonical owners may warrant 20,000 or more substantive words when the job genuinely requires definitions, mechanisms, worked examples, misconceptions, diagnosis, transfer, practice, evidence and professional reference. That depth is a standard of coverage, not a search-engine trick. A concise router or calculator should not be inflated to the same length. Preserve strong legacy owners unless a specific rewrite is authorised.

Give deep resources several reading levels of commitment rather than pretending one reading level fits everybody: the immediate answer; a short route for the present decision; worked examples; then the full technical treatment. Use meaningful headings and internal anchors. A parent should understand what the tutor is doing without learning the tutor’s full operating system. A student should be able to leave after completing the relevant task.

Original examples, careful distinctions and well-tested explanations create value. Repeated introductions, keyword substitutions, padded FAQs and multiple location names attached to the same advice do not. Resident illustrative students may create continuity within an established series, but invented examples must not be presented as real case outcomes or testimonials.

8. Research informs the design; it does not certify our results

The research layer distinguishes official rules, empirical findings, evidence syntheses, design frameworks, expert interpretation and eduKate’s own hypotheses. Each has a different use. A framework can clarify a purpose without proving an intervention works. A study in a different population can suggest a design without predicting an individual student’s result. A local before-and-after improvement does not establish that the website caused it.

The IES/What Works Clearinghouse guide, Organizing Instruction and Study to Improve Student Learning (2007), supports practices including spaced learning, retrieval through quizzing and alternating worked examples with problem solving, while assigning different evidence levels to its recommendations. It supports a carefully designed practice-and-return architecture; it does not justify a blanket claim that mixed practice always beats every other form of instruction.

The OECD Learning Compass 2030 is an aspirational framework connecting agency with knowledge, skills, attitudes, values and wellbeing. CAST’s Universal Design for Learning guidelines offer ways to plan for different means of engagement, representation, action and expression. Here they inform learner agency and access. Neither is cited as a trial demonstrating eduKate’s effectiveness, nor as a basis for labelling children with fixed learning styles.

For every substantial empirical claim, record the source, publication date, population, task, outcome and important limitations. Cite the claim where it is used. Where evidence is contested, identify the disagreement. When evidence is weak, say what is uncertain and how a small, ethical local evaluation could reduce it. Do not borrow the prestige of an institution in place of reading what its evidence actually supports.

The federation should sustain a two-way research–practice relationship. A learning difficulty can generate a classroom question; a source can suggest a mechanism; an implementation can reveal whether the method was actually used; a fresh observation can revise the route. Existing Research–Practice Partnerships and Teacher Inquiry owners should carry that deeper explanation rather than a new competing research hub.

9. Measure capability without turning learners into a dashboard

For each evaluated route, specify the intended population and outcome before interpreting the results. Record the baseline task, later task, support conditions, time interval and relevant changes in teaching or context. A correct answer after a hint, a fresh answer after a delay and a transferred performance should remain distinguishable. Different subjects need different rubrics; one universal learner score would conceal important differences.

Reports must show dates, numerators, denominators, missing data and who participated. Voluntary feedback is not automatically representative. Attrition must not disappear from success claims. “Not measured” is an honest state; it is neither zero nor permission to imply success. A small case study should remain a small case study unless a stronger design supports a larger inference.

Safety-critical or decision-critical defects receive their own treatment. A wrong admissions rule, inaccessible submission step or harmful diagnostic inference cannot be averaged away by attractive engagement metrics. Collect the minimum information needed for a legitimate evaluation, and do not publish identifiable student work merely because it illustrates an appealing story.

10. Access, privacy and safeguarding begin with the first release

Accessibility is not a final infrastructure prize. The engineering target is WCAG 2.2 Level AA for the defined product scope, including complete relevant processes. A conformance claim requires evidence against applicable criteria, not an overlay, a favourable automated score or a few accessible pages. Plan for keyboard use, readable structure, reflow, focus, accessible forms, alternative representations and media alternatives from the beginning; test with automated checks and human task review. Publish known barriers rather than claiming the estate already conforms.

