<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[The Information Continuity Lab - Medium]]></title>
        <description><![CDATA[Designing Information Continuity from Business Intent to Digital Implementation - Medium]]></description>
        <link>https://medium.com/demand-to-data-knowledge-graphs?source=rss----bcba237793cc---4</link>
        <image>
            <url>https://cdn-images-1.medium.com/proxy/1*TGH72Nnw24QL3iV9IOm4VA.png</url>
            <title>The Information Continuity Lab - Medium</title>
            <link>https://medium.com/demand-to-data-knowledge-graphs?source=rss----bcba237793cc---4</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Wed, 18 Nov 2026 18:28:30 GMT</lastBuildDate>
        <atom:link href="https://medium.com/feed/demand-to-data-knowledge-graphs" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[From Open Data to Working Applications — The Next Step Is Learning From What We Build]]></title>
            <link>https://medium.com/demand-to-data-knowledge-graphs/from-open-data-to-working-applications-the-next-step-is-learning-from-what-we-build-cd3fc5eb6092?source=rss----bcba237793cc---4</link>
            <guid isPermaLink="false">https://medium.com/p/cd3fc5eb6092</guid>
            <category><![CDATA[demand-to-data]]></category>
            <category><![CDATA[claude]]></category>
            <category><![CDATA[openai-codex]]></category>
            <category><![CDATA[decision-making]]></category>
            <category><![CDATA[mrdm]]></category>
            <dc:creator><![CDATA[Dr. Jarkko Moilanen]]></dc:creator>
            <pubDate>Fri, 09 Oct 2026 09:46:54 GMT</pubDate>
            <atom:updated>2026-10-09T10:09:00.911Z</atom:updated>
            <content:encoded><![CDATA[<h4>Stream 1–05</h4><p><em>Open Data Value already connects datasets, business opportunities, and AI-generated implementation packages. The next phase brings implementation results and business-user feedback back into the decisions that started the work.</em></p><h3>We Have Built the First Half of the Process</h3><p>Over the past months, I have been developing <a href="https://create.opendatavalue.com/">Open Data Value</a>, a platform that connects published open data to business opportunities and helps people turn those opportunities into working applications. The original motivation was straightforward. Governments and other organizations publish large quantities of data, but the path from finding a dataset to producing useful business outcomes remains unnecessarily difficult.</p><p>At the time of writing, the platform covers 66 open data portals, with more than 134,000 datasets analyzed, over 1,100 data products, and more than 1,300 potential use cases. These numbers represent a growing collection of opportunities connected to actual published data. The important part is not the size of the catalog, but the relationships between available data, business objectives, use cases, and candidate data products.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*UNNC5lQK1amZYz3VOkthHg.png" /></figure><p>Instead of asking users to start with a dataset and figure out what to do with it, Open Data Value provides business-oriented starting points. A use case describes the intended decision, potential users, relevant datasets, and expected value. The relationships between those elements provide context for evaluating whether an opportunity deserves further attention.</p><p>This was the first part of the work. The next challenge was converting those opportunities into something a software development agent would understand well enough to implement.</p><h3>Value Forge: Preparing the Work Before Coding Starts</h3><p>To address that challenge, I built Value Forge, a multi-agent system that converts use cases into implementation-ready handover packages for coding assistants such as Codex and Claude Code.</p><p>The system does not simply rewrite a use case into a longer prompt. It uses specialist agents to examine the business context, available datasets, feasibility, architecture, user experience, requirements, and verification criteria. Multiple implementation approaches compete, reviewers evaluate them, and the strongest candidate becomes the basis for the final package.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*wbpU41_azoyL3lHHM9pGNA.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*qWFaWxag3gDUL-Z59ohVGg.png" /></figure><p>The output contains structured implementation guidance, data dependencies, requirements, constraints, acceptance criteria, and supporting evidence. The coding assistant receives a much stronger starting point than a short description of what someone wants built.</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2Fgvmj5gJqyJg%3Ffeature%3Doembed&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3Dgvmj5gJqyJg&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2Fgvmj5gJqyJg%2Fhqdefault.jpg&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/c22b4fc7c29a5e7044c29eaceb73ef8f/href">https://medium.com/media/c22b4fc7c29a5e7044c29eaceb73ef8f/href</a></iframe><p>I have already tested this process with Codex and produced working proofs of concept from generated packages. The results have reinforced my view that a coding assistant benefits considerably from having the work prepared before implementation starts.</p><p>Value Forge does not attempt to replace coding agents. It focuses on the earlier stage, where business opportunities become sufficiently clear, evidence-bound implementation instructions.</p><p>The existing workflow is therefore relatively simple: select a use case, generate a package with Value Forge, hand the package to a coding assistant, and build the application.</p><p>That process works. But it also exposes an important limitation. What happens to everything we learn after handing the package over?</p><h3>We Are Not the First to Think About Durable Context</h3><p>The idea of preparing durable specifications for coding agents is not new, and I do not want to present it as something invented in Open Data Value.</p><p>I recently read Adnan Masood’s article, <a href="https://medium.com/@adnanmasood/specifications-as-durable-context-for-coding-agents-e7dbb806f7d2">Specifications as Durable Context for Coding Agents</a>. It examines three approaches to spec-driven development: BMAD, GitHub Spec Kit, and OpenSpec. These approaches share a concern I have encountered in my own work. Coding agents need context that survives individual conversations and remains available throughout implementation.</p><p><a href="https://github.com/bmad-code-org/BMAD-METHOD">BMAD</a> provides structured development methods and specialist roles. <a href="https://github.com/github/spec-kit">GitHub Spec Kit</a> organizes specification, planning, implementation, and verification into a repeatable process. <a href="https://github.com/Fission-AI/OpenSpec">OpenSpec</a> focuses on maintaining specifications and recording changes as software evolves.</p><p>These are different solutions to related problems. Value Forge already applies several of the same principles through its multi-agent preparation process, layered context, machine-readable requirements, deterministic validation, and evidence-based review. There is little value in replacing that architecture with another framework simply because it uses familiar terminology.</p><p>What particularly interests me about OpenSpec is the way changes become explicit records. Instead of allowing an implementation to drift away from its original specification without explanation, a change has a reason, affected requirements, proposed modifications, and a history. That is a useful principle to bring into the next phase of Value Forge.</p><p>Masood also makes an important distinction between describing behavior and enforcing it. A specification might state that an application must satisfy a particular business rule, but the implementation still needs technical constraints and executable tests to verify that behavior. A well-written handover package is not proof that the resulting application is correct.</p><p>These observations help clarify what I want to improve. We already have a process for preparing implementation packages. Now we need stronger mechanisms for preserving those packages, recording what changes during implementation, verifying the result, and returning useful findings to the original business context.</p><p>The direction is related to spec-driven development, but its scope extends beyond software specifications. I want to connect implementation learning back to the business decision that justified the work in the first place.</p><h3>The Missing Connection Between Implementation and Business Decisions</h3><p>Consider a use case aimed at helping city planners identify areas where public services should be expanded. Open Data Value identifies relevant datasets, and Value Forge generates a package describing the proposed application. Codex receives the package and implements a working interface containing maps, comparisons, filters, and recommendations.</p><p>From the implementation perspective, the experiment appears successful. The application runs, the interface behaves as expected, and the tests pass. Yet while developing the application, the coding agent discovers that an important dataset is refreshed only twice a year.</p><p>This changes the feasibility of the original idea. The application might support long-term planning but not operational decisions that require current information. The implementation package might also need stronger instructions about displaying data freshness and avoiding unsupported claims.</p><p>Now suppose business users examine the resulting application. They find the maps attractive and the interface easy to operate, but the application does not help them determine where to allocate next year’s budget. They expected comparisons between service demand, existing capacity, and projected population growth. The implementation provides only current service locations.</p><p>We now have several different findings. One concerns available data. Another concerns the implementation requirements. A third concerns how accurately the original use case described the business decision.</p><p>If these findings remain buried in a coding conversation or repository, the next implementation starts without the benefit of what we learned. The organization risks repeating the same mistakes, even when the application itself was technically successful.</p><h3>Introducing Living Decision Context</h3><p>I am developing a model called <strong>Living Decision Context.</strong> The goal is to maintain an evolving, evidence-backed understanding of a business opportunity while preserving the original use case and the implementation history.</p><p>The model connects four durable objects.</p><ul><li>The <strong>Use Case</strong> represents the original business hypothesis, including the intended decision, users, and expected value.</li><li><strong>Decision Context</strong> maintains the current understanding of relevant data, assumptions, limitations, feasibility, risks, and outcomes.</li><li>The <strong>Build Contract</strong> represents the immutable implementation package generated from a particular context version.</li><li>Finally, the <strong>Build Outcome</strong> captures what happened during implementation and what evidence was produced.</li></ul><p>The original use case should not be rewritten whenever a coding agent reports something new. The generated package should not silently change after handover either. Both need stable identities and traceable histories, while Decision Context evolves through reviewed learning.</p><p><strong>My intention is to make Decision Context part of a connected graph rather than another isolated document. </strong>Business objectives, use cases, datasets, candidate data products, implementations, and findings already have meaningful relationships. New evidence should strengthen those relationships and improve the understanding available for future decisions.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/790/1*pQIADfh9KhzzkFWmGn2CAw.png" /><figcaption>From Idea to Impact. The six-stage Living Decision Context process.</figcaption></figure><p>An implementation becomes more than a delivery result in this model. It becomes an experiment that tests some of the assumptions behind the original business opportunity. The experiment might support those assumptions, expose missing evidence, or reveal that the opportunity needs a different approach.</p><p>The next decision might lead to another implementation package, but it might also narrow the use case, identify a missing dataset, redirect investment, or stop the work altogether.</p><h3>Three Learning Loops With Different Responsibilities</h3><p>An important design decision is to separate three types of learning. Not every implementation finding should return to the business level, and not every business observation requires a software change.</p><p><strong>The implementation loop concerns the software itself.</strong> It asks whether the application functions correctly, follows its technical constraints, and passes the relevant tests. A failing calculation, broken API connection, or missing error state belongs here. These findings guide technical repairs and verification.</p><p><strong>The build-contract loop concerns the quality of the handover package</strong>. It asks whether the package contained sufficient instructions to guide the coding assistant correctly. Missing acceptance criteria, unclear data transformations, weak UX instructions, and ambiguous requirements should lead to proposed contract improvements. This is where an OpenSpec-like change record is particularly valuable. Each accepted amendment should create a new package version with a clear explanation of what changed and why.</p><p><strong>The decision-context loop concerns the business opportunity itself</strong>. It asks what the implementation revealed about the intended decision, expected value, data feasibility, business assumptions, or actual user needs. These findings should become reviewed learning associated with the use case and its wider decision context.</p><p>The same finding might affect more than one loop. Discovering stale data might require a visible freshness indicator in the implementation contract while also changing the business assessment of whether the use case is suitable for operational decisions.</p><p>Keeping these loops separate allows us to improve implementation quality without confusing technical progress with business value. It also avoids automatically generating another build whenever an experiment fails to meet expectations.</p><h3>The Business User Must Be Part of the Learning Process</h3><p>Technical verification is only one part of evaluating a working application. The people who need the solution should also assess whether it meets their expectations and supports the decision they originally wanted to make.</p><p>For that reason, the next phase includes a new Implementation Feedback tab inside the existing Use Case view in Open Data Value. Once a proof of concept is available, users will be able to inspect the implementation, review screenshots or a working preview, and evaluate the result against the original business need.</p><p>The planned feedback model includes ratings for business fit, usability and user experience, support for the intended decision, clarity and trust in the data, and overall usefulness. Users will also be able to leave written comments about what works, what is missing, and what should change.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/920/1*WFF21XV3LD14WxNixFZETQ.png" /><figcaption>Implementation Feedback tab concept showing the POC preview, ratings, and written feedback.</figcaption></figure><p>The purpose is not to produce another generic satisfaction score. Each evaluation should be connected to the exact implementation and package version. We should be able to compare subsequent versions, see which concerns were addressed, and identify problems that continue to affect the user experience.</p><p>A successful build that receives poor business-fit ratings is an important learning outcome. It suggests that passing technical tests and implementing documented requirements was insufficient to satisfy the original business need. The problem might lie in the implementation, the package, or the business assumptions that shaped it.</p><p>User feedback will enter the same governed learning process as technical findings. Ratings and comments represent user observations, not automatically verified business facts. When a finding has implications beyond the implementation, it should be reviewed before becoming part of the evolving Decision Context.</p><p>This closes a gap in the original workflow. A business user identifies a need, Value Forge prepares the implementation package, a coding assistant builds the application, and the user evaluates how well the result satisfies the need. The learning then informs future decisions and, where appropriate, future packages.</p><h3>Evidence Must Accompany the Feedback</h3><p>One principle I want to preserve throughout this development is that AI-generated explanations should not automatically become accepted evidence.</p><p>A coding agent might report that all requirements were implemented. That statement is useful as a submission, but it is not sufficient proof. The platform should receive test results, implementation references, observed data limitations, screenshots, and other supporting artifacts wherever appropriate.</p><p>For user-facing proofs of concept, visual inspection is particularly useful. The feedback process should retain evidence of the primary interface, the main user journey, and relevant limitations or error states. A future automated preview-deployment process would allow the application to be tested in a browser rather than evaluated entirely from source code and generated reports.</p><p>Each Build Outcome should also preserve its relationship to the original package and implementation version. The repository commit, package identifier, contract version, evaluation results, and supporting evidence form an audit trail explaining what was tested and what was observed.