Why curriculum metadata now shapes publisher advantage in education AI
Education AI is moving past model access. The operational question now resides within the content estate. Can a curriculum portfolio be segmented, described, connected, reviewed, and served to AI systems without losing editorial intent, assessment validity, rights control, or source provenance? For large education publishers and digital learning providers, this is a portfolio value question. Prior investment in pedagogy, assessment, accessibility, rights management, and efficacy work carries forward once assets are structured as a machine-readable learning layer.
Digital Promise's June 2026 scan of education AI public goods found more than 1,500 candidate data-bearing records across repositories. It also reported basic usability gaps, including inconsistent licensing, uneven documentation, weak artifact-type tagging, and 61% of records missing grade-span metadata. The implication for commercial publishers is clear. AI readiness depends on metadata quality as much as content volume.
Integra has worked inside this gap through production-scale content engineering programmes with large education publishers. Those programmes show a repeatable pattern. The structural work required for reliable digital delivery overlaps with the work AI systems need for grounded retrieval, tutoring, feedback, and governance. Content granularity, structured metadata, and traversable learning outcomes are features of mature content operations, not AI add-ons.
This whitepaper defines the Learning Asset Graph as a practical model for converting trusted publisher portfolios into AI-ready learning infrastructure. Education knowledge graphs are not new. The narrower claim here is that publisher portfolios need a production-grade graph of instructional assets, rights, standards, relationships, and evidence if AI features are to scale beyond isolated demonstrations.
This whitepaper is for
Digital delivery was only the starting point
Before the current GenAI wave, publishers and EdTech organizations mainly approached digital curriculum as a delivery challenge. Content had to flow into platforms. Chapters became screens. Instructor resources became downloadable files. Item banks moved into test engines. Standards maps lived in spreadsheets or product databases. The focus was on access, packaging, navigation, interoperability, and measurement.
That phase created a useful foundation. It also left many assets organised around product containers rather than instructional meaning. A page in a digital lesson may include a concept explanation, a worked example, a misconception warning, two practice items, a diagram, an accessibility note, and teacher guidance. To a learner or teacher, the page is coherent. To an AI system, it may be an undifferentiated chunk unless its pieces are separated, typed, tagged, and connected.
Applied AI changes the unit of value. A learning assistant must retrieve the right explanation rather than the nearest page. A teacher co-pilot must generate a scaffold that respects the standard, grade level, and instructional sequence. An assessment tool must keep formative checks separate from summative claims. A remediation engine must distinguish an arithmetic slip from a misconception. These capabilities depend on content granularity and rich pedagogical metadata.
| Earlier digital curriculum model | AI-ready curriculum model |
|---|---|
| Assets packaged as chapters, modules, pages, screens, and item banks | Assets represented as granular learning objects, including objectives, explanations, examples, items, hints, misconceptions, rubrics, and evidence records |
| Metadata supports discovery, LMS integration, standards tagging, and reporting | Metadata operates during retrieval, tutoring, personalisation, generation control, rights checks, and evaluation |
| Platform logic decides the user journey | AI systems need structured context to reason about instructional intent and next best action |
| Quality assurance focuses on editorial accuracy, display, accessibility, and platform behavior | Quality assurance adds source grounding, retrieval fit, AI-use permissions, hallucination risk, and evaluation sets |
| Assets often remain tied to a specific courseware or CMS model | Assets can be packaged as portable, governed knowledge bundles usable by multiple AI and platform systems |
For publishers, the goal is not to discard the pre-2022 digital estate. The goal is to expose more of the instructional structure already embedded in it. Editorial teams have written scope and sequence. Authors have designed explanations and examples. Assessment teams have created distractors and rubrics. Learning science teams have studied efficacy. Much of that knowledge is present, although it is often implicit, scattered, or locked inside platform-specific schemas.
The hidden asset gap in publisher portfolios
Large publishers hold a material advantage in applied AI. Their portfolios combine trusted content, author networks, peer review practices, editorial standards, rights processes, accessibility workflows, subject-matter expertise, assessment design, and outcome alignment. Those capabilities matter more as AI tools move into classrooms and courseware.