Core explanations and ordinary learning routes should remain usable without an account wherever practicable. Offer printable and low-bandwidth access when it serves the task. Do not make video, an AI chat or a paid service the only route to essential understanding. Translation requires review of meaning, examples, cultural context and qualification rules; a translated page is not automatically a valid country edition.

Before accounts, dashboards or personalised AI are released, resolve purpose, data minimisation, access roles, security, retention, deletion and the applicable legal requirements. A statement about parental consent for children under thirteen is not a complete global privacy policy. Singapore’s PDPC guidance on children’s personal data in the digital environment must be read with the relevant PDPA guidance and obligations. Obtain qualified review for the actual service and jurisdictions rather than assuming one age threshold settles every case.

Educational probes identify what may need teaching or further observation. They must not present themselves as clinical diagnosis, formal disability assessment, official placement or guaranteed admission. Wellbeing and inclusion resources require competent review and clear limits. Help routes should be relevant, current and location-appropriate; a generic helpline footer cannot make an otherwise unsafe article responsible.

11. AI is a bounded service, not the architecture’s source of authority

UNESCO’s Guidance for Generative AI in Education and Research calls for human-centred, age-appropriate consideration of pedagogical suitability, privacy and agency. It is guidance, not a declaration that a particular product is safe or lawful. The federation’s proposed AI services must earn permission through their own defined scope and evaluation.

A useful first AI service may be narrow: locate a verified owner, clarify a checked explanation or provide a graduated hint on an approved task. It should distinguish source-grounded explanation from generated inference, show uncertainty and return to the source. It must not invent official rules, publish raw generated articles automatically, diagnose a child, manufacture credentials or turn unvalidated feedback into a formal grade.

“Never give the answer” is not a universal pedagogy. Sometimes a learner needs a worked model. At other times, giving the answer destroys the intended opportunity to think. Define the task purpose first: teaching, practice, diagnosis or verification. During a check of unaided subject reasoning, remove answer-generating help while preserving legitimate access arrangements. Record what assistance was allowed.

Before release, test correctness, source fidelity, misleading confidence, age suitability, scope breaches, prompt injection from retrieved material, privacy and escape to human help. Keep logs proportionate and protected. A concept page about eduKateAI is not proof that an evaluated runtime exists. Each live feature must have its own documented release state and rollback path.

12. Search and metadata must describe the actual resource

Use titles, opening paragraphs, headings and descriptions that naturally identify the real reader, question and value. Research the language people use, but do not make a keyword’s appearance in three paragraphs and an H2 a universal requirement. Preserve established intent and useful URLs. Discoverability should lead to a truthful, useful page rather than a template produced to occupy a query.

Google’s guidance for AI features in Search says there are no special additional optimisations or special schema required for inclusion. That supports a focus on accessible, useful, indexable content and clear internal links—not an invented AI-ranking formula. Crawling for search, use in other AI systems, access permissions and licensing remain distinct questions. Do not promise visibility, citations, rankings or traffic from markup.

Google’s 2025 structured-data simplification retired specific Search features including Course Info and Learning Video. That does not mean every related schema.org vocabulary term is invalid, nor that Course List is the same feature. Use the current Search feature gallery and the individual eligibility requirements when implementing markup. Article, Breadcrumb, Course List, Education Q&A or Math Solver must be considered according to the real content and functionality, not attached indiscriminately to every page.

Human-readable FAQs can still help readers. Their usefulness is separate from rich-result eligibility. A static worked example does not become a functioning solver because a type is declared. A future course is not a completed course because a card exists. Author, reviewer, date and organisation fields must describe real people, work and relationships.

13. A small registry that can survive a large library

The technical record should be simple enough to maintain and precise enough to prevent drift. For each canonical object, keep a stable ID; object type; exact job; audience; owner URL and responsible estate; prerequisites; related practice or resources; return route; applicable language, jurisdiction, course and cohort; source and rights information; review status; and change history. A route, article, tool, dataset and translated edition need distinguishable types.

For learning activities, add the target capability, prompt, allowed support, answer or rubric, common errors and next action. For calculators, add the versioned rule, authoritative source, assumptions, valid input range, boundary tests and limitations. For claims, add the source date and applicability. For sensitive services, keep non-public operational and safeguarding records separate from public content.