</p><p>At the same time, we need to distinguish what each type of evidence demonstrates. A screenshot proves that a particular interface state was rendered. A passing test establishes that a defined condition was satisfied under the test conditions. A positive user rating records an assessment by that user. None of these independently proves that the application delivered measurable business value.</p><p>The review process should consider the strength and meaning of the available evidence before promoting findings into Decision Context or accepting changes to a Build Contract.</p><h3>Closing the Feedback Cycle</h3><p>The resulting architecture brings together two sources of implementation learning. Technical evidence comes from coding agents, tests, deployment checks, and implementation review. Business feedback comes from people evaluating whether the resulting application helps them achieve the intended outcome.</p><p>Both sources contribute to a reviewed learning process, but their findings need different destinations. A defect might remain within the implementation loop. An ambiguous requirement might lead to a new Build Contract version. A finding about feasibility or business usefulness might change the Decision Context and influence the next investment decision.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*XjTqXfDB0XsSEvNzssYiuA.png" /><figcaption>The complete feedback cycle, combining Technical Evidence and Business Feedback into Reviewed Learning and Living Decision Context.</figcaption></figure><p>There is also an opportunity for learning to extend beyond individual use cases. If one implementation identifies a serious limitation in a dataset shared by several use cases, that information should become available wherever the dataset is referenced. Similarly, repeated failures caused by a particular type of handover instruction should inform improvements to Value Forge itself.</p><p>This will require clear evidence provenance, relationship tracking, and rules for propagating findings. A reported problem in one implementation should trigger reassessment where relevant, rather than silently modifying every related use case.</p><p>Over time, this creates the basis for understanding recurring patterns. We should be able to identify which data limitations prevent implementation, which package instructions produce repeated corrections, which use cases satisfy users, and which opportunities fail to deliver the expected value.</p><h3>What I Am Building Next</h3><p>I plan to develop this in stages, preserving the current Open Data Value and Value Forge workflow while introducing the new capabilities around it.</p><p>The first priority is establishing the durable model for Use Case, Decision Context, Build Contract, and Build Outcome. This includes immutable package versions, source-context references, evidence provenance, and rules for reviewing findings. I also want explicit change records that explain why a requirement or acceptance criterion changes between package versions, drawing on the useful principles demonstrated by OpenSpec without replacing the existing Value Forge architecture.</p><p>The second stage brings feedback into the portal. The new Implementation Feedback tab will allow users to evaluate working proofs of concept and submit ratings and notes connected to specific implementation versions. Reviewers will assess technical findings and user feedback before promoting relevant learning into Decision Context or creating proposed changes to implementation packages.</p><p>The third stage introduces external Build Outcome submissions and automated implementation verification. Coding agents and development pipelines will be able to return structured results, while preview deployments and browser checks provide additional evidence about the application that was produced. External agents will submit findings, but they will not have permission to rewrite the original use case or silently change its business context.</p><p>The first milestone is deliberately limited. I want to take an existing Value Forge package, build the application in Codex, collect implementation evidence, obtain business-user feedback, review the findings, and update the Decision Context. A subsequent package should then be traceable to the evidence and decisions that motivated its changes.</p><p>That will demonstrate whether the full learning cycle works before expanding the automation further.</p><h3>The Real Objective Is Better Decisions</h3><p>Open Data Value began by connecting open data to potential business value. Value Forge extended that process by turning business use cases into implementation-ready packages for coding agents. The next phase is about learning from what happens after those packages leave the platform.</p><p>The broader spec-driven development community has already established useful principles for durable context, structured handovers, change management, and verification. I see no reason to reinvent those ideas. The important work is applying the relevant principles to our existing architecture and connecting the implementation process to the business decisions that motivated it.</p><blockquote>What I want to build is a system where technical results and business-user feedback both contribute to a growing understanding of which opportunities are feasible, useful, and worth pursuing.</blockquote><p>That understanding should guide future implementations, improve the instructions given to coding agents, and help organizations decide where their resources create the most value.</p><p>For open data portal owners, this means moving toward an evidence-backed view of what their published data makes possible and which obstacles prevent its use. For people using AI coding agents, it means preserving the knowledge gained during implementation instead of losing it between projects, prompts, and package versions.</p><p>A working application is an important result, but it is not always the right final outcome. Sometimes the most valuable result of an implementation is learning that the original problem needs a different solution.</p><blockquote>Every implementation should leave the business context better informed than it was before the work started.</blockquote><p>Explore the current platform at <a href="https://create.opendatavalue.com/">create.opendatavalue.com</a>. If you work with open data, data products, or AI-driven implementation and want to discuss this direction, <a href="https://jarkkomoilanen.com/engage">book a 30-minute conversation with me</a>.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=cd3fc5eb6092" width="1" height="1" alt=""><hr><p><a href="https://medium.com/demand-to-data-knowledge-graphs/from-open-data-to-working-applications-the-next-step-is-learning-from-what-we-build-cd3fc5eb6092">From Open Data to Working Applications — The Next Step Is Learning From What We Build</a> was originally published in <a href="https://medium.com/demand-to-data-knowledge-graphs">The Information Continuity Lab</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Minimum Lovable Semantics: Do AI Agents Need a Different Approach?]]></title>
            <link>https://medium.com/demand-to-data-knowledge-graphs/minimum-lovable-semantics-do-ai-agents-need-a-different-approach-25a771ec0d63?source=rss----bcba237793cc---4</link>
            <guid isPermaLink="false">https://medium.com/p/25a771ec0d63</guid>
            <category><![CDATA[ontology]]></category>
            <category><![CDATA[semantics]]></category>
            <category><![CDATA[data]]></category>
            <category><![CDATA[research]]></category>
            <dc:creator><![CDATA[Dr. Jarkko Moilanen]]></dc:creator>
            <pubDate>Fri, 09 Oct 2026 08:57:06 GMT</pubDate>
            <atom:updated>2026-10-09T08:57:08.202Z</atom:updated>
            <content:encoded><![CDATA[<p><em>Why graph-first thinking, proportional semantic rigor, and AI-agent reliability might change how we approach enterprise semantics.</em></p><h3>The divide between ontology and graph practitioners</h3><p>During a recent discussion with data practitioners, someone made an interesting observation. Ontology-driven approaches tend to be more rigid and less agile than graph-first approaches. I have noticed this division before. Some practitioners see formal ontologies as the foundation for trustworthy enterprise knowledge, while others prefer to start connecting real business entities and let the model evolve through practical use cases.</p><p>Both sides have reasonable arguments. Ontologies establish shared meaning, concepts, relationships, and logical rules. Graphs make connections between actual entities explicit and accessible. The distinction is not as clear-cut as some discussions suggest, because ontologies themselves are often represented as graphs. RDF is a graph-based data model, and formal semantic technologies such as OWL build on that foundation.</p><p>What interests me is not which technology wins the argument. I am more interested in the consequences for organizations trying to build useful systems. How much shared meaning do we need before creating something valuable? How much modeling is necessary to avoid ambiguity? And now that AI agents are beginning to consume enterprise knowledge and operate across multiple systems, are we asking the right questions about semantic modeling?</p><p>These discussions led me to an idea I have started calling <strong>Minimum Lovable Semantics (MLS)</strong>. I see it as a potential design approach for the growing number of organizations that need trustworthy semantic context without turning every AI initiative into an enterprise ontology project.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*kuI9ysSo0ebZ_MAETBV6NA.png" /></figure><h3>The problem with defining everything before using anything</h3><p>Traditional enterprise semantic programs often aim to establish a consistent understanding of business concepts across organizational boundaries. That is a legitimate objective. When different departments use different definitions for customers, revenue, services, or operational events, shared semantics help systems exchange information without changing its meaning.</p><p>Problems emerge when the ambition to establish consistency expands beyond what current business needs justify. Organizations might spend considerable time aligning conceptual models, designing hierarchies, debating relationship definitions, and establishing governance structures before the first application demonstrates measurable value. The model might be technically excellent while the business is still waiting for its original question to be answered.</p><p>There is nothing inherently rigid about ontologies. Formal ontology engineering has supported incremental approaches for years. Paola Di Maio’s <a href="https://arxiv.org/abs/1108.1488"><em>Just Enough Ontology Engineering</em></a>, published in 2011, explicitly proposed lightweight and agile practices. W3C’s <a href="https://www.w3.org/TR/skos-primer/">SKOS</a> also supports less formal knowledge organization without requiring comprehensive logical modeling.</p><p>The rigidity practitioners complain about often comes from the implementation and governance process, not the underlying semantic technology. We should distinguish between formal modeling where precision is necessary and organizational habits that require everything to be modeled before anything is delivered. Those are different problems, even though they frequently arrive at the same meeting.</p><h3>Minimum Viable Semantics was already an important step</h3><p>The idea of building only the semantics needed for a particular problem is not new. In a <a href="https://www.linkedin.com/pulse/how-do-semantic-practitioners-tackle-enterprise-peat-layer-jonker-r8rze">2024 discussion of enterprise semantic practices</a>, Ilse Jonker described Minimum Viable Semantics as starting with a small collection of shared concepts and metadata, getting something into production, and expanding from there. Jon Cooke has also described <a href="https://www.linkedin.com/posts/jon-cooke-096bb0_minimum-viable-semantics-part-5-of-building-activity-7460697728797425666-HQIG">Minimum Viable Semantics</a> as applying the smallest amount of semantic structure needed to close a measurable accuracy gap in a specific decision.</p><p>I agree with that direction. It moves semantic engineering away from abstract completeness and toward measurable business results. Rather than modeling an entire domain before implementing an application, we introduce meaning where it improves decisions, interpretation, or integration. That makes the work easier to justify and its results easier to evaluate.</p><p>The discussion is also becoming relevant to AI architecture. In August 2026, Gartner published research titled <a href="https://www.gartner.com/en/documents/8304653"><em>How to Enable the Context Layer With a Minimum Viable Ontology</em></a>. Its argument centers on providing reliable enterprise context for autonomous agents without unnecessary upfront modeling. The industry is clearly questioning how much semantic infrastructure AI systems need.</p><p>Minimum Lovable Semantics therefore needs a stronger justification than being another name for lightweight modeling. Minimum Viable Semantics already addresses business outcomes, incremental delivery, and even AI-agent accuracy. If MLS deserves a separate identity, its difference must be visible in what we design, how consumers use it, and how we measure success.</p><h3>What changes when the consumer is an AI agent?</h3><p>Traditional applications generally follow relationships and decision rules that developers have explicitly implemented. Human analysts bring contextual knowledge that rarely appears in a data model. When something is unclear, they consult documentation, ask colleagues, or apply their own understanding of the organization.</p><p>AI agents change this arrangement. They increasingly discover available capabilities, select data products, interpret business relationships, plan workflows, and interact with other software systems. Their decisions depend on how well they understand the context supplied to them, including information that developers did not explicitly anticipate in every possible workflow.</p><p>Consider an AI agent asked to identify which data products support a business objective. Finding a connection in a graph is relatively straightforward. Understanding what that connection means is another matter. A relationship called supports might indicate that a product supplies source data, provides analytical output, or merely contributes evidence to a decision. An agent that treats these meanings as interchangeable might select the wrong product while producing an explanation that sounds entirely reasonable.</p><p>The challenge becomes more serious when the agent moves from answering questions to initiating actions. It must distinguish authoritative evidence from supporting information, understand how current the information is, and recognize situations where the available context does not justify a conclusion. It also needs to identify the relevant operational capabilities and restrictions. Actual permissions and execution controls must remain enforced by the systems responsible for them, rather than relying on the agent’s interpretation of descriptive metadata.</p><p>This changes the role of semantics. We are no longer concerned solely with whether information is represented correctly. We also need to understand how independently operating software interprets that information and what happens when its interpretation is incomplete or wrong.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*V1zehPh9WkUos8g6e7tIvw.png" /></figure><h3>A practical example: From an energy objective to an agent decision</h3><p>Suppose an energy organization wants to reduce peak electricity demand in a district. It has several data products covering historical consumption, weather forecasts, customer segments, network capacity, and demand predictions. An AI agent receives a request to identify which products support forecasting next week’s peak demand and whether the available evidence is sufficient.</p><p>A traditional domain-first initiative might begin by developing a broad energy ontology covering network infrastructure, customer categories, measurement concepts, operational events, and business processes. That investment might make sense for an organization supporting hundreds of related applications, particularly when formal inference and interoperability are important.</p><p>A Minimum Viable Semantics approach would start with the forecasting decision and define the terms needed to answer it correctly. The team might clarify what peak demand means, how districts are identified, how forecasts differ from actual measurements, and which units apply. The immediate objective would be to eliminate ambiguity affecting the forecasting task.</p><p>Minimum Lovable Semantics would retain that discipline while asking what happens when the semantic model becomes a shared capability used by additional people and agents. An independent agent should identify the same authoritative sources, interpret important relationships consistently, and understand the limits of the available evidence. A business owner should understand why particular data products were selected. A development team should reuse the definitions without arranging another workshop with their original authors.</p><p>The first model might be viable because one agent successfully completes the task. The second becomes lovable when its meaning is accessible, reusable, trustworthy, and manageable as new consumers and requirements appear. The difference is not how many concepts we model. It is whether the semantic capability continues delivering value beyond its first successful implementation.</p><h3>Why I see graphs as the natural starting point</h3><p>This is where graph-first thinking fits particularly well. Business objectives, use cases, data products, datasets, owners, contracts, policies, and evidence do not exist independently. Their relationships explain how information contributes to business outcomes. A graph represents these connections directly and allows different consumers to explore the relevant portions without requiring a single application-specific hierarchy.