The gap usually appears between content quality and AI usability. A portfolio may be instructionally strong while still being difficult for AI systems to use safely. Legacy production formats, edition drift, inconsistent metadata, platform-specific identifiers, incomplete standards mappings, and weak provenance can limit what AI systems can do with high-quality material.
| Asset gap | How it shows up | Why it matters for applied AI |
|---|---|---|
| Granularity gap | Content is stored as pages, chapters, PDFs, or large HTML blocks | Retrieval may return too much context, miss the instructional unit, or mix teacher-only and student-facing content |
| Metadata gap | Grade, objective, difficulty, cognitive demand, reading level, and item purpose are inconsistent | AI systems cannot reliably select, adapt, or evaluate content for the right learner, task, and instructional context |
| Relationship gap | Prerequisites, downstream skills, misconceptions, and remediation assets are implicit | Tutors and recommendation systems struggle to choose the next instructional move |
| Assessment gap | Items are mapped to topics or standards, but distractor rationale, rubric logic, and misconception mapping are incomplete | AI-generated feedback and practice can become generic, poorly aligned, or difficult to defend |
| Rights and provenance gap | Usage permissions and source versions are separated from the content chunks used by AI | AI systems may reuse content outside approved boundaries or cite the wrong version |
| Evidence gap | Content usage, item performance, and learning outcomes are tracked in separate systems | Publishers cannot easily connect AI behavior to learning evidence and product improvement |
What the work revealed
Integra has delivered content engineering programmes that encountered every gap in the table above. Working with a large UK education publisher over three academic years, the team enriched more than 224,000 metadata rows across GCSE assessment archives, tagging content at worksheet, question, sub-question, and nested question levels against assessment objectives, skills, and learning outcomes.
A parallel programme restructured 238 assessments into more than 1,400 atomic question items. Multi-part questions had to be split before they could support item-level analysis. Learning outcome mappings had to be rebuilt against a new standardised framework before downstream systems could traverse skill relationships consistently. Metadata fields for grade, skill, and objective were inconsistent across the archive and required SME-reviewed enrichment before AI-enabled features could use them reliably.
These programmes were commissioned for content delivery, not AI readiness. Their outputs nevertheless match the assets required by AI-enabled search, tutoring, item analysis, and evaluation. That makes the Learning Asset Graph a grounded extension of production work, not a speculative architecture.
A portfolio audit should answer five questions
- Which assets are usable for AI grounding without additional cleanup, segmentation, or rights review?
- Which metadata fields are present, reliable, missing, duplicated, or platform-specific?
- Which instructional relationships are explicit, and which remain inside prose, PDFs, tables, or author notes?
- Which assessment assets include enough metadata to support tutoring, feedback, remediation, and mastery inference?
- Which evaluation artifacts can test whether AI systems retrieve, cite, explain, and generate within approved educational boundaries?
Metadata as runtime intelligence
Education metadata has a long history. The sector did not begin this work with ChatGPT. Metadata standards and interoperability specifications have supported discovery, alignment, exchange, and analytics for more than a decade. Applied AI makes that foundation more valuable and raises the required level of detail.
LRMI provides vocabulary for describing educational resources. CASE gives standards, competencies, and learning outcomes consistent digital identifiers. QTI enables assessment items, tests, and results data to move across assessment systems. Caliper Analytics defines learning-event profiles for activities such as reading, assessment, media use, grading, and search.
| Standard | Original function | Applied AI relevance |
|---|---|---|
| LRMI / Schema.org | Describe educational resources for discovery, markup, exchange, and resource description | Provides a base vocabulary for resource type, audience, educational level, and alignment that AI systems can use in retrieval and filtering |
| CASE | Exchange academic standards, competencies, and learning outcomes using consistent identifiers | Allows AI tutors, item generators, recommendation systems, and analytics layers to reference the same objective or standard |
| QTI | Exchange item and test content, item banks, tests, results, and assessment data across systems | Supports assessment-aware AI, including item selection, practice generation, scoring logic, feedback boundaries, and interoperability with existing banks |
| Caliper Analytics | Describe learning activities with consistent event profiles | Connects content interactions, practice attempts, feedback, reading, grading, and mastery signals to the learning asset graph |
Catalog metadata helps people find content. Runtime metadata helps systems act on content. In applied AI, metadata performs six operational jobs. It narrows retrieval, selects context, constrains generation, controls rights, guides personalisation, and supports evaluation.
What research tells us about AI, tutoring, and structure
Research does not support the loose claim that any AI chatbot improves learning. The defensible reading is more specific. Structured, instructionally aligned systems can improve outcomes when they are designed around pedagogy, content, assessment, learner state, and implementation quality.
A major intelligent tutoring systems meta-analysis reviewed 50 controlled evaluations and reported a median effect of 0.66 standard deviations, moving the median student roughly from the 50th to the 75th percentile. A Harvard physics RCT compared a custom AI tutor with an active-learning class format; students using the AI tutor achieved better learning outcomes in less time. Khan Academy found that Khanmigo improved next-item correctness when it received structured summaries of a learner's recent problem-solving history.
| Evidence stream | Finding | Design implication for publishers |
|---|---|---|
| Intelligent tutoring systems | Structured tutoring can produce substantial learning gains, with outcomes shaped by alignment and implementation quality | Preserve learning objectives, scope and sequence, assessment alignment, and instructional intent as metadata and relationships |
| GenAI tutoring RCTs | AI tutors can outperform some instructional conditions in bounded contexts when designed around pedagogy and course content | Avoid generic chatbot deployment. Build tutor context from curated, course-specific assets and evaluation sets |
| Khanmigo experiments | Structured learner signals improved next-item correctness; raw data was less useful | Convert learner history, skill gaps, and content relationships into clear signals AI systems can use |
| Metadata-aware RAG | Metadata can improve retrieval when content is repetitive or highly structured | Treat curriculum metadata as a retrieval feature, not a decorative field |
| Knowledge graph approaches | Structured domain knowledge can help ground tutoring and improve adaptation | Represent prerequisites, misconceptions, objectives, standards, and remediation assets as explicit graph relationships |
The practical lesson is consistent across the evidence. AI-enabled learning depends on the quality of the instructional context that reaches the model. Strong metadata and graph relationships do not guarantee learning gains by themselves. They give product teams the substrate needed to build, test, govern, and improve applied AI systems.
How leading publishers are responding
Recent publisher product pages and announcements show a clear pattern. AI features are being positioned around proprietary content, platform integration, course alignment, and learner support. That pattern supports the argument while also showing the unresolved operational need. Product teams need a reusable content intelligence layer beneath individual interfaces.
| Organization | Current AI positioning | Signal for the Learning Asset Graph |
|---|---|---|
| Pearson | AI-powered study tools presented as grounded in reliable Pearson content, course-aligned for learners, and integrated into platforms such as MyLab, Mastering, Revel, and Pearson+ | Trusted content and course alignment are core to the value proposition. The next question is how reusable the underlying content intelligence layer is across use cases |
| McGraw Hill | AI Reader is positioned inside Connect and McGraw Hill GO, with explanations and study support tied to McGraw Hill instructional content and course materials | The publisher content base becomes the grounding layer for AI reading, explanation, and self-quizzing |
| Cengage Group | AI products span more than 100 higher education products and more than one million students, with assistants embedded in course homepages, eBooks, and assignments | AI is being embedded into learning workflows; graph-level metadata can make these assistants more portable and assessment-aware |
| HMH | Instruction-aligned AI connected to curriculum, assessment data, secure infrastructure, and its Ed platform | K-12 AI value is tied to instructional coherence, standards alignment, and data signals rather than standalone tool adoption |
These examples do not prove that every publisher needs the same architecture. They show why the architecture question now matters. Grounded AI features require source boundaries, standards links, assessment context, rights control, and product portability. A Learning Asset Graph gives those requirements a shared operating model.
The AI-Ready Learning Asset Graph
Knowledge graphs in education have a long research history. Work on intelligent tutoring systems, knowledge spaces, and educational knowledge graphs has been active for decades. Adaptive learning platforms have used curriculum graph models commercially since the 2010s. The Learning Asset Graph is grounded in that tradition. The term is used here for the production layer large publishers need now. It focuses on how existing assets are structured, which metadata fields are attainable at scale, how rights and provenance travel with each asset, and how AI-ready packages can be built from publisher-owned content rather than rebuilt from scratch.
The Learning Asset Graph is a proposed content infrastructure model for education publishers and digital learning providers. It represents curriculum as a network of typed instructional assets with metadata, relationships, rights, provenance, and evidence. The model can be implemented in several technical stacks, including graph databases, search indexes, vector databases with metadata, OKF-compatible knowledge bundles, CMS extensions, data warehouses, or hybrid architectures, allowing publishers to adopt the model within their existing technology ecosystem.
Core asset types
| Asset type | Typical metadata | AI use case |
|---|---|---|
| Learning objective | Standard ID, grade band, skill statement, prerequisite list, mastery criteria | Anchor tutoring, recommendations, item selection, and progress analytics |
| Concept explanation | Topic, objective, reading level, approved source span, examples, misconceptions | Ground AI explanations and course-aligned Q&A |
| Worked example | Problem type, solution steps, strategy, difficulty, prerequisite concepts | Support step-by-step tutoring and scaffolded feedback |
| Assessment item | Objective, item type, difficulty, answer key, distractor rationale, rubric, scoring rule | Generate practice, diagnose misconceptions, and provide formative feedback |
| Misconception | Linked concept, error pattern, diagnostic cues, remediation assets | Detect learner errors and select targeted remediation |
| Rubric | Criteria, performance levels, scoring guidance, aligned task | Guide AI-assisted feedback and human review |
| Media asset | Modality, alt text, transcript, rights, accessibility status, source | Support multimodal tutoring and accessible alternatives |
| Evidence record | Usage signal, item performance, learner-event data, efficacy claim, population | Connect content use to product improvement and outcome analysis |
Relationship types
| Relationship | Meaning |
|---|---|
teaches | A lesson, explanation, example, or activity supports a learning objective |
assesses | An item, task, or rubric measures a learning objective or standard |
requires | One concept or objective depends on another as prerequisite knowledge |
remediates | A resource addresses a misconception, skill gap, or error pattern |
misconceptionOf | An error pattern is associated with a concept or objective |
alignedTo | An asset maps to a standard, competency, or curriculum framework |
supportsAccommodation | An asset supports a learner accommodation or accessibility pathway |
hasEvidence | A content asset links to usage, assessment, efficacy, or review evidence |
The five-layer model
| Layer | Function | Example output |
|---|---|---|
| Source-of-truth layer | Preserve canonical publisher assets and versions | XML, ePub, InDesign exports, assessment banks, teacher guides, videos, standards maps |
| Segmentation and compilation layer | Create granular units and extract instructional meaning | Objective pages, concept pages, item records, misconception records, rubric records |
| Metadata and relationship layer | Apply education-specific fields and typed relationships | Grade, subject, standard ID, difficulty, prerequisite, item purpose, rights, provenance |
| AI-ready packaging layer | Prepare assets for retrieval, graph traversal, and AI runtime use | OKF-compatible bundles, JSONL, graph exports, vector indexes, evaluation sets |
| Runtime and evidence layer | Power AI products and feed learning signals back into the graph | Tutor responses, teacher co-pilot outputs, remediation recommendations, analytics dashboards |
LLM Wiki, OKF, and portable learning knowledge bundles
Recent work in AI knowledge representation gives useful language for the Learning Asset Graph. Andrej Karpathy's LLM Wiki proposes a pattern in which LLMs maintain a persistent, cross-linked knowledge base over source material. Google's Open Knowledge Format formalizes this into portable knowledge bundles made of Markdown files with YAML frontmatter. These bundles are readable by humans, parseable by agents, diffable in version control, and portable across tools.
Example OKF-style learning asset record
| Generic OKF capability | Education-specific requirement from the Learning Asset Graph |
|---|---|
| Markdown file with YAML frontmatter | Typed learning asset with grade, subject, standard, objective, difficulty, source, and rights metadata |
| Links among documents | Typed relationships such as requires, assesses, teaches, remediates, and misconceptionOf |
| Human-readable knowledge bundle | Editorially reviewed, source-grounded curriculum bundle with review status and versioning |
| Portable across agents and tools | Usable by AI tutors, teacher assistants, semantic search, assessment workflows, analytics, and content operations |
| Extensible metadata keys | Compatibility with LRMI, CASE, QTI, Caliper, accessibility metadata, and publisher-specific taxonomies |
From pilot to production
A common failure mode in education AI projects is the attractive demo that cannot be scaled across the portfolio. A single title can be manually cleaned and prompted into a prototype. Production needs a repeatable operating model covering schema design, segmentation rules, extraction methods, human review, QA gates, rights checks, version control, evaluation, and integration.
The move from pilot to production should be staged. Each stage creates an asset that can be inspected by editorial, product, data, engineering, and compliance teams.
| Stage | Work performed | Production output |
|---|---|---|
| Stage 1 Portfolio inventory | Identify source formats, product systems, asset types, metadata fields, rights constraints, and known quality issues | AI-readiness inventory and risk register |
| Stage 2 Schema design | Define required fields, controlled vocabularies, relationship types, governance rules, and mappings to existing standards | Publisher-specific learning asset schema |
| Stage 3 Segmentation | Split chapters, lessons, assessments, media, and teacher guides into instructionally meaningful units | Granular content units with stable IDs |
| Stage 4 Metadata enrichment | Extract, normalize, and review grade, objective, standard, difficulty, modality, rights, source, and accessibility fields | Machine-actionable metadata layer |
| Stage 5 Relationship mapping | Connect objectives, concepts, standards, prerequisites, items, rubrics, misconceptions, remediation, and evidence | Learning Asset Graph export |
| Stage 6 AI-ready packaging | Prepare indexed, source-grounded bundles for RAG, graph traversal, OKF-style use, and API integration | Corpus packs, OKF-compatible bundles, JSONL, graph files, vector-ready chunks |
| Stage 7 Evaluation and QA | Test retrieval accuracy, source citation, generation boundaries, instructional appropriateness, and rights behavior | Evaluation set, scorecard, editorial QA report |
| Stage 8 Scale roadmap | Define production roles, tools, throughput targets, QA sampling, governance, and release cadence | Portfolio rollout plan |
Quality gates for AI-ready content
Instructional fit
The content unit accurately supports the stated objective, grade band, and learner need.
Retrieval fit
The unit is neither too broad nor too narrow for AI grounding and can be retrieved by meaningful queries.
Assessment fit
Items include enough metadata to support feedback, remediation, scoring boundaries, and mastery interpretation.
Rights fit
The asset has an explicit AI-use policy and source boundary.
Accessibility fit
Media includes alt text, transcripts, captions, or approved alternatives where required.
Governance fit
Source, version, review status, and change history remain connected to the asset.
The RFP playbook for AI-ready curriculum
Large publishers and digital learning providers should treat AI-ready content as a procurement category rather than a contained internal experiment. The right partner should understand curriculum, metadata, assessment, accessibility, rights, AI retrieval, editorial QA, and scale. The RFP should make those capabilities visible and testable.
Questions buyers should ask
- How will you segment curriculum assets into instructionally meaningful units without losing editorial intent?
- Which metadata fields are required for tutoring, teacher support, assessment, remediation, search, and governance?
- How will your model connect or map to LRMI, CASE, QTI, Caliper, accessibility metadata, and our internal taxonomy?
- How will you represent prerequisites, misconceptions, remediation assets, and assessment relationships?
- How will you preserve rights, source provenance, version history, and editorial approval status through AI packaging?
- What evaluation set will prove that AI systems retrieve the right assets and stay within approved content boundaries?
- Can the output be delivered in multiple formats, including OKF-compatible bundles, JSONL, graph exports, and platform-specific imports?
- What QA workflow will combine automation, subject-matter review, editorial review, and sampling at production scale?
- How will learner-event data and item performance be linked back to content improvement cycles?
- What operating model supports portfolio-level rollout across grades, subjects, editions, and products?
A practical pilot shape
| Pilot component | Suggested scope | Decision value |
|---|---|---|
| Corpus sample | One unit, chapter, course module, or assessment bank with representative asset types | Shows real content constraints and avoids abstract demonstrations |
| Schema and graph sample | Define 20 to 40 metadata fields and 6 to 10 relationship types for the selected corpus | Tests whether the model fits product, editorial, and AI runtime needs |
| OKF-compatible bundle | Create a small bundle of Markdown files with YAML metadata and links | Demonstrates portability and reviewability |
| RAG or tutor evaluation | Create 50 to 100 test prompts with expected retrieval behavior and approved answer boundaries | Shows whether the AI system can act within curriculum constraints |
| QA report | Review source accuracy, instructional fit, rights, accessibility, metadata consistency, and retrieval fit | Gives procurement and product teams a basis for scale decisions |
| Scale estimate | Estimate throughput, staffing, tool needs, QA sampling, and delivery cadence for a full series or product line | Turns the pilot into an RFP-ready production plan |
The commercial opportunity is specific. Publishers do not need to rebuild their curriculum from scratch to participate in applied AI. They need to make their best assets computable, governed, and connected. The Learning Asset Graph provides the model for that conversion.
The Learning Asset Graph in practice
The asset types and relationship types defined in the Learning Asset Graph become concrete when traced through a student's actual difficulty. The two scenarios below show what changes when a content estate carries typed relationships and misconception records, compared with one that does not.
Priya is studying for her GCSE Chemistry exam. Her school uses an AI-assisted practice tool built on content that a publisher has structured as a Learning Asset Graph. She works through a question set on calculating the concentration of a diluted solution.
The original multi-part question has been split into three atomic items. Each carries its own metadata, learning objective link, and misconception records created by chemistry editors during the enrichment stage. Priya gets the calculation step wrong. She has added the two solution concentrations rather than applied the dilution formula. This is a documented error pattern. Students who have recently worked with quantities in other contexts may carry additive logic into dilution problems where it does not apply.
The misconceptionOf record on this item names that error pattern and links it via the remediates relationship to a worked example. The worked example contrasts additive mixing (which applies to amounts) with dilution (which applies to concentration). Before serving it, the system checks the requires relationship and asks whether Priya has demonstrated mastery of molar mass calculations, the prerequisite this topic depends on. Her recent evidence records confirm she has. The gap is located at the dilution step, not below it.
Priya reviews the worked example and attempts a parallel item. She gets it right. The outcome is stored as an evidence record connected to both the item and the misconception. In a content estate without this structure, the platform would have returned the nearest page about concentration. That page would cover material Priya already understands alongside the material she needs. The misconception that caused the error would go unnamed.
Priya receives instruction that targets the specific wrong mental model she holds. Her teacher's diagnostic view shows that she missed a dilution question and triggered the additive-concentration misconception twice before resolving it. That is an actionable signal. A score alone would not show the cause of the error.
James is a first-year History undergraduate preparing an essay on the causes of the First World War. His university's AI writing support tool draws on a course content base structured as a Learning Asset Graph.
His draft argues that the assassination of Archduke Franz Ferdinand caused the war. The claim is factually grounded but insufficient for the learning objective his essay is assessed against, which requires a multi-causal argument distinguishing long-run structural causes from short-run triggering events.
The essay rubric is stored as a machine-readable asset, linked via the assesses relationship to the learning objective for historical argument construction. One criterion requires the student to identify multiple causal factors and evaluate their relative significance. The tool compares James's draft against this criterion. His draft does not address the alliance system, imperial competition, or militarism as structural causes.
The tool retrieves the concept explanation connected via the teaches relationship to the learning objective. The explanation describes how historians distinguish structural from contingent causes, with two contrasting examples. Its reading level is calibrated at first-year undergraduate. The feedback James receives identifies the rubric criterion his draft does not meet and grounds it in that concept explanation. It does not rewrite his argument. In a content estate without this structure, the tool would have retrieved a broad section on WWI causes. James would have received generic feedback to include more factors.
James revises. His tutor reviews an evidence trail showing the first draft, the criteria the tool identified as unmet, the concept explanation it retrieved, and the revision. The tutor evaluates the quality of how James addressed the gap, not whether a gap existed.
James's revision demonstrates multi-causal reasoning, not additional facts. The rubric shaped his learning at the point of drafting as well as the point of grading. His tutor receives a structured evidence record that informs marking and the design of the next assignment. The assessment has functioned as instruction.
What's next?
For large publishers, the practical work is now clearer. Trusted content only becomes useful to AI systems when its instructional structure is explicit and governed. Large publishers already hold much of the required material, including curated explanations, carefully designed examples, standards maps, item banks, rubrics, teacher guidance, accessibility work, and outcome research. The work ahead is to expose the structure inside those assets and package it so AI systems can use it responsibly.
The Learning Asset Graph gives that work a concrete shape. It connects content units, metadata, standards, assessments, prerequisites, misconceptions, evidence, rights, and provenance. It can be expressed through existing standards and emerging formats. It supports current product needs while preparing curriculum for more agentic learning workflows.
The content engineering work documented in Integra's case study demonstrates this model in production. A programme commissioned as structured content delivery produced the granular, metadata-rich foundation the Learning Asset Graph describes. That equivalence was observed in the work, not asserted in advance. See how Integra approaches this work.
LLM Wiki shows how knowledge can be maintained. OKF shows how knowledge can travel across tools. The Learning Asset Graph applies that pattern to curriculum, so AI systems can act within the boundaries publishers define.
Plan an AI-readiness programme with Integra →Further reading
The following sources were reviewed for this whitepaper and are useful for readers who want to examine the standards, research, product direction, and knowledge-format patterns referenced.
| Source | URL |
|---|---|
| Digital Promise, Mapping the Public Goods Behind Education AI | digitalpromise.org |
| Dublin Core, About LRMI | dublincore.org/about/lrmi |
| Schema.org LearningResource | schema.org/LearningResource |
| 1EdTech CASE | 1edtech.org/standards/case |
| 1EdTech QTI | 1edtech.org/standards/qti |
| 1EdTech Caliper Analytics | 1edtech.org/standards/caliper |
| ERIC, Effectiveness of Intelligent Tutoring Systems | eric.ed.gov/?id=EJ1090502 |
| Scientific Reports, AI tutoring outperforms in-class active learning | nature.com/articles/s41598-025-97652-6 |
| Khan Academy, Building a Better AI Tutor | blog.khanacademy.org |
| arXiv, Utilizing Metadata for Better RAG | arxiv.org/html/2601.11863 |
| arXiv, Adaptive AI Tutor Using KG-RAG | arxiv.org/abs/2311.17696 |
| Pearson AI-Powered Study Tools | pearson.com |
| McGraw Hill AI Reader | mheducation.com |
| Cengage Group, Scales AI Products Across Portfolio | cengagegroup.com |
| HMH Instruction-Aligned AI in Classrooms | hmhco.com |
| Karpathy LLM Wiki | gist.github.com/karpathy |
| Google Cloud, Introducing Open Knowledge Format | cloud.google.com |
| GitHub OKF Specification | github.com/GoogleCloudPlatform |
| IntegraNXT Content Engineering for AI | integranxt.com/content-engineering-for-ai |