These are record requirements, not a claim that a database, public API or learner model has already been deployed. Begin with the least complex maintainable record. An API becomes justified when there is a real reuse need, a rights policy, a version contract, security and support capacity. A public page being free to read does not automatically make its images, text or datasets openly licensed for every reuse.

14. Release gates that distinguish intention from reality

Apply these gates to a bounded route or feature, not only to the entire federation. One complete English route can be tested while another wing is still being mapped. Do not wait for every hub to be full before learning from real use. Equally, do not call a hub operational merely because its headings and empty table exist. A useful minimum hub can route to existing owners immediately; unfinished destinations must be labelled rather than represented as available.

Publication verification, usability testing and learning evaluation are separate receipts. A successful save is not public verification. Public verification is not evidence of learning impact. A checklist is not a clinical validation. No automated status transition may collapse those distinctions.

15. Corrections must travel through the dependency graph

Maintain a second loop alongside the learner’s loop: claim → source → review → release → observe → correct → notify affected owners → verify the repair. A change in a qualification rule may affect an overview, a calculator, an FAQ and a translation. Correcting only the first page leaves the system internally inconsistent.

Review frequency follows risk and change, not one universal annual interval. Stable conceptual explanations can have a different cycle from admissions deadlines, fees, examination formats, safety information and AI product comparisons. A material official change triggers a review even when the scheduled review date is distant. Expired or disputed content should display a limitation or be removed from the active route while a correction is resolved.

Preserve enough revision information to understand what changed, why it changed and which routes were affected. Separate a source correction from an editorial improvement. Do not quietly convert a proposed title into a live success claim, erase an old owner to resolve a spreadsheet conflict, or overwrite another editor’s newer work. Use small reversible changes and check the current version before saving.

16. Binding corrections to proposals 1–282

The original row numbers are retained so earlier work can still be traced. The following instructions govern their next use. They are dispositions and review requirements, not a claim that all proposed destinations have been built or independently checked.

A hold is not abandonment. It protects the reader until the missing evidence, reviewer or ownership decision is available. Where a proposed title embeds an unsupported claim, revise the proposal before publishing; do not preserve a misleading title merely to satisfy an old keyword plan.

17. Reclassifying all 48 Mark 2 upgrades

The 48 proposals remain useful as a backlog, but are no longer a universal checklist to impose on every owner. Some are foundational controls, some are task-specific instruments and some are later services that need their own authority, resources and evaluation.

18. The ten-year horizon: 2026–2036

The horizon is directional, not a promise that a calendar will make a capability real. The federation should become more useful through accumulated, reviewed knowledge and tested routes. It should not become more complicated merely because another maturity level has been named.

2026–2027: make the existing estate coherent

Resolve owners, remove contradictory instructions, repair high-value routes and distinguish live material from proposals. Build a small set of complete English, Mathematics, tutoring and recovery journeys with accessible entrances and genuine checks. Connect existing Science work to inquiry rather than declaring another empty wing complete. Protect the designs and owners already approved. Establish the correction ledger before increasing the publication rate.

2027–2029: extend patterns that have earned their place

Expand subject and life-course coverage where a real unmet job, competent reviewer and maintenance capacity coincide. Develop stronger practice sequences, alternative representations and teacher reference material. Use local and school contexts to make knowledge usable without creating location clones. Compare how routes work for different readers; improve the model before reproducing it across every grade and subject.

2029–2032: connect learning to inquiry and contribution

Grow carefully reviewed project, research, professional-learning and adult-return pathways. Release bounded tools and AI services only when their evidence, access and governance justify them. Publish reusable materials and datasets where rights and safeguards permit. Maintain the distinction between an educational model, a research finding, an operational service and a recognised credential.

2032–2036: make the institution durable and the knowledge portable

Develop locally reviewed world editions, succession of editorial responsibility, stable records, preservation and export. A durable public resource should not depend on one platform, one search feature, one generation of AI or one person’s memory. International growth should increase contextual accuracy rather than erase local knowledge. The civilisation ambition is realised through people being able to investigate, maintain, improve and pass on useful understanding.

19. The next build packets

Packet 1: ownership and truth. Reconcile the existing atlases against this charter, attach a disposition to proposed rows, check high-risk current facts and distinguish live links from plans. Completion means an auditable map for the selected scope, not an inflated estate-wide completion claim.

Packet 2: two complete subject routes. Choose one bounded English need and one bounded Mathematics need from existing owners. Add only genuinely missing explanation, practice, feedback or return components. Test the route from an unfamiliar learner’s entrance to a fresh attempt. Document what is still unmeasured.

Packet 3: the human support connection. Connect national discovery to the Tutor System, its parent and student routes, the technical Tutor Handbook and the subject owners. Verify that a family can understand fit, review evidence and recognise when support should change without reading the entire professional textbook.

Packet 4: one inquiry-to-life demonstration. Use existing Science, Mathematics, language and inquiry owners to build a small project with an inspectable output and a clear return into subject learning. Record the project as proposed until its actual materials and routes are live. Do not create a new master hub merely to name the demonstration.

These packets establish sequence and acceptance criteria. They do not override an existing lane’s explicit authorisation, appoint reviewers who have not agreed, or claim the work has already happened. Select additional proposals by unmet reader need, educational dependency, risk, distinctiveness and capacity—not solely keyword volume.

20. Maintainability is part of educational quality

Every new owner creates a future obligation: source checks, broken-link repair, curriculum updates, accessibility work, correction handling and sometimes translation. Budget for that obligation before adding another series. A backlog without review capacity is not a library; it is accumulated editorial risk.

For each release packet, identify who is responsible for content, subject review, access, technical operation and correction. One person may hold several roles, but the roles must be real. For a high-stakes or specialist resource, an unfilled necessary review role means the release is blocked or its scope must be narrowed. Do not solve the gap by inventing a reviewer biography.

Preserve stable URLs and useful history. Keep content exportable and maintain an appropriate recovery process. Distinguish editorial updates from homepage, navigation and software changes so an approved writing task does not become an uncontrolled redesign. Use each site’s strengths without requiring every estate to contain every capability.

21. What success will mean

A learner can enter without already knowing the diagnosis. A parent can understand a decision without becoming the tutor. A tutor can reach a rigorous technical explanation without losing the subject. A teacher can adapt a resource without guessing its evidence or rights. An adult can return to learning without being forced into a child’s school-year identity. A reader can challenge a claim and find the source, the limitation and a correction route.

The system can admit uncertainty, recognise that another professional or institution owns the next decision, and allow a person to finish. It can show what is live, what is tested, what is still planned and what needs repair. It can connect a school skill to a larger human problem without flattening either into a slogan.

Build the explanation deeply. Make the entrance simple. Preserve the owner. Test the handoff. Return capability to the person. Carry useful knowledge forward.

22. Scope of this revision and source record

Mark 3 reconciles the paper with the federation’s established responsibilities, the existing Civilisation Atlas and the live Tutor System, and incorporates the primary-source guidance linked at the relevant claims above. Research was checked on 26 September 2026. Source publication dates and the applicability of official rules remain distinct from that check date. The IES guide, OECD and CAST frameworks, W3C standard, UNESCO guidance, Google documentation, PDPC guidance and MOE admissions information have different authority and are not interchangeable evidence.

This release upgrades the white paper and its build instructions. It does not certify every article in the estate, enact a privacy programme, validate a diagnostic, implement structured data, deploy an AI service, establish WCAG conformance, create a partnership or demonstrate a learning effect. Those claims require their own recorded work and verification.

Preserved Mark 1 and Mark 2 planning registers

Historical planning material follows. Retained titles, tables, horizons, the 28-archetype survey, article proposals 1–282 and upgrades U1–U48 preserve the development record. They are not automatically current factual claims, completed services or fresh publication instructions. Where they conflict with Mark 3, use the ownership charter, correction matrix and release gates above. Preserve the useful proposal; verify its facts and owner before acting.

Mark 2: the level upgrades

Added 26 September 2026 as version 2.0. Everything above this line is Mark 1 and stands unchanged.

Mark 1 answered the question every education site must answer first: what to write. Mark 2 answers the harder one: what level to run at. A complete category map is still only a library, and libraries do not become the premier education website on earth by adding shelves. They become it by climbing five levels, each with its own upgrades, its own gate and its own measure. This part lists all forty-eight upgrades, then reports the second sweep of the world map, which found six hubs Mark 1 missed.

Mark 2 changes the totals: 36 hubs (12 owned, 24 new), 282 new articles (218 from Mark 1, 64 from the second sweep), 1,044 articles across the estate, and 48 upgrades on a five-level ladder.

How Mark 2 was scoured

Four lenses, beyond the three Mark 1 used.

Product layers. What the world’s best platforms do beyond publishing: Khan Academy’s mastery practice and its Khanmigo tutor, which since August 2026 draws interactive diagrams for maths and science and keeps teachers in control of what students see; Duolingo’s streaks and spaced review; Quizlet’s flashcards; Brilliant’s interactive problems; Coursera’s certificates; BBC Bitesize’s exam-board filters and revision podcasts. Each is a level above a page.

The 2026 search change. AI answer engines now sit between the site and its reader. AI Overviews appear on roughly two in five Google searches, and Google’s May 2026 guidance says optimising for them is still SEO. What earns a citation is consistent across engines: the answer in the first paragraph, named authors, statistics with named sources, schema markup, fast server-rendered pages and crawler access. Citations are fragmented across ChatGPT, Perplexity, Gemini and AI Overviews, so share of voice across a fixed query set replaces rank as the measure.

Google’s education rich results. Search supports structured data built for learning sites: Practice problems (Quiz markup, at least two problems per concept), Math solver, Education Q&A for flashcards, Course info and Course list, Learning video and Dataset. A premier site should qualify for every one of them.

MOE’s own direction. The Student Learning Space now carries an Adaptive Learning System for Primary 5 to Secondary 2 Mathematics and Upper Secondary Geography, a Short Answer Feedback Assistant for Mathematics and an Authoring Copilot for teachers. MOE keeps the early years print-first, introduces AI progressively, and protects productive struggle so that thinking is not outsourced. A premier site in Singapore matches that stance rather than racing past it.

The re-read. Mark 1’s map was read again by life stage and by reader. The university years had admissions but no academics. The family paying for school had no money hub. The child living the school day had no rules hub. The expatriate family had nothing. The competitor beyond mathematics had nothing. The reader who reads for pleasure had picture books only. Six hubs, listed below.

The five levels

A level is passed, not claimed. Each gate is pass or fail, and the estate reports its level each quarter in this paper’s revision log.

The upgrade register: 48 upgrades

Horizons are the same as Mark 1’s (H1 to March 2027, H2 to September 2027, H3 to March 2028). Every upgrade names one owner and one measure, so it can be scored.

The second sweep: six hubs Mark 1 missed

Numbering continues from Mark 1 (219 to 282). Same rules: one owner, a check-before-build line, a Status cell that becomes the live link.

N19 · University-Level Academics

Owner: eduKateSingapore.com · Horizon: 3 · Check before build: university mathematics rows are checked against Bukit Timah Tutor’s Estate G (entrance and bridging); essay writing against SETC’s academic-writing estate.

N20 · Competitions and Olympiads Beyond Mathematics

Owner: eduKateSingapore.com · Horizon: 2 · Check before build: mathematics olympiads stay with Bukit Timah Tutor (Estate B); computing competitions stay in N4 (row 53).

N21 · Student Finance and Support Schemes

Owner: eduKateSingapore.com · Horizon: 1 · Check before build: scheme rules and amounts come from the official page with a dated check; row 249 is checked against the tuition pages on eduKatePunggol and the fee planner in N13.

N22 · School Life and Rules

Owner: eduKateSingapore.com · Horizon: 2 · Check before build: DSA and Primary 1 registration are checked against the Punggol pathways first; the school calendar against Project Atlas Estate F.

N23 · Reading, Books and Ideas

Owner: Singapore English Tuition Centre · Horizon: 3 · Check before build: examination literature stays in SETC’s reading-and-literature estate; picture books for the under-sevens stay in N6 (row 69). This hub is reading for its own sake.

N24 · International and Expatriate Families in Singapore

Owner: eduKateSingapore.com · Horizon: 2 · Check before build: international schools are in N7 (row 84); Mother Tongue exemption is in N3 (row 42). Admission rules and fees carry the official-source line with a dated check.

Mark 2 build order

Articles. Horizon 1 rises to 106 (N21 joins, because Edusave, the Financial Assistance Scheme and bursaries are searched every posting season). Horizon 2 rises to 132 (N20, N22 and N24 join). Horizon 3 rises to 44 (N19 and N23 join).

Upgrades. Horizon 1 carries the cheap, compounding ones: the Level 1 hygiene (U1 to U6, U9), the first instruments (U12, U13, U15), the diagnostic entry (U20), the policies (U26, U27, U28), the newsletter and entity work (U35, U37), and the two infrastructure moves that cost little and pay early (U39, U43), with the ladder scored from the first quarter (U48). Horizon 2 builds the instruments and the runtime. Horizon 3 builds the institution and the infrastructure.

The one rule that orders everything. No site opens a Level 3 upgrade before it has passed the Level 1 gate. A tutor on top of an unlinked, undated library is a faster way to be wrong.

Mark 2 in numbers

PEW.MAP@2.0
levels          = [Library, Instrument, Runtime, Institution, Infrastructure]
gate_rule       = pass, not claim; scored quarterly in the revision log
hubs            = 36   (owned=12, new=24)
articles        = 1044 (atlases=762, mark1=218, mark2=64)
upgrades        = 48   (L1=9, L2=9, L3=9, L4=11, L5=10)
horizons        = H1 106 articles | H2 132 | H3 44
order_rule      = no Level 3 before Level 1 gate
ai_rule         = print-first for the young; struggle before the tutor; no answer-dumping
citation_rule   = answer first, named author, sourced statistic, schema, crawlable
measure_rule    = one owner, one measure per upgrade

Risk register

• Velocity outruns verification. A thousand articles written faster than they are checked is a thousand liabilities. The official-source line and the freshness register are the brake.
• The tutor that gives answers. An AI that solves the homework destroys the site’s reason to exist. Socratic guardrails and an answer-dump rate are non-negotiable.
• Cannibalisation across sites. Two pages on the same intent split the authority of both. The owner map and the check-before-build lines are the cure, and a quarterly duplicate audit the check.
• Stale seasonal pages. Cut-offs, dates and fees age in months. Pages past their refresh date are counted, not hoped about.
• Health-adjacent overreach. SEN, wellbeing and milestone pages that drift into diagnosis. The evidence gate stays on every one of them.
• Dependence on one engine. AI citations are fragmented across engines and rankings no longer guarantee them. Share of voice across a fixed query set, measured monthly, replaces any single ranking as the health check.

Mark 2 questions

What is a level upgrade?
A change that moves the site from doing one thing to doing the next: from answering to practising, from practising to remembering the learner, from remembering to being cited, from being cited to being built upon. Articles add breadth; upgrades add altitude.

Why levels rather than more articles?
Because the sites that own each slice of the world today are not the ones with the most pages. They are the ones whose pages do something, remember something, or are trusted for something. The map from Mark 1 is necessary; the ladder from Mark 2 is what makes it premier.

What is the first thing to do on Monday?
U1 and U3: open the ten most visited pages on each site, put the answer in the first three lines, and add a dated official-source line. Then publish the editorial policy (U28) and score the ladder for the first time (U48).

Mark 2 sources

• Google Search Central: the structured-data gallery (Practice problems, Math solver, Education Q&A, Course, Learning video, Dataset) and the 2021 launch note on practice-problem and math-solver markup.
• Google’s May 2026 guidance on optimising for generative AI features in Search, and 2026 industry analyses of AI-answer citation patterns (answer-first structure, entity authority, sourced statistics, schema, crawler access, cross-engine fragmentation).
• Google.org and Khan Academy, August 2026: Khanmigo’s interactive diagrams and teacher-controlled practice tool.
• Ministry of Education, Singapore: forum replies of 11 February and 18 September 2026 on AI in schools; Student Learning Space overview and its Adaptive Learning System, Short Answer Feedback Assistant and Authoring Copilot.