</p><p>In our energy example, the business objective connects to a forecasting use case. That use case depends on several data products, which expose data through defined interfaces and carry their own quality commitments, ownership information, and operational conditions. Evidence from those products supports a forecast, and the forecast contributes to a business decision. The agent follows these relationships to assemble the context necessary for its task.</p><p>But graph structure alone is not enough. A connection called dependsOn must have a defined meaning. The difference between measures, supports, and contributesTo must be understandable where it affects interpretation. Relationships should have stable identifiers, appropriate provenance, and clear ownership. Otherwise, we create an impressive collection of connections that different consumers interpret differently.</p><p>This suggests a practical architectural principle: <strong>graph-first development, semantics as needed, and agent-validated reliability.</strong></p><p>The graph evolves as business use cases introduce new entities and relationships. Shared vocabulary defines the meaning required for interoperability and trustworthy interpretation. More formal semantic constraints are introduced when ambiguity, complexity, or operational risk demands them. The organization does not abandon ontologies. It uses the appropriate level of semantic formality for the problem.</p><p>The underlying graph technology is not the deciding factor. RDF graphs, property graphs, and other representations might all support parts of this approach. MLS should remain a modeling and operating philosophy rather than become an argument for a particular database or semantic language.</p><h3>Minimum Lovable is different from Minimum Viable</h3><p>I think the distinction becomes clearer when we treat semantics as something consumed by people and AI agents, rather than merely a model produced by semantic experts.</p><p>A minimum viable semantic model demonstrates that a defined task works. A minimum lovable semantic capability should demonstrate that intended consumers independently understand and use the model, that its definitions remain consistent when reused, and that changes do not silently undermine existing applications.</p><p>The term lovable might sound unusual in semantic engineering, but there is a practical meaning behind it.</p><ul><li>People should prefer reusing existing semantic assets over creating their own competing definitions.</li><li>Developers should find them understandable and accessible.</li><li>AI agents should receive machine-readable context that supports reliable interpretation without requiring excessive prompting or undocumented assumptions.</li></ul><p>Lovability is therefore not about making a graph visually appealing or making ontology documentation friendlier. It concerns the quality of the experience of consuming, maintaining, and extending shared meaning.</p><p>It also changes the stopping condition. The smallest model is not necessarily the best model. Adding a few well-defined relationships, examples, stable identifiers, or compatibility rules might slightly increase modeling effort while substantially reducing the work required by every subsequent consumer.</p><p>This is the difference I would like to explore further. Minimum Viable Semantics focuses on enough structure to deliver value for a defined problem. Minimum Lovable Semantics would explicitly require evidence that the shared meaning is suitable for independent consumption, repeated use, and continued evolution. Those practices are not individually new, but treating them as essential acceptance criteria might provide a useful distinction.</p><h3>Semantic context should follow the agent’s task</h3><p>There is another architectural opportunity here. The semantic information maintained by an organization is not necessarily the same information an AI agent needs to receive during a particular task.</p><p>An organization might maintain thousands of concepts, relationships, contracts, and policies. An agent performing a narrowly defined task only needs the relevant portion of that knowledge, together with whatever dependencies and constraints are required for correct interpretation.</p><p>This creates two related but separate design concerns. The shared semantic foundation must be sufficiently complete and maintainable for its intended organizational scope. The context supplied to an agent must be sufficiently relevant and precise for its current task. Reducing the amount of context provided to an agent should never mean removing information necessary for a safe or correct decision.</p><p>I see task-aware semantic delivery as an important part of MLS. Instead of exposing an entire enterprise model to every agent, we assemble relevant context from the graph, vocabulary, product definitions, and operational metadata. The agent receives enough information to interpret the task and identify appropriate capabilities, without carrying unrelated semantic material through every reasoning step.</p><p>This is close to the architectural thinking we have already applied in the Open Data Products SDK, where structured artifacts, graph relationships, and agent-oriented context extraction support more focused consumption. It is also consistent with a broader design preference I have followed in AI systems: deterministic operations should handle work that does not require generative reasoning, while AI reasoning is reserved for tasks where interpretation and judgment add value.</p><p>A useful semantic architecture should support both. It should allow an agent to follow explicit, validated relationships where they exist and recognize when a question genuinely requires additional reasoning or human judgment.</p><h3>The missing feedback loop: Agents testing semantics</h3><p>Perhaps the most interesting opportunity is to use AI agents as a source of evidence about semantic quality.</p><p>Traditional semantic quality checks examine definitions, consistency, structural constraints, and logical validity. Those remain essential. But agent-based systems introduce additional evidence through their operational behavior. We observe which relationships agents misunderstand, where they select inappropriate products, when they disagree about concept meanings, and how frequently they lack sufficient context to complete a task.</p><p>Those observations might help us identify semantic weaknesses that static validation does not expose. An agent that repeatedly confuses forecast values with measured observations might reveal a missing distinction in the shared vocabulary. Two agents that reach incompatible interpretations of the same relationship might indicate ambiguity in its definition or differences in the context supplied to them.</p><p>This does not mean agents should autonomously rewrite authoritative business definitions. Semantic ownership and change approval remain human and organizational responsibilities. It means operational failures become inputs to controlled semantic improvement.</p><p>For MLS, I would want to test semantic correctness, independent consumability, cross-agent interpretation consistency, context efficiency, and the effort required to incorporate changing business requirements. We also need tests for situations where information is missing or contradictory. An agent that confidently invents an unknown relationship should fail, even if it produces a convincing explanation.</p><p>These measurements would help determine whether MLS offers something beyond familiar engineering principles. Without measurable acceptance criteria and practical results, it risks becoming another attractive label for work we were already supposed to be doing.</p><h3>Where the Open Data Products family fits</h3><p>I see a natural connection between MLS and the <a href="https://opendataproducts.org/">Open Data Products standards family</a>, which we have been developing under LF AI &amp; Data.</p><p>ODPV provides shared vocabulary and semantic alignment. ODPG specifies relationships between data products, business objectives, use cases, policies, and other enterprise objects. ODPS defines individual data products and their commitments, while ODPC supports catalog and portfolio structures. ODPR, currently a draft, addresses repeatable workflows, and the Python SDK provides tooling for validation, traversal, and agent-oriented operations.</p><p>These components follow a deliberate separation of concerns. Business meaning does not need to be buried inside every individual data product specification. Relationships do not need to be duplicated across disconnected metadata documents. Agent workflows do not need to reinvent definitions that already exist in the shared vocabulary and product contracts.</p><p>The <a href="https://opendataproducts.org/framework/">Data Product Operating Framework</a> brings these elements into an organizational operating model. It addresses the responsibilities, governance, lifecycle, and evidence required to make the connected specifications useful in practice. That gives us a possible home for MLS as a design principle rather than another specification competing for attention.</p><p>I would not position MLS as an alternative to formal enterprise ontologies or existing semantic platforms. Quite the opposite. Organizations should reuse authoritative semantic assets where they exist and introduce additional formalism when necessary. A lightweight approach that ignores established definitions is not agile. It is simply creating future integration problems.</p><h3>Where this approach might fail</h3><p>Before advocating MLS as a methodology, we need to understand its limits. Graph-first development might encourage teams to create local relationships faster than they establish shared definitions. A model that works for one use case might become confusing when its relationships spread into unrelated domains. Different agents might interpret the same graph differently because of their models, prompts, or retrieved context, even when the underlying vocabulary is consistent.</p><p>There are also environments where more formal semantics are necessary from the beginning. Regulatory reporting, scientific knowledge integration, complex logical inference, and high-risk autonomous operations might require rigorous constraints and established ontologies. In these settings, a broad formal model might be the most economical long-term choice, especially when many applications already depend on it.</p><p>MLS should not turn minimalism into a dogma. The necessary level of formality depends on the consequences of incorrect interpretation, the number of consumers, and the need for interoperability. Sometimes the minimum responsible solution is a substantial formal ontology.</p><p>The challenge is finding a repeatable way to make that decision rather than allowing individual preferences for graphs, ontologies, or modeling tools to determine the architecture.</p><h3>What should Minimum Lovable Semantics become?</h3><p>I am treating Minimum Lovable Semantics as a proposal for discussion, not as a completed methodology or a new semantic standard. Much of its foundation already exists in lightweight ontology engineering, Minimum Viable Semantics, knowledge graphs, semantic interoperability, and AI-agent architecture.</p><p>The possible contribution is their combination into a consumer-centered approach where semantic models are built around business use cases, represented through connected knowledge, delivered as relevant context, and continuously evaluated through the behavior of people and AI agents.</p><p>But important questions remain unresolved. What evidence would demonstrate that a semantic capability is genuinely lovable rather than merely viable? Should independent consumption by another team or another agent be mandatory? How much cross-agent agreement is sufficient, especially when agents use different models and reasoning strategies?</p><p>We also need to determine where graph-first development should give way to more formal ontology engineering. How do we establish the right level of semantic rigor for different operational risks? How should authoritative business definitions be governed while still allowing use-case teams to move quickly? And which parts of semantic context must be available at runtime, rather than only during design?</p><p>There is also the question of measurement. If MLS is worth adopting, it should produce observable benefits in semantic reliability, consumer adoption, reuse, development effort, and change management. We need practical comparisons against existing approaches, including cases where MLS performs worse.</p><p>I would like to develop this thinking with the broader community before attempting to formalize it within the Data Product Operating Framework. I am particularly interested in the experience of ontology engineers, knowledge graph practitioners, enterprise architects, data product teams, and developers building AI agents in production.</p><p>I suspect the most useful outcome would not be another argument about whether graphs or ontologies are better. It would be a shared understanding of how much meaning our systems need, how to make that meaning useful to independent consumers, and how to verify that AI agents interpret it reliably.</p><p><strong>The open question is whether Minimum Lovable Semantics gives us a better way to get there, and what we would need to prove before calling it an approach worth adopting.</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*S4Pc6FK8tpzmjBFhtuJMvA.png" /></figure><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=25a771ec0d63" width="1" height="1" alt=""><hr><p><a href="https://medium.com/demand-to-data-knowledge-graphs/minimum-lovable-semantics-do-ai-agents-need-a-different-approach-25a771ec0d63">Minimum Lovable Semantics: Do AI Agents Need a Different Approach?</a> was originally published in <a href="https://medium.com/demand-to-data-knowledge-graphs">The Information Continuity Lab</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Model Is Part of the Method: Testing LLM Sensitivity in a Demand-to-Data Research Pipeline]]></title>
            <link>https://medium.com/demand-to-data-knowledge-graphs/the-model-is-part-of-the-method-testing-llm-sensitivity-in-a-demand-to-data-research-pipeline-a608c2ebd161?source=rss----bcba237793cc---4</link>
            <guid isPermaLink="false">https://medium.com/p/a608c2ebd161</guid>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[mrdm]]></category>
            <category><![CDATA[demand-to-data]]></category>
            <category><![CDATA[decision-making]]></category>
            <dc:creator><![CDATA[Dr. Jarkko Moilanen]]></dc:creator>
            <pubDate>Mon, 05 Oct 2026 06:14:11 GMT</pubDate>
            <atom:updated>2026-10-09T09:48:36.539Z</atom:updated>
            <content:encoded><![CDATA[<h4>Stream 1–04</h4><p>This experiment is part of a <a href="https://medium.com/demand-to-data-knowledge-graphs">larger research program</a> (and PhD Research) looking at whether demand for data products can be derived systematically from available metadata. The broader work starts from open-data portals because they provide a large, distributed, and accessible source of metadata across countries, sectors, and platform types. The method does not depend on open data itself. The same approach could be applied anywhere similar metadata is available. Below you find spoken thoughts on the topic and below that the experiment details as text.</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FG0kisSy9ArQ%3Ffeature%3Doembed&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DG0kisSy9ArQ&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FG0kisSy9ArQ%2Fhqdefault.jpg&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/ade94c902c41a939e1d36da3e93013d1/href">https://medium.com/media/ade94c902c41a939e1d36da3e93013d1/href</a></iframe><h3>Overview of the context and experiment</h3><p>The core research pipeline takes that metadata and builds a structured representation of potential demand. It derives business objectives, use cases, KPIs, information needs, candidate data products, relationships between them, and gaps in the available data. The aim is to understand whether we can move from available data toward evidence-based candidate use cases and data products in a repeatable way.</p><p>That creates an important methodological question. Part of this pipeline depends on an LLM to interpret the evidence and construct parts of that representation. If I keep the evidence, prompts, schemas, thresholds, deterministic processing, and pipeline configuration fixed, then change only the model, how much does the resulting representation change? This model-sensitivity experiment was designed to answer that question.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*yn8-Kl1dpB1rdxDVBG6gpQ.png" /></figure><p>The test was therefore not about asking three models the same question and comparing their prose. Each model operated inside the same end-to-end Demand-to-Data and MRDM pipeline. What I wanted to know was whether changing the model changes only the wording around the result, or whether it changes the structure of the problem that the research pipeline produces.</p><p>When an LLM sits inside a research or decision pipeline, it is tempting to treat the model as infrastructure. You define the process, fix the prompts, define the schemas, connect the data, then select whichever model seems appropriate for the job. If the pipeline stays the same, the assumption is that the underlying method stays broadly the same too.</p><blockquote>“I wanted to test that assumption directly by asking how sensitive Demand-to-Data and MRDM generation is to the model condition when the evidence, prompts, schemas, thresholds, deterministic processing, and pipeline configuration remain constant.”</blockquote><p>This was not a benchmark where</p><ul><li>three models received the same question and their answers were compared.</li><li>Each model operated inside the same multi-step generation pipeline.</li></ul><p>The pipeline derives a structured representation of demand from available metadata, including objectives, use cases, KPIs, information needs, candidate data products, graph relationships, explicit gaps, and links to available datasets. The model therefore participates in constructing the representation that later analysis and decisions depend on.</p><blockquote>If model choice changes that representation materially, it becomes difficult to argue that the model is merely an interchangeable technical component.</blockquote><p>The experiment was designed to isolate that question as tightly as possible. Metadata richness and the presence of schema information were included as secondary interaction questions, while cross-portal pattern discovery was deliberately excluded because that belongs to a separate study. Stage 1 focused only on structural and computational sensitivity. Semantic quality was reserved for a separate blinded human evaluation.</p><h3>A controlled sample rather than a convenient sample</h3><p>The final experiment used 120 frozen dataset identities arranged into 30 portal subsets, with four records in each subset. The portal subset was defined as the experimental unit. Individual records, generated nodes, graph edges, or candidate products were not treated as independent observations because doing so would artificially increase the apparent sample size and ignore the structure of the experiment.</p><ul><li>Each of the 30 portal subsets was processed under</li><li>three exact model conditions, and</li><li>each condition was executed three independent times.</li></ul><p>The complete design therefore contained 270 execution units, calculated as 30 portal subsets multiplied by three model conditions multiplied by three replicates. Seed control was unavailable, so the replicates were genuine independent executions rather than deterministic reruns using fixed generation seeds.</p><p>The sample was constrained before any experimental generation took place. It contained exactly</p><ul><li>40 low-richness,</li><li>40 medium-richness, and</li><li>40 high-richness records.</li></ul><p>At least 20 records had to contain schema information, and the final sample contained 49. The portal distribution was also fixed at</p><ul><li>17 Huwise portals,</li><li>12 CKAN portals, and</li><li>one Data Europa Search portal.</li></ul><p>Eight portals were classified as small, 14 as medium, and eight as large. Geographic coverage required at least 16 countries across five represented continents, with France capped at eight portals and every other country capped at three.</p><p>There was another constraint that matters for the interpretation of the experiment. Every selected record had to share a deterministic capability with another record inside its four-record portal subset. The intent was to avoid building portal subsets from records that had no meaningful basis for comparison inside the deterministic part of the pipeline. Sampling therefore reflected the requirements of the research design rather than simply drawing 120 convenient records from the source population.</p><h3>The first two experimental designs failed</h3><p>The final experiment was version 3, but versions 1 and 2 were not deleted or rewritten. Both failed before experimental execution, and both were preserved as part of the research record. That history matters because the problems were found before any model results existed, which prevented the sampling rules from being adjusted in response to desirable or undesirable outputs.</p><p>Version 1 imposed a universal maximum of three portals per country. That constraint turned out to be incompatible with the required Huwise allocation. The experiment therefore stopped before execution. Version 2 made one specific amendment, allowing up to eight French portals while retaining the three-portal maximum for every other country. That solved the geographic feasibility problem but exposed another issue in the underlying sampling population.</p><p>Duplicate portal and dataset identities existed in the eligible population. Most were repeated identities whose source-content hashes were identical, which meant they represented duplicated instances of the same underlying content. One economie-france identity was different because the same identity appeared with conflicting source-content hashes. Treating those conflicting versions as one record, choosing one version, or repairing them during preparation would have introduced an arbitrary research decision into the sample.</p><p>Version 3 therefore introduced an explicit source-identity rule. The preparation process collapsed 379 repeated identities whose source-content hashes were identical and excluded the single conflicting economie-france identity completely. It did not merge the conflicting records, repair them, or select the version that happened to be more convenient. After that resolution, richness and sampling were recomputed from 134,099 eligible identities. The earlier protocol versions remain preserved as failed pre-execution versions rather than being overwritten by the version that eventually worked.</p><h3>Freezing the experiment before seeing the outputs</h3><p>The execution gate was intentionally strict because model-sensitivity research becomes difficult to defend if the sample, prompts, model aliases, thresholds, or analysis rules drift during execution. Before any experimental generation call, the process verified parent-run and Article 1 checksums, recomputed identity resolution and the eligible population, recomputed richness medians and strata, completed the sampling-feasibility checks, and generated the deterministic sample.</p><p>The experiment then froze the sample bytes, prompts, thresholds, schemas, configuration, protocol, analysis plan, execution plan, failure taxonomy, pricing assumptions, source snapshot, and predecessor evidence. A total of 227 frozen files and all 270 planned execution units were verified before execution. The sample SHA-256 was confirmed, the exact model snapshots were checked for availability, and a separate checksum-bound execution authorization was required before model calls were permitted.</p><p>The protocol also defined what should happen if a contradiction appeared after freezing. The experiment was not allowed to modify version 3 in place. It had to stop and create a new protocol version. That sounds procedural, but it addresses a basic problem in LLM experiments: it is easy to make small adjustments during execution and later forget that those adjustments changed the experiment.</p><h3>Three exact model conditions</h3><p>The final experiment used three fixed model snapshots.</p><ul><li>Condition A used gpt-4o-mini-2024–07–18.</li><li>Condition B used gpt-5.4-mini-2026–03–17 with reasoning effort set to none.</li><li>Condition C used gpt-5.4–2026–03–05 with reasoning effort also set to none.</li></ul><p>Moving aliases and silent model substitution were prohibited, and every returned model identifier matched its authorized snapshot.</p><p>Provider retries were set to zero, with one application attempt per call. The intention was to observe the behaviour of the experimental condition rather than hide difficult calls behind repeated attempts until a desired result appeared. <strong>The completed experiment contained 1,192 model calls across the 270 planned execution units.</strong> Every unit reached a terminal state, there were no failed terminal units, no application retries, and no checksum failures.</p><h3>What Stage 1 measured</h3><p>Stage 1 retained every replicate, valid empty result, warning, raw call, token count, latency measurement, execution duration, and structural outcome. A replicate could not be discarded because it was empty, anomalous, expensive, or contrary to expectations. This matters because empty results and unusual outputs are part of model behaviour, not inconvenient observations to remove from the dataset.</p><p>The structural analysis compared objectives, use cases, KPIs, information needs, candidate data products, nodes, edges, explicit gaps, linked datasets, and structural overlap. The computational analysis included calls, tokens, latency, duration, and estimated cost. The main model-condition comparisons used a portal-blocked Monte Carlo permutation procedure with 9,999 draws. Confidence intervals used 5,000 bootstrap draws, and the three pairwise comparisons between A and B, A and C, and B and C were corrected using the Holm procedure. Within-model replicate variation was kept separate from between-model-condition variation.</p><p>This stage was intentionally limited to structure and computation. A model producing more KPIs, more graph nodes, or more information needs does not mean it produced better KPIs, better nodes, or better information needs. Output quantity, graph size, product formation, and gap counts are not semantic-quality measures, and the Stage 1 protocol explicitly prevents those numbers from being used to declare a model superior.</p><h3>The structures changed substantially</h3><p>The differences between the model conditions were not small. Condition A produced an average of 2.04 objectives per portal-replicate unit, Condition B produced 2.43, and Condition C produced 2.10. The differences became larger further down the generated structure. A produced 3.06 use cases on average, compared with 4.68 for B and 4.41 for C. For KPIs, A produced 5.61 while B and C both produced slightly above 12.</p><p>Information needs showed one of the strongest differences. Condition A generated an average of 14.97 information needs, Condition B generated 49.81, and Condition C generated 22.99. The resulting graphs followed the same pattern. A produced an average of 27.01 nodes and 51.41 edges, B produced 63.61 nodes and 109.62 edges, and C produced 35.90 nodes and 89.92 edges. Explicit gaps averaged 19.01 for A, 78.04 for B, and 28.33 for C.</p><p>These were not merely descriptive differences in the observed means. For use cases, KPIs, information needs, nodes, edges, and explicit gaps, the pre-specified omnibus test produced p = 0.0001. Objectives also differed across model conditions, with p = 0.0053. Candidate data products were the important exception, with an omnibus p-value of 0.1219.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*auYMQznsDFnou545yvl4Bw.png" /></figure><p>The practical implication is difficult to dismiss. The evidence, prompts, deterministic processing, schemas, thresholds, sample, and pipeline stayed fixed, yet the inferred representation of demand changed substantially when the model changed. The difference was not limited to wording.</p><blockquote>“Model selection affected how the system decomposed business demand and how much structure it created around that demand.”</blockquote><h3>The full model did not simply produce more</h3><p>There is a second result that complicates the usual assumption that larger models simply produce richer versions of what smaller models produce. Condition C used the full GPT-5.4 model, while Condition B used GPT-5.4-mini. If model scale translated into straightforward structural expansion, C should have generated at least as much structure as B. That is not what happened.</p><p>Condition B generated an average of 49.81 information needs compared with 22.99 for C. It produced 63.61 graph nodes compared with 35.90, 109.62 edges compared with 89.92, and 78.04 explicit gaps compared with 28.33. The adjusted pairwise tests found significant B versus C differences for information needs, nodes, edges, and explicit gaps. Their KPI counts were nearly identical, while their use-case counts were not significantly different after correction.</p><p>That does not establish that the mini model performed better. It establishes something more interesting for this stage of the research:</p><blockquote>“model families and model sizes do not appear to sit on a simple scale where increasing model capability produces progressively larger versions of the same representation. The models appear to decompose the same evidence differently.”</blockquote><h3>Candidate data products resisted the obvious story</h3><p>Candidate data products produced a result worth examining because the descriptive numbers invite a conclusion that the statistical analysis does not support. Condition A produced an average of 1.01 candidate data products, B produced 0.76, and C produced 0.47. Looking only at those means, it would be easy to say that the larger models generated fewer candidate products.</p><p>The pre-specified omnibus comparison does not support that conclusion. Its p-value was 0.1219. The A versus C pairwise comparison produced an unadjusted p-value of 0.0206, but after applying the planned Holm correction it became 0.0618. That comparison therefore did not meet the adjusted significance threshold.</p><p>This is a useful example of why the analysis plan had to exist before execution. Without it, the descriptive difference is attractive enough to become a headline. With the pre-specified correction applied, the appropriate conclusion is much narrower: candidate-product counts differed descriptively, but Stage 1 did not establish a statistically significant model-condition effect for that outcome.</p><blockquote>The upstream structures changed dramatically while candidate-product formation did not show the same statistically supported model effect. That suggests a possible distinction between “how much structure the model generates” and “how many product opportunities the pipeline ultimately forms.” More decomposition does not automatically create more candidate products.</blockquote><h3>Model choice mattered more than rerunning the same model</h3><p>Because LLM output is stochastic, the obvious challenge is whether these differences simply reflect normal variation from one run to another. The three-replicate design allowed that question to be examined separately. Between-model-condition variance exceeded within-model replicate variance for every structural measure where variation existed.</p><p>The between-to-within variance ratio was 1.82 for objectives, 4.90 for use cases, 4.26 for KPIs, 6.30 for information needs, 6.15 for nodes, 5.22 for edges, and 6.34 for explicit gaps. Candidate data products also had a ratio above two even though the omnibus model-condition comparison for product counts was not statistically significant.</p><p>This is one of the more consequential findings in the experiment:</p><blockquote>“Repeating the same model did produce observable variation, as expected, but changing the model condition produced substantially greater variation across most of the generated structure.”</blockquote><p>In this experiment, model selection was therefore a larger source of structural variation than replicate-to-replicate stochasticity.</p><h3>The empty outputs tell another part of the story</h3><p>The execution results also revealed a pattern that disappears if analysis focuses only on average graph size. Condition A produced structural output in 59 of its 90 execution units. Condition B did so in 26, and Condition C in 22. The remaining executions were valid empty results rather than failures.</p><p>At the same time, B and C produced substantially larger structures across several aggregated measures. That suggests two separate behaviours are operating. One concerns whether the model and pipeline produce a structural result for a portal subset at all. The other concerns how much structure the system generates once it proceeds.</p><p>A model might cross the generation threshold more frequently while producing relatively compact representations, while another might proceed less frequently but generate much denser structures when it does. That distinction matters for practical systems.</p><blockquote>“If one model produces outputs for many evidence sets while another declines more often but expands aggressively when it does proceed, describing either model as simply more productive or more conservative loses important information about its behaviour.”</blockquote><h3>Cost introduces another uncomfortable question</h3><p>The computational differences were also substantial. Across the three replicates, Condition A cost about $0.24, Condition B about $2.85, and Condition C about $10.93 based on the recorded token-derived rates. The full experiment consumed 3,459,233 tokens, with a total estimated cost of $14.02.</p><p>Condition C therefore cost almost four times as much as B, yet B generated substantially more information needs, nodes, edges, and explicit gaps. Again, this does not establish that B delivered more value. Stage 1 has no basis for making that claim. It does show that price, model size, and structural output do not move together in a simple way.</p><p>For organizations building large AI pipelines, that distinction has practical consequences:</p><blockquote>“A more expensive model might produce a smaller representation because it rejects weaker inferences, consolidates concepts more aggressively, or behaves differently around evidence thresholds. It might also miss useful opportunities that another model surfaces.”</blockquote><p>Structural counts alone cannot separate those possibilities, which is why cost-effectiveness cannot be evaluated properly until semantic quality enters the analysis.</p><h3>What the experiment changes for me</h3><p>Before running this experiment, I would have described the LLM as one component of the Demand-to-Data pipeline. After seeing these results, that description feels incomplete. When a model participates in turning source evidence into objectives, KPIs, requirements, gaps, products, and relationships, changing the model changes part of the mechanism through which the problem itself is represented.</p><p>That distinction becomes especially important when generated structures feed human decisions.</p><ul><li>If one model derives 15 information needs from a body of evidence and another derives 50, those representations create different work.</li><li>If one identifies 19 gaps and another identifies 78, they create different governance backlogs.</li><li>If one representation contains 27 graph nodes and another contains 64, people downstream are no longer looking at slightly different wording around the same problem.</li></ul><p>They are looking at materially different decompositions of that problem. And this also creates a reproducibility issue for production systems.</p><blockquote>“A team might freeze its prompts, schemas, deterministic logic, and source data, then update the underlying LLM six months later because a newer model becomes available. From a software perspective, that looks like an upgrade. If the resulting knowledge structure changes substantially, the change starts to look more like a methodological revision that deserves validation.”</blockquote><h3>Stage 1 does not tell us which representation is better</h3><p>The most important limitation of these results is also deliberate. More structure does not mean better structure. A model generating twice as many KPIs might have identified useful dimensions that another model missed, or it might have fragmented a smaller set of strong KPIs into unnecessary detail. A model reporting four times as many gaps might be more sensitive to missing evidence, or it might be generating gaps that human reviewers consider irrelevant.</p><p>Stage 1 therefore authorizes conclusions only about structural and computational sensitivity. No semantic evaluator was used, no LLM judge ranked the models, and the analysis manifest explicitly states that no model-superiority claim is authorized from this stage. Stage 2 remains a separate blinded independent human evaluation.</p><p>That second stage is where the difficult questions move from quantity to usefulness. Human reviewers need to assess whether the additional use cases are relevant, whether the KPIs are meaningful, whether information needs are supported by the evidence, whether identified gaps are genuine, and whether candidate data products represent useful opportunities. Only then does it become possible to compare the extra computational cost with the semantic value produced.</p><h3>The question I think is worth debating</h3><p>The result I am comfortable defending at this stage is that model choice is an experimental variable, not merely an implementation detail, when an LLM participates in creating the structures that a study or decision process later analyzes. <strong>In this experiment, changing the model changed the generated representation more than repeating the same model changed it.</strong></p><p>That leaves a broader question for AI systems already operating inside organizations. Teams increasingly use LLMs to generate requirements, strategies, risk assessments, knowledge graphs, use cases, investment proposals, and other structured material that affects decisions. <strong>If replacing the model materially changes the representation of the same evidence, then organizations need to decide when a model upgrade stops being ordinary maintenance and becomes a methodological change</strong> that needs revalidation.</p><p>Stage 2 should make that discussion more interesting. The structural differences are now measurable. The remaining question is whether humans judge those differences to represent more useful understanding, more unnecessary structure, or some mixture of both.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*S4Pc6FK8tpzmjBFhtuJMvA.png" /></figure><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=a608c2ebd161" width="1" height="1" alt=""><hr><p><a href="https://medium.com/demand-to-data-knowledge-graphs/the-model-is-part-of-the-method-testing-llm-sensitivity-in-a-demand-to-data-research-pipeline-a608c2ebd161">The Model Is Part of the Method: Testing LLM Sensitivity in a Demand-to-Data Research Pipeline</a> was originally published in <a href="https://medium.com/demand-to-data-knowledge-graphs">The Information Continuity Lab</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[What 134,480 Dataset Records Produced]]></title>
            <link>https://medium.com/demand-to-data-knowledge-graphs/what-134-480-dataset-records-produced-2389d1196d95?source=rss----bcba237793cc---4</link>
            <guid isPermaLink="false">https://medium.com/p/2389d1196d95</guid>
            <category><![CDATA[demand-to-data]]></category>
            <dc:creator><![CDATA[Dr. Jarkko Moilanen]]></dc:creator>
            <pubDate>Sun, 04 Oct 2026 06:27:14 GMT</pubDate>
            <atom:updated>2026-10-04T12:59:41.649Z</atom:updated>
            <content:encoded><![CDATA[<h4>Stream 1–03</h4><p>I have completed the first descriptive analysis for the first study in my new PhD research.</p><p>The research starts from a problem I have encountered repeatedly while working with open-data platforms. Data owners already have substantial data inventories, but they do not necessarily know what those assets could support. The practical question is often not “<em>What data do we need for this use case?</em>” but “<em>Given the data we already have, what could we do with it?</em>”</p><p>This study investigates that inverse direction. Instead of starting with a predefined business objective or use case and searching for data, the implemented artifact starts with existing dataset metadata. It uses that metadata as evidence for constructing candidate Machine-Readable Decision Models, or MRDMs (<a href="https://medium.com/demand-to-data-knowledge-graphs/from-dataset-metadata-to-decision-models-537a79c2cb51">MRDM concept discussed in more detail in previous post</a>). The first large-scale generation and descriptive analysis is now complete.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/721/1*DhclToPAnSdrfqYtq9RGjQ.png" /><figcaption>Figure 1. Descriptive scale of the completed generation run. The figures describe generated structures and do not represent semantic-quality measures.</figcaption></figure><h3>From existing data supply to candidate decision models</h3><p>An MRDM is more than a generated use-case description. It represents a candidate decision context connecting a potential use with the decision supported, expected value, Business Objectives, Information Needs, candidate Data Products, supporting datasets, provenance, and explicit gaps.</p><p>The distinction between available data and required information is deliberate. A dataset can provide sufficient evidence to suggest a potentially useful decision without providing all the information required to support that decision. The model therefore allows Information Needs to exist independently of the data currently available.</p><p>The completed analysis covers 65 heterogeneous open-data portal snapshots and 134,480 dataset metadata records. The pipeline generated 538 candidate Business Objectives, 1,405 candidate Use Cases, 1,091 candidate KPIs, 5,333 normalized Information Needs, and 661 candidate Data Products. The resulting graphs contain 130,514 nodes and 174,394 edges. All graph edges retain artifact-level provenance. These numbers establish the generated population. They do not establish its semantic quality.</p><h3>Information demand extends beyond existing supply</h3><p>The structural results provide an initial view of what happens when candidate decision contexts are derived from existing data supply.</p><p>The 1,405 candidate Use Cases contain an average of 3.80 Information Needs. The generated population contains approximately 0.47 candidate Data Products per Use Case, while each candidate Data Product is composed of approximately 3.06 datasets.</p><p>More interestingly, 3,358 of the 5,333 Information Needs, approximately 63%, are represented with at least one explicit gap.</p><p>This result should not be interpreted as evidence that 63% of the required data is missing. A graph gap has a narrower meaning within the artifact. It indicates that some part of the generated information demand is not satisfied by the relationships represented in the graph.</p><p>The result nevertheless illustrates why Information Need is modeled separately from Dataset. If the artifact generated use cases only around data already available, information demand would be constrained by supply. An apparently plausible use could then look complete simply because requirements unsupported by existing datasets were never represented. The current model does not make that assumption.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*lo1jMDmzbGRQlnc_FsKXLg.png" /><figcaption>Figure 2. Of 5,333 generated Information Needs, 3,358 have at least one explicit graph gap. A gap indicates unsupported demand within the generated graph and does not establish that the corresponding information is absent from the real world.</figcaption></figure><h3>The graph contains 20,422 explicit gaps</h3><p>The completed graphs contain 20,422 explicit gaps. These consist of 875 dataset gaps, 41 KPI gaps, 449 ownership gaps, 3,505 product gaps, and 15,552 unused-data opportunities.</p><p>At this stage, these are structural observations rather than findings about the underlying data ecosystems. A gap could result from genuinely unavailable information, insufficient metadata, the boundaries of the retained input, unsuccessful matching, or a generated Information Need that later proves unsupported.</p><p>This distinction is important. The purpose of the descriptive analysis is to characterize what the artifact produced before independent semantic evaluation begins.</p><p>One negative case is already visible. Argentina is the only portal for which the pipeline produced no candidate Data Product. Its 1,287 analyzed records still produced one objective, three Use Cases, four KPIs, nine Information Needs, 1,310 graph nodes, 1,332 edges, and 20 gaps. The absence of a Data Product was retained rather than corrected after generation.</p><p>Negative outcomes remain part of the empirical record rather than being repaired to make the generated structures appear complete.</p><h3>Metadata richness varies substantially</h3><p>The second research question concerns the relationship between source metadata and the validity of the resulting decision models.</p><p>The first analysis now provides a deterministic metadata profile for every one of the 134,480 analyzed records. Titles are present in 100% of the records, descriptions in 88.8%, tags or categories in 98.0%, and publisher information in 99.8%. License information is present in 83.8%, while machine-readable access information is available for 91.8%. Geographic information is present in 52.9% of the records. Schema or field-level information appears in only 8.4%.</p><p>The median normalized description contains 265 characters and 26 unique word tokens. The latter is retained as a lexical-richness proxy rather than treated as a measure of semantic quality.</p><p>The low availability of schema or field information is particularly relevant. The artifact is attempting to construct candidate decision contexts primarily from descriptive catalog metadata rather than from detailed profiling of dataset contents.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/642/1*Qbn3Yqe5TKM3sl0i9z-B1A.png" /><figcaption>Figure 3. Presence of selected metadata elements across the 134,480 analyzed records. Temporal and update fields indicate field presence in the normalized snapshots and should not be interpreted as evidence that the underlying datasets are current.</figcaption></figure><p>Whether this evidence is sufficient remains unanswered. If independently assessed MRDM validity remains strong despite limited schema-level metadata, that would provide evidence about how much useful semantic context can be derived from existing catalog descriptions. If validity varies substantially with metadata richness, that instead establishes a boundary condition for the approach. Both outcomes are scientifically useful.</p><h3>The portals behave differently</h3><p>The 65 portal snapshots are heterogeneous in size and generated structure. The median portal contributes 500 analyzed records, with an interquartile range from 200 to 1,610 and a maximum of 10,000. The median number of gaps per portal is 112, while the maximum is 4,169.</p><p>The descriptive outlier analysis also produces different portals depending on the measure considered. Boston, Long Beach, and Seine-Saint-Denis have high numbers of generated Use Cases relative to input size. Boston, Bretagne, and Electricity North West have high Information Needs per Use Case. Other portals appear as outliers for Data Products per Use Case, datasets per Data Product, gap ratios, or graph structure.</p><p>These differences should not yet be interpreted as differences in quality. A system generating more Use Cases from the same number of metadata records is not necessarily performing better. It could be identifying more legitimate opportunities, or it could simply be generating more aggressively. Semantic evaluation is required to distinguish those explanations.</p><h3>The next dataset contains human judgments</h3><p>The technical generation phase is now separated from the semantic evaluation phase. A reproducible sample of 130 complete MRDMs has been prepared for independent review. The sample contains exactly two cases from each portal and is evenly divided between 65 Use Case paths connected to a candidate Data Product and 65 paths without a linked candidate Data Product.</p><p>A second evaluation dataset contains 258 individual graph relationships: 65 Business Objective to Use Case relationships, 65 Use Case to Information Need relationships, 64 Information Need to Data Product relationships, and 64 Data Product to Dataset relationships. Argentina accounts for the two missing product-related cases because its graph contains no candidate Data Product.</p><p>The reviewer-facing artifacts contain the source context and provenance required for assessment, but generated confidence and score fields have been removed. Human judgment fields remain blank.</p><p>The research boundary is therefore clear. The completed analysis establishes what the artifact generated, how the generated structures differ across environments, what metadata was available as evidence, and how every generated relationship can be traced. It does not establish that the generated decision models are semantically valid.</p><p>The next phase introduces independent human judgment. The resulting evaluation will allow the structural observations reported here to be compared with MRDM-level and relationship-level semantic validity and, subsequently, with the characteristics of the source metadata.</p><p>The first study has therefore moved from artifact construction to empirical evaluation. Time to celebrate! :D</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*S4Pc6FK8tpzmjBFhtuJMvA.png" /></figure><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=2389d1196d95" width="1" height="1" alt=""><hr><p><a href="https://medium.com/demand-to-data-knowledge-graphs/what-134-480-dataset-records-produced-2389d1196d95">What 134,480 Dataset Records Produced</a> was originally published in <a href="https://medium.com/demand-to-data-knowledge-graphs">The Information Continuity Lab</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[From Dataset Metadata to Decision Models]]></title>
            <link>https://medium.com/demand-to-data-knowledge-graphs/from-dataset-metadata-to-decision-models-537a79c2cb51?source=rss----bcba237793cc---4</link>
            <guid isPermaLink="false">https://medium.com/p/537a79c2cb51</guid>
            <category><![CDATA[jarkko-moilanen]]></category>
            <category><![CDATA[mrdm]]></category>
            <category><![CDATA[demand-to-data]]></category>
            <category><![CDATA[research]]></category>
            <dc:creator><![CDATA[Dr. Jarkko Moilanen]]></dc:creator>
            <pubDate>Sat, 03 Oct 2026 10:58:34 GMT</pubDate>
            <atom:updated>2026-10-04T13:00:23.799Z</atom:updated>
            <content:encoded><![CDATA[<h4>Stream 1–02: Designing Graph-Supported Machine-Readable Decision Models for Supply-First Data Discovery</h4><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FEvmF7XoJ0zI%3Ffeature%3Doembed&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DEvmF7XoJ0zI&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FEvmF7XoJ0zI%2Fhqdefault.jpg&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/fdb9e82be00caa4227cf309a2fa60482/href">https://medium.com/media/fdb9e82be00caa4227cf309a2fa60482/href</a></iframe><h4>Abstract</h4><p><em>Organizations increasingly possess large volumes of cataloged data but often lack structured knowledge about what decisions those data assets might support. Data catalogs primarily describe available data supply, while use cases, information requirements, expected value, and business objectives are commonly developed through separate processes. Research on dataset discovery has developed increasingly sophisticated methods for finding data in response to known information needs or natural-language queries, including semantic data discovery, dataset recommendation, and knowledge-graph-based search (Castro Fernandez et al., 2018; Nargesian et al., 2018; Viswanathan et al., 2023). This paper investigates the inverse problem: whether existing dataset metadata can serve as evidence for discovering potential uses of data and representing those uses as graph-supported Machine-Readable Decision Models.</em></p><p><em>We define a Machine-Readable Decision Model (MRDM) as a structured, machine-processable representation of a candidate decision context that explicitly connects a potential use with the decision it supports, expected value, business objectives, information needs, reusable data products, available datasets, provenance, and identified information gaps. MRDMs are represented within a knowledge graph, allowing entities to be shared across decision contexts and supporting traversal between inferred information demand and existing data supply. The construct builds on established work in goal-oriented requirements engineering, information-needs modeling, data-product design, dataset knowledge graphs, and decision knowledge graphs rather than treating these individual elements as novel (Horkoff et al., 2019; van den Homberg et al., 2018; Hao et al., 2021; Färber &amp; Lamprecht, 2021; Hasan et al., 2026).</em></p><p><em>The implemented artifact operates on dataset metadata obtained through source-specific ingestion adapters and combines deterministic processing with LLM-supported semantic inference. A complete downstream execution across 65 heterogeneous open-data portal snapshots processed 134,480 dataset metadata records and generated 538 candidate objectives, 1,405 candidate use cases, 1,091 KPIs, 5,333 information needs, 661 candidate data products, 130,514 graph nodes, 174,394 typed relationships, and 20,422 explicit information gaps. All 65 portal runs completed on their first attempt, all 174,394 relationships contain artifact-level provenance references, and checksum manifests for all 65 runs passed independent verification. The source catalogs reported 341,634 datasets in total, but ten inputs were capped. The study therefore does not claim that all reported source datasets were ingested or processed. All generated data products remain candidates requiring human review.</em></p><p><em>The study follows Design Science Research and distinguishes technical execution from semantic validation. The completed heterogeneous execution establishes technical feasibility, structural outputs, provenance coverage, and reproducible run artifacts at multi-portal scale. The next empirical stage evaluates the semantic validity of generated MRDMs and individual graph relationships and examines how source metadata richness and quality affect model validity. The expected contribution is empirically grounded design knowledge concerning whether existing data supply can be transformed into explicit, traceable, and machine-processable hypotheses about potential organizational use.</em></p><p><strong>Keywords:</strong> machine-readable decision models, metadata, data ecosystems, information needs, data products, knowledge graphs, use discovery, supply-first discovery, design science research</p><h3>Introduction</h3><p>Organizations increasingly possess more data than they know how to use. Data catalogs, open-data portals, enterprise platforms, and data-product architectures make datasets easier to inventory, describe, and access. Standardized metadata approaches improve findability, accessibility, interoperability, and reuse, while catalog vocabularies such as DCAT support dataset discovery and metadata aggregation across catalogs (Wilkinson et al., 2016; W3C, 2024). Discoverability, however, does not answer a different question: what could an organization do with the data it already has?</p><p>A catalog containing hundreds or thousands of datasets can describe titles, publishers, classifications, geographic coverage, temporal properties, schemas, and other metadata. It does not necessarily explain which decisions those datasets could support, what organizational value those decisions might create, what additional information would be required, or which reusable data products should bring the necessary information together. Existing research addresses important parts of this problem. Semantic data-discovery systems support navigation and retrieval across heterogeneous data assets (Castro Fernandez et al., 2018), table-search approaches identify related or combinable datasets from large collections (Nargesian et al., 2018), dataset recommendation systems such as DataFinder begin with a natural-language description of a research need and retrieve relevant datasets (Viswanathan et al., 2023), and dataset knowledge graphs integrate metadata to improve representation and discovery (Färber &amp; Lamprecht, 2021). These approaches primarily improve the path from an expressed need, query, task, or data-discovery operation toward relevant data.</p><p>The practical motivation for the present study originates from the opposite direction. In work with open-data platforms, data owners repeatedly raised a similar question: given the data already available, what valuable uses can be developed from it? In these settings, the data assets already existed, while their potential uses were unclear. The requirement was therefore not solely to locate data for an established business objective. It was to support systematic discovery of possible uses from existing data supply. We describe this as a supply-first discovery problem, in which the conventional direction from business objective to use case, information need, and data is inverted so that existing data becomes the starting point for asking what the available supply might support.</p><p>This inversion matters because prior research has demonstrated that information demand and available data frequently do not align. Van den Homberg et al. (2018a), for example, explicitly mapped disaster-response information needs against available datasets and identified information gaps. Their work demonstrates why information needs should remain conceptually distinct from datasets and why unsupported needs should remain visible rather than disappear when matching data cannot be found.</p><p>The artifact examined in this paper addresses the supply-first problem by starting with existing dataset metadata and asking what potential decisions and uses that metadata provides evidence for. The output is not intended to be a list of textual use-case suggestions. The artifact constructs richer candidate decision contexts containing the potential use, decision supported, expected value, information requirements, candidate data products, available supporting datasets, objectives, provenance, confidence information, and gaps in the information required to realize the proposed use. We describe this artifact as a graph-supported Machine-Readable Decision Model (MRDM), in which a use case is one component of the model rather than the complete artifact.</p><p>The study addresses two research questions. RQ1 asks: How can heterogeneous dataset metadata be transformed into graph-supported Machine-Readable Decision Models representing potential uses of existing data? RQ2 asks: How do the richness and quality of source metadata affect the validity of the resulting Machine-Readable Decision Models? The second question is grounded in an established metadata-quality problem. Large-scale studies of open-data portals have found substantial variation in metadata quality and completeness across independently operated portals, with consequences for discovery and reuse (Neumaier et al., 2016; Kubler et al., 2018). Metadata quality is therefore treated as an explanatory condition of the artifact rather than merely an implementation limitation.</p><p>The study makes three intended contributions. First, it develops the MRDM as a candidate design construct for representing potential data use in a machine-processable form. Second, it provides a graph-supported design connecting potential decisions and expected value with information requirements and existing data supply. Third, it evaluates whether such models can be derived from heterogeneous catalog metadata and identifies metadata-related boundary conditions affecting their validity. The study does not assume that MRDM is novel merely because this terminology is introduced here. Goal models, decision models, decision knowledge graphs, information-needs models, data products, dataset knowledge graphs, semantic data discovery, and AI-assisted semantic inference all have established research foundations. The candidate contribution concerns their configuration around a different starting point: existing data supply.</p><h3>2. From Use Cases to Machine-Readable Decision Models</h3><p>The term use case is insufficient to describe the artifact generated by the implemented system. A conventional use-case description can identify a potential application of data, whereas the artifact examined here contains a broader decision context. For example, one candidate generated by the current implementation is Geographic Analysis of Commercial Activity. The artifact does not stop at a title and description. It represents a decision concerning where resources should be directed based on commercial potential and business-support needs, associates expected value with local economic growth and employment, identifies information requirements including current business statistics by geography, market-demand data, and local industry trends, and connects this context with a candidate data product, a business objective, domains, confidence information, and the source dataset from which the discovery process originated.</p><p>This distinction matters because the original dataset does not necessarily provide all the information required by the candidate use. A dataset can provide enough semantic evidence to suggest a potentially valuable decision while supplying only part of the information necessary to make that decision. The artifact must therefore represent both what exists and what would be required.</p><p>For this study, a Machine-Readable Decision Model is defined as a structured, machine-processable representation of a candidate decision context that explicitly connects a potential use with the decision it supports, expected value, business objectives, information needs, reusable data products, available datasets, provenance, and identified information gaps. The model is machine-readable because its components and relationships are represented structurally rather than existing only as natural-language prose. It is explicitly a candidate because a metadata-derived model represents a hypothesis about potential use rather than an established organizational requirement. It also separates information demand from current data supply, allowing the information required to realize a potential decision to exceed what the originating dataset currently provides.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*8x4XIjamc60EAiTtYJVjLg.png" /><figcaption>Example of MRDM card generated from Ajman portal datasets.</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*xlkxgzP20gu7rvST-IuRZQ.png" /><figcaption>Example visualized graph representation of one Machine-Readable Decision Model manifestation. Source Ajman open data portal.</figcaption></figure><p>This separation has foundations in information-needs research. Van den Homberg et al. (2018a) demonstrate that explicit information requirements can be mapped to available datasets while retaining unmet requirements as gaps. Their subsequent work further considers selection of datasets against information coverage, quality, and cost (van den Homberg et al., 2018b). The Data Product Canvas provides a complementary perspective. Hasan et al. (2026) connect data-product design with consumer needs, business value, desirability, feasibility, and viability, establishing that data products should be considered in relation to intended use and value rather than solely as technical data assets. MRDM differs in purpose from both approaches because it attempts to construct a candidate decision context computationally from existing metadata and represent the resulting relationships in machine-processable form. It does not replace human requirements elicitation or data-product design.</p><h3>3. Supply-First Discovery</h3><p>The construction direction differs from the conceptual demand-to-data ordering of the resulting graph. The implemented discovery process begins with existing data supply and proceeds from dataset metadata through normalization, analysis and semantic inference to a candidate MRDM, graph integration, evaluation, and human selection. This distinction is fundamental because the graph describes semantic relationships represented after discovery; it does not imply that discovery itself begins with a Business Objective.</p><p>Dataset-discovery research generally assumes that some expression of demand already exists. DataFinder, for example, receives a natural-language research description and recommends relevant datasets (Viswanathan et al., 2023). Semantic data-discovery systems similarly support users in locating appropriate data across complex data environments (Castro Fernandez et al., 2018). Information-needs mapping starts with identified requirements and determines which datasets satisfy them (van den Homberg et al., 2018a). The present artifact investigates the inverse starting point. In demand-first discovery, a known objective, task, or information requirement leads toward data supply. In supply-first discovery, known data supply is used as evidence from which potential information demand and candidate decision contexts are inferred.</p><p>The distinction is directional rather than absolute. Once an MRDM has been generated, the resulting graph supports demand-to-supply traversal as well. The contribution proposition therefore concerns the construction process rather than an assertion that demand-to-data analysis is absent from prior research.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/629/1*x0T9e7K-mh1YnrV5o458mA.png" /><figcaption>Figure 1. Supply-first discovery and MRDM construction. Metadata from supported source platforms is transformed through source adapters into a common representation, processed through deterministic functions and LLM-supported semantic inference, and represented as a candidate MRDM. Human evaluation and selection form the boundary between discovery and subsequent development.</figcaption></figure><p>Human selection creates a clear boundary for the artifact. Candidate MRDMs represent possibilities that a user can inspect and select for further investigation or implementation. Selection does not establish that the proposed decision will create the expected value. It establishes that a candidate has been judged sufficiently relevant to proceed to a subsequent stage.</p><h3>4. Graph-Supported Machine-Readable Decision Models</h3><p>An individual MRDM represents one candidate decision context, but the implemented artifact does not treat these models as isolated documents. MRDMs are supported by a knowledge graph containing constructs including Business Objective, Use Case, Decision Supported, Expected Value, Information Need, Data Product, and Dataset, together with attributes concerning provenance, confidence, status, domains, and information gaps. The central demand-to-data relationship connects Business Objective to Use Case, Use Case to Information Need, Information Need to Data Product, and Data Product to Dataset.</p><p>Information Need occupies a central position between potential use and data supply because it represents what must be known, measured, observed, or inferred for a candidate use to operate. Decision Supported and Expected Value enrich this structure by making explicit why the use might matter and which decision the information is intended to support.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/674/1*nuxhcyxX-3JE-b9Y20YV3w.png" /><figcaption>Figure 2. Graph-supported Machine-Readable Decision Model. An MRDM connects a candidate use and its business context with the information required to support a decision, candidate data products, and available datasets. Information needs without adequate supporting supply remain represented as explicit gaps. Entities can be shared across multiple MRDMs.</figcaption></figure><p>Graph representation becomes significant when multiple MRDMs exist. The same Information Need can recur across several decision models, a reusable Data Product can address needs associated with multiple candidate uses, a Dataset can support several products, and multiple candidate uses can contribute to the same Business Objective. The graph therefore represents an ecosystem of potential decisions rather than a collection of independent generated documents.</p><p>Knowledge-graph representation itself is established prior art. Hao et al. (2021) developed a Decision Knowledge Graph for engineering design, converting textual decision-support problems into semantic graph structures and supporting navigation of decision-related knowledge. Wang et al. (2025) extend this direction through a Concept-Decision-Knowledge model and knowledge graph for product conceptual design, extracting decision-relevant knowledge from heterogeneous textual resources. These studies demonstrate that decision knowledge can be represented and processed through graphs, but they begin from explicit decision problems, requirements, or domain knowledge rather than from a catalog of existing datasets whose potential uses remain unknown.</p><p>Dataset knowledge graphs provide another neighboring stream. Färber and Lamprecht (2021) constructed the Data Set Knowledge Graph by integrating metadata about scholarly datasets and linking them to publications and other research entities. Their work demonstrates the value of semantic integration across dataset metadata but does not represent the business decision, expected value, information-demand, and data-product structures investigated here. MRDM therefore sits between decision-oriented and dataset-oriented knowledge graphs while introducing supply-first generation as the mechanism connecting them.</p><h3>5. Metadata Ingestion and Normalization</h3><p>The current implementation deliberately operates on dataset metadata rather than dataset contents. Metadata is obtained from two currently supported portal families, CKAN-based catalogs and Huwise-based catalogs, each of which exposes source-specific metadata through REST APIs. Source-specific adapters transform this metadata into a common internal representation before candidate discovery takes place.</p><p>Metadata normalization across heterogeneous open-data portals is established practice. Neumaier et al. (2016), for example, assessed metadata quality across more than 260 open-data portals and normalized metadata originating from different portal technologies into a common representation for cross-portal comparison. Their work demonstrates both the feasibility and necessity of separating portal-specific metadata acquisition from downstream analysis. The present architecture follows the same separation principle but uses the normalized representation as input to candidate decision-context discovery.</p><p>CKAN and Huwise should therefore not be understood as conceptual requirements of MRDM. They are the two source families for which ingestion adapters currently exist. The architectural requirement is broader: a source must expose sufficiently descriptive metadata that an adapter can transform into the common representation expected by the discovery pipeline. DCAT provides a related standard foundation for interoperable descriptions of datasets and data services (W3C, 2024). The present artifact does not require DCAT specifically, but follows the broader principle of separating source-specific metadata from a normalized representation suitable for subsequent processing.</p><h3>6. Empirical Research Environment and Architectural Extensibility</h3><p>Open data was selected as the initial empirical environment for practical and methodological reasons. The study requires access to large numbers of independently maintained datasets and metadata records spanning different organizations, domains, catalog sizes, and metadata practices. Open-data ecosystems provide such an environment without requiring privileged access to confidential enterprise catalogs.</p><p>Prior research demonstrates the heterogeneity available in this environment. Neumaier et al. (2016) studied hundreds of open-data portals and large numbers of dataset descriptions across different platform technologies. Kubler et al. (2018) likewise examined metadata quality across a large international collection of open-data portals. These studies demonstrate that open-data ecosystems provide substantial variation in metadata practices and quality, making them suitable for examining metadata-dependent behavior.</p><p>Open data is therefore the empirical research environment rather than the conceptual scope of the artifact. Enterprise data catalogs, internal data-product portfolios, data marketplaces, data spaces, and other metadata repositories can face the same supply-first problem. The architecture separates the metadata source, source-specific adapter, common metadata representation, discovery pipeline, MRDM construction, and knowledge-graph representation. Additional adapters can therefore map metadata from other catalog technologies into the common representation.</p><p>This establishes architectural extensibility, not empirical generalizability. The fact that another metadata source can be connected through an adapter does not establish that MRDM discovery will perform equivalently in that environment. Performance in enterprise catalogs or other non-open-data environments remains to be tested empirically.</p><h3>Metadata as Evidence and Boundary Condition</h3><p>The decision to operate on metadata creates both an advantage and a constraint. The advantage is reuse. Organizations already invest resources in cataloging and describing data, and a metadata-driven discovery process can operate on these descriptions without downloading, profiling, and processing every underlying dataset before potential uses are explored. This is consistent with the wider role assigned to metadata in FAIR data management, where rich metadata supports findability and reuse independently of immediate access to underlying data (Wilkinson et al., 2016).</p><p>The constraint is that the artifact only sees what the metadata communicates. Sparse, generic, outdated, incomplete, or ambiguous metadata provides weaker evidence for semantic interpretation than metadata containing clear descriptions, classifications, field information, geographic scope, temporal coverage, publisher context, and other meaningful properties. This problem is established in open-data research. Neumaier et al. (2016) found substantial variation in metadata quality across open-data portals and developed automated quality-assessment measures covering dimensions relevant to dataset discovery and reuse. Kubler et al. (2018) similarly documented variation in metadata quality across a large international sample of portals.</p><p>The implication for MRDM is direct: metadata is not simply input; it is evidence. The validity and specificity of a metadata-derived MRDM are therefore expected to vary with the semantic richness and quality of the source metadata. The study treats this relationship as an empirical question rather than merely a limitation. Variation across heterogeneous catalogs provides an opportunity to investigate whether measurable metadata characteristics are associated with variation in MRDM validity.</p><p>A future extension could supplement metadata with analysis of the underlying dataset contents. Such an extension could provide additional evidence when catalog metadata is sparse by examining schemas, field names, distributions, values, or other properties of the data itself. This would, however, alter the design substantially. Content-level analysis introduces requirements concerning data access, transfer, format handling, profiling, privacy, computational resources, and potentially domain-specific interpretation. The present study therefore deliberately evaluates metadata-driven discovery first. Its purpose is to establish how far organizations can proceed by reusing metadata they already maintain before introducing the additional cost and complexity of dataset-content analysis.</p><p>This design choice also creates a useful empirical boundary. Poor performance under semantically sparse metadata has a different interpretation from poor performance under rich metadata. The former can indicate a limitation of the available evidence, while the latter provides stronger reason to question the inference or representation mechanism. Evaluation should therefore characterize the metadata available for each evaluated MRDM rather than treating all source records as equivalent.</p><h3>8. Hybrid Generation Process</h3><p>The implemented pipeline combines deterministic processing with LLM-supported semantic inference. Deterministic processing is used where operations can be specified explicitly, including source ingestion, normalization, filtering, structural transformations, and parts of matching and graph construction. LLM-based processing supports operations requiring semantic interpretation and candidate generation. This division allows predictable operations to remain rule-based while using language models where interpretation of heterogeneous natural-language metadata is required.</p><p>The role of the LLM must be distinguished from the research contribution. The language model does not establish the truth of an inferred use case, information need, objective, or relationship. It generates candidate semantic interpretations from the evidence available in metadata. Those candidates are subsequently normalized, linked, represented in the graph, and subjected to evaluation. Recent work on LLM-supported ontology and knowledge-graph engineering similarly identifies concept extraction, semantic alignment, relation generation, and ontology population as promising applications while emphasizing the continuing need for systematic evaluation of generated structures (Garijo et al., 2025).</p><p>The conceptual processing sequence is therefore Normalized Metadata → Deterministic Processing → LLM-Supported Candidate Generation → Normalization and Linking → MRDM → Knowledge Graph → Evaluation. This hybrid architecture also allows the study to investigate failures more precisely. An invalid relationship can originate from insufficient source metadata, deterministic transformation, model-based semantic inference, normalization, matching, or graph construction. Treating the complete pipeline as an undifferentiated AI system would make those sources of error difficult to distinguish.</p><p>The reproducible research snapshot will therefore document the division between deterministic and model-based operations together with prompts, model identifiers, configurations, thresholds, timestamps, source identifiers, and other relevant parameters. The purpose is not merely operational reproducibility. These records provide provenance needed to determine how a candidate semantic assertion was produced.</p><h3>9. Relationship to Existing Research and Candidate Research Gap</h3><p>The proposed artifact intersects several established research streams. The scientific contribution cannot therefore rest on any individual component of the artifact. Business objectives, use cases, information requirements, data products, machine-readable decision structures, knowledge graphs, metadata normalization, dataset discovery, and LLM-supported semantic inference all have substantial prior foundations.</p><p>Goal-Oriented Requirements Engineering provides mature approaches for representing stakeholder goals and connecting organizational intentions with system requirements. The systematic mapping study by Horkoff et al. (2019) demonstrates the breadth of goal-oriented approaches and their role in connecting stakeholder intentions, requirements, and system behavior. Business Objective and Use Case are therefore established concepts rather than contributions of the present study. The difference investigated here concerns how candidate demand-side structures originate. Instead of beginning with stakeholder goals that have already been elicited, the artifact attempts to derive candidate decision contexts from descriptions of existing data supply.</p><p>Information-needs research provides a second foundation. Van den Homberg et al. (2018a) explicitly map information needs against available datasets and identify information needs for which available supply is inadequate. Their subsequent quantitative work considers the optimization of data selection against information coverage, data quality, and cost (van den Homberg et al., 2018b). These studies establish both the value of representing information demand independently from datasets and the analytical importance of unmet demand. MRDM adopts this separation but reverses the initial direction of discovery. Rather than beginning with elicited information needs, it attempts to infer candidate decision contexts and associated information needs from available data descriptions.</p><p>Data-product research provides a third foundation. Hasan et al. (2026) propose the Data Product Canvas for designing data products around sustained enterprise value. Their framework connects business problems, use cases, consumer information needs, desirability, feasibility, viability, and the resources required to create data products. This work is close to MRDM at the conceptual level because it connects data products with use and value. The distinction lies in the artifact form and construction process. The Data Product Canvas is a collaborative design artifact used to reason about data products, whereas the present study investigates computational generation of machine-readable candidate decision contexts from existing metadata.</p><p>Dataset discovery and recommendation constitute a fourth neighboring stream. AURUM constructs a discovery structure over heterogeneous enterprise data and supports exploratory search across organizational data assets (Castro Fernandez et al., 2018). Table Union Search identifies datasets that can be combined with a query table in large open-data collections (Nargesian et al., 2018). DataFinder uses natural-language descriptions of research needs to retrieve relevant scientific datasets and demonstrates the value of semantic retrieval over conventional dataset search (Viswanathan et al., 2023). These systems substantially advance data discovery, but their principal direction remains from a query, task, example, or known data requirement toward relevant data assets.</p><p>Knowledge graphs provide a fifth foundation. Färber and Lamprecht (2021) demonstrate that metadata describing datasets can be integrated into a Data Set Knowledge Graph that connects research datasets with publications and related scholarly entities. Their work establishes that heterogeneous dataset metadata can support a richer semantic representation than isolated catalog records. Decision-oriented knowledge graphs address the opposite side of the problem. Hao et al. (2021) represent engineering decision-support problems as a Decision Knowledge Graph, while Wang et al. (2025) construct a Concept-Decision-Knowledge graph to support product conceptual-design decisions. Together, these streams establish both dataset-centered and decision-centered graph representations. Neither graph representation by itself constitutes the proposed contribution.</p><p>Metadata-quality research provides a sixth foundation and an important boundary condition. Neumaier et al. (2016) demonstrate automated cross-portal metadata-quality assessment and show substantial variation among open-data portals. Kubler et al. (2018) likewise examine open-data metadata quality across an international portal population. These findings are important for the present artifact because MRDM generation relies on metadata as its primary evidence. The quality of semantic inference should therefore be evaluated in relation to the quality of the descriptions from which inference is performed.</p><p>The emerging use of LLMs for semantic and ontology-engineering tasks provides an implementation foundation rather than the conceptual contribution. The literature summarized by Garijo et al. (2025) indicates that language models can support activities such as concept extraction, semantic alignment, and knowledge-graph construction while also raising concerns about hallucination, consistency, and evaluation. MRDM uses LLM-supported semantic inference within a constrained pipeline, but the contribution does not depend on a claim that LLM-based metadata interpretation is itself new.</p><p>Across the literature examined, we did not identify an equivalent artifact combining the complete configuration studied here: starting from heterogeneous dataset metadata, deriving candidate decision contexts, expanding those contexts into explicit information needs that may exceed current supply, linking those needs to reusable data products and datasets, retaining unmet demand as explicit gaps, and integrating multiple candidate contexts through shared graph entities. The contribution claimed in this paper therefore concerns this supply-first configuration rather than any of its individual components.</p><p>This conclusion is deliberately bounded by the literature examined. It does not establish that no equivalent work exists. It establishes a candidate research gap that the present study positions against the closest identified streams and that should continue to be tested as the research progresses.</p><h3>10. Research Method</h3><p>The study follows Design Science Research because it investigates an implemented artifact intended to address an observed information-systems problem and seeks transferable knowledge through systematic evaluation of that artifact. Hevner et al. (2004) position design science around the construction and evaluation of artifacts that address relevant organizational problems. Peffers et al. (2007) provide a Design Science Research Methodology spanning problem identification, definition of solution objectives, design and development, demonstration, evaluation, and communication.</p><p>The relevance problem originates from repeated practical situations in which data owners possess substantial cataloged data but lack structured knowledge about potential uses. The rigor base draws from requirements engineering, information-needs research, data-product design, metadata and data-discovery research, decision modeling, and knowledge graphs. The design cycle has already produced an operational artifact. The present research therefore focuses on formalization, reproducibility, independent evaluation, refinement, and extraction of transferable design knowledge rather than initial software construction.</p><p>This distinction is methodologically important. Working software demonstrates technical feasibility but does not establish semantic validity. A system can successfully generate a candidate use while that use remains poorly supported by source metadata. Likewise, a graph can contain a technically valid path between an objective and dataset while one or more semantic relationships composing that path are invalid. Evaluation therefore separates technical operation, structural graph properties, and semantic validity.</p><h3>11. Standards Boundary and Artifact Provenance</h3><p>The implementation uses concepts and structures from the Open Data Products standards family maintained under the Linux Foundation AI &amp; Data Foundation. The lead researcher founded and maintains this standards work. The standards are therefore treated as pre-existing technical infrastructure and artifact provenance rather than independent evidence supporting the research claims.</p><p>This separation is important because the Information Need layer examined in the present study currently resides in the application graph and is not preserved in the existing Open Data Product Graph standards export. The MRDM is therefore not presented as an existing Linux Foundation standard. It is a research construct implemented partly using standards-based infrastructure.</p><p>The relationship between research and standardization should consequently proceed from evaluation toward possible standards evolution rather than in the opposite direction. Existing standards can provide implementation infrastructure, but their existence does not validate the MRDM construct. If empirical evaluation identifies transferable structures or design principles, those findings can subsequently inform standardization.</p><h3>12. Current Technical Instantiation and Empirical Boundary</h3><p>The artifact has now been executed through the complete downstream pipeline across 65 heterogeneous open-data portal snapshots. All 65 portal runs completed on their first attempt. The retained inputs contained 134,480 analyzed dataset metadata records, from which the pipeline generated 538 candidate business objectives, 1,405 candidate use cases, 1,091 KPIs, 5,333 information needs, and 661 candidate data products. The resulting graphs contain 130,514 nodes, 174,394 typed relationships, and 20,422 explicit information gaps.</p><p>The source catalogs reported 341,634 datasets in total. This number represents the dataset counts reported by the source catalogs and must be distinguished from the 134,480 records processed by the pipeline. Ten portal inputs were capped during acquisition. The present study therefore does not claim that all 341,634 reported source datasets were ingested or analyzed. The empirical unit for the completed execution is the set of 134,480 retained and analyzed metadata records across the 65 portal snapshots.</p><p>The completed execution also provides evidence concerning technical reliability and provenance. All 65 portal snapshots completed downstream processing without requiring a rerun. All 174,394 generated relationships contain artifact-level provenance references, and checksum manifests for all 65 portal runs passed independent verification. The complete execution required 950 model calls, consumed 4,812,798 tokens, and took approximately three hours and 27 minutes. These measurements characterize the completed artifact execution rather than its semantic quality.</p><p>One portal, Argentina, produced no candidate Data Product satisfying the pipeline’s acceptance criteria. The absence of a DataProduct node was retained in the resulting artifact rather than repaired or suppressed and constitutes the sole warning in the 65-portal execution. This outcome is analytically relevant because successful pipeline execution does not require every portal to produce every artifact type. An absence can itself represent an observable result of the generation and filtering process.</p><p>The generated semantic structures remain candidates. In particular, the 661 generated Data Products have not yet been established as valid data products through independent human assessment. The same distinction applies to generated objectives, use cases, information needs, and their semantic relationships. The completed execution therefore establishes heterogeneous full-pipeline operation, graph construction, explicit gap representation, and provenance coverage at the scale reported here. It does not establish the semantic validity or organizational usefulness of the generated MRDMs.</p><p>This distinction defines the current empirical boundary of the study. Technical feasibility and heterogeneous full-pipeline execution have been demonstrated across the 65 retained portal snapshots. Independent semantic validation, relationship-level assessment, and analysis of the effect of metadata quality on MRDM validity constitute the next empirical stage.</p><h3>13. Evaluation Design</h3><p>The completed 65-portal execution changes the role of evaluation in the study. Technical feasibility across heterogeneous portal snapshots no longer constitutes planned evaluation. It is an observed property of the current artifact execution. The remaining evaluation therefore concentrates on whether the generated semantic structures are defensible and on the conditions under which their quality varies.</p><p>The evaluation separates three levels: complete MRDM validity, relationship-level semantic validity, and metadata sensitivity. This separation is necessary because successful pipeline execution does not imply semantic validity. Likewise, a candidate MRDM that appears coherent as a whole can contain individual relationships that are weak or unsupported by the source evidence.</p><h4>13.1 MRDM-Level Validity</h4><p>Independent reviewers will evaluate a stratified sample of generated MRDMs against the source metadata and provenance available for each candidate. The evaluation will assess whether the available metadata provides reasonable evidence for the proposed use, whether the proposed decision is coherent with that use, whether the stated expected value is reasonably connected to the decision, whether the identified Information Needs are relevant to realizing the use, whether the candidate Data Product forms a coherent reusable information resource, and whether the proposed Business Objective has a defensible relationship with the candidate use.</p><p>The purpose of this evaluation is not to determine whether reviewers prefer a generated idea or believe an organization should implement it. An MRDM is a candidate hypothesis about potential use. The evaluation therefore asks whether the semantic interpretation represented by the model is defensible given the evidence from which it was generated.</p><p>Sampling should account for the heterogeneous structure of the completed execution. The 65 portal snapshots differ in size, metadata characteristics, generated artifact counts, and graph structure. A stratified design will therefore be used to avoid allowing the largest portals to dominate the evaluation and to ensure that candidate structures from different metadata environments are represented.</p><h4>13.2 Edge-Level Semantic Validity</h4><p>The completed graphs contain 174,394 typed relationships. All of these relationships now contain artifact-level provenance references, making relationship-level evaluation possible. A stratified sample will be drawn from the principal semantic relationship types, including Business Objective to Use Case, Use Case to Information Need, Information Need to Data Product, and Data Product to Dataset.</p><p>Independent reviewers will assess whether each sampled relationship is semantically supported by the available evidence. This local evaluation is necessary because complete-model assessment can conceal individual weak links. For example, a plausible use case can contain an unsupported Information Need, or a coherent Data Product can be linked to a dataset whose metadata provides insufficient evidence for that relationship.</p><p>Where the evaluation design and response scale support it, agreement among reviewers will be measured. Reviewer disagreement will not automatically be treated as noise. Persistent disagreement can indicate ambiguous metadata, underspecified relationship semantics, or insufficient evaluation criteria and can therefore provide evidence for refinement of the artifact.</p><h4>13.3 Metadata Sensitivity</h4><p>The third evaluation examines whether MRDM validity varies systematically with characteristics of the source metadata. This question is particularly important because the artifact deliberately uses metadata rather than dataset contents as its primary evidence. Prior studies have documented substantial variation in metadata quality across open-data portals (Neumaier et al., 2016; Kubler et al., 2018).</p><p>Metadata profiles will therefore be constructed for evaluated records and portal snapshots. Candidate dimensions include description completeness, semantic specificity, classification availability, temporal and geographic context, field-level information where available, and other characteristics that can be measured consistently across the retained inputs. These characteristics will then be compared with MRDM-level and relationship-level evaluation outcomes.</p><p>This analysis will help distinguish different failure mechanisms. Lower validity associated with sparse metadata would identify a boundary condition of metadata-driven discovery. Weak validity despite semantically rich metadata would provide stronger evidence of problems in semantic inference, normalization, matching, or the MRDM representation itself. Metadata quality is therefore treated as an explanatory variable rather than as a generic limitation stated after evaluation.</p><p>The completed execution also provides several naturally occurring cases for qualitative analysis. The 20,422 explicit gaps provide evidence about where generated information demand is not connected to sufficient supply. The Argentina portal, which completed processing but produced no acceptable candidate Data Product, provides a boundary case in which absence was retained rather than repaired. Such cases will complement aggregate validity measures by helping explain how the artifact behaves when the available evidence does not support generation of the expected downstream structure.</p><h3>14. Structural Graph Evaluation</h3><p>The completed 65-portal execution provides a substantially larger basis for structural analysis than the earlier single full-chain instantiation. Across the retained portal snapshots, the generated graphs contain 130,514 nodes, 174,394 typed relationships, and 20,422 explicit information gaps. Structural evaluation will use these generated artifacts to characterize how the proposed MRDM representation behaves across heterogeneous metadata environments.</p><p>The analysis will examine graph connectivity, reachability, realized relationship cardinalities, entity reuse, provenance coverage, and the distribution of explicit gaps. Particular attention will be given to whether Information Needs, Data Products, Datasets, and Business Objectives are reused across candidate decision contexts. This addresses one of the principal reasons for representing MRDMs through a graph rather than as independent documents. If semantically meaningful entities recur across multiple candidate contexts, the graph provides an explicit mechanism for representing shared information demand and reusable supply. If such reuse is limited, the assumed value of graph integration must be reconsidered.</p><p>The 20,422 explicit gaps provide a second structural property for analysis. A gap represents information demand for which the generated graph does not contain sufficient supporting supply under the implemented matching and acceptance rules. Gap counts alone do not establish that information is objectively unavailable. A gap can result from genuinely absent supply, incomplete portal coverage, capped source inputs, insufficient metadata, failed matching, or the generation of an unsupported Information Need. Gap analysis must therefore distinguish structural observations from semantic conclusions. Independent evaluation of sampled gaps will be needed before they can be interpreted as evidence about deficiencies in the underlying data ecosystems.</p><p>The completed run also allows provenance coverage to be treated as an observed structural property rather than a future requirement. All 174,394 generated relationships contain artifact-level provenance references. Structural analysis will examine whether this provenance is sufficient to trace evaluated relationships back through the generation pipeline and to the relevant source artifacts. Provenance coverage is particularly important for MRDM because generated relationships represent candidate semantic assertions rather than established facts.</p><p>Cross-portal variation will provide an additional level of structural analysis. The 65 portal snapshots differ in input size and generated graph composition. Comparing distributions of objectives, use cases, information needs, candidate data products, relationships, and gaps across portals can identify structural patterns that warrant subsequent semantic investigation. Such differences should not be interpreted directly as differences in portal quality or value because input sizes vary and ten source inputs were capped. They instead provide descriptive evidence about how the artifact behaves across heterogeneous environments.</p><p>The Argentina result provides a specific structural boundary case. The portal completed the pipeline but produced no candidate Data Product satisfying the implemented acceptance criteria. The absence of a DataProduct node was retained and constitutes the sole warning in the completed execution. This case will be examined alongside the broader graph population because it demonstrates that the artifact can preserve absence rather than forcing every execution into a structurally complete decision model.</p><p>Prior dataset and decision knowledge graphs establish the general value of graph structures for connecting metadata and decision-related knowledge (Färber &amp; Lamprecht, 2021; Hao et al., 2021; Wang et al., 2025). The present structural evaluation addresses a narrower proposition: whether graph integration provides meaningful reuse, traceability, and gap representation when candidate decision contexts are derived supply-first from heterogeneous dataset metadata.</p><h3>15. Reproducibility and Provenance</h3><p>The completed 65-portal execution provides concrete evidence for reproducibility and provenance that was previously described as planned work. All 65 retained portal snapshots completed the full downstream pipeline on their first attempt, and checksum manifests for all 65 runs passed independent verification. The resulting artifacts therefore provide a consistent technical basis for subsequent structural and semantic evaluation.</p><p>Provenance is particularly important because the graph contains generated semantic assertions rather than only source metadata. The completed execution produced 174,394 typed relationships, and all 174,394 relationships contain artifact-level provenance references. This allows generated relationships to be traced to the artifacts involved in their construction and provides a basis for examining how individual semantic assertions were produced during independent evaluation.</p><p>The execution itself also produced measurable operational provenance. Processing the 134,480 retained metadata records required 950 model calls and consumed 4,812,798 tokens. End-to-end execution took approximately three hours and 27 minutes. These measurements characterize the computational execution of the artifact and provide a baseline for subsequent replications, refinements, and comparisons. They should not be interpreted as measures of semantic quality or efficiency until suitable comparison conditions have been established.</p><p>The distinction between reported source size and processed input size is retained explicitly in the reproducibility record. The 65 source catalogs reported 341,634 datasets, while the completed pipeline analyzed 134,480 retained records. Ten catalog inputs were capped. Reproduction of the reported experiment therefore concerns the retained portal snapshots and their 134,480 analyzed records rather than the complete source populations reported by those catalogs.</p><p>The run also preserves exceptional outcomes rather than silently correcting them. The Argentina portal completed processing but produced no candidate Data Product satisfying the implemented acceptance criteria. This absence remains represented in the artifacts as the sole warning from the 65 runs. Retaining such outcomes is important for reproducibility because manual repair of structurally inconvenient results would make the evaluated artifact differ from the artifact produced by the pipeline.</p><p>The generated semantic entities remain candidates despite the completeness of the technical provenance. Provenance establishes how an artifact or relationship was produced; it does not establish that the semantic assertion is correct. The 661 candidate Data Products, 1,405 candidate use cases, 5,333 Information Needs, 538 candidate objectives, and their associated relationships therefore remain subject to independent semantic evaluation.</p><p>For the research snapshot used in formal evaluation, the provenance record should preserve the source snapshots, checksum manifests, code version, configurations, prompts, model identifiers, thresholds, generation parameters, and timestamps associated with the completed run. This provides the connection between technical reproducibility and scientific evaluation: an independently assessed MRDM or graph relationship should be traceable to the exact artifact state and processing conditions under which it was generated.</p><h3>16. Remaining Work Before ECIS 2027</h3><p>The status of the study has changed substantially following completion of the 65-portal execution. Full downstream processing across heterogeneous portal snapshots is no longer planned work. The artifact has processed 134,480 retained metadata records across all 65 portal snapshots and produced the complete downstream structures required for subsequent evaluation. The remaining work therefore shifts from demonstrating technical feasibility toward establishing semantic validity and extracting defensible design knowledge.</p><p>The first remaining activity is independent semantic evaluation of the generated artifacts. A stratified sample of complete MRDMs will be assessed against the metadata and provenance from which they were generated. This evaluation will examine the defensibility of candidate use cases, decisions supported, expected value, information needs, candidate data products, and business objectives. A separate edge-level evaluation will assess sampled Business Objective to Use Case, Use Case to Information Need, Information Need to Data Product, and Data Product to Dataset relationships. The completed provenance layer makes this relationship-level assessment technically feasible because all 174,394 generated edges now contain artifact-level provenance references.</p><p>The second activity is metadata-sensitivity analysis. The 65 portal snapshots provide heterogeneous metadata environments against which the relationship between metadata characteristics and semantic validity can be examined. Metadata profiles will be constructed using characteristics that can be measured consistently across the retained inputs, and these profiles will be compared with MRDM-level and relationship-level evaluation outcomes. This analysis is intended to establish whether metadata richness and quality constitute observable boundary conditions for supply-first discovery rather than merely assumed limitations.</p><p>The third activity is structural analysis of the completed graph population. The 130,514 nodes, 174,394 relationships, and 20,422 explicit gaps provide a basis for examining connectivity, realized cardinalities, entity reuse, gap distributions, provenance coverage, and cross-portal structural variation. Particular attention will be given to whether Information Needs and candidate Data Products recur across multiple decision contexts, because such reuse constitutes one of the principal arguments for graph-supported rather than document-based representation.</p><p>The fourth activity is analysis of exceptional and negative cases. The 20,422 explicit gaps require interpretation because a structural gap can arise from absent data supply, capped source input, insufficient metadata, failed matching, or an unsupported generated Information Need. Similarly, the Argentina portal completed processing but produced no candidate Data Product satisfying the implemented acceptance criteria. Such cases will be retained and examined rather than excluded because they provide evidence about the boundaries and failure behavior of the artifact.</p><p>The fifth activity is refinement of the MRDM semantics and extraction of design knowledge from the evaluation results. Candidate-versus-validated relationship semantics, permitted cardinalities, gap interpretation, provenance requirements, and criteria for accepting generated structures will be refined in response to empirical findings. The objective is to move from an implemented artifact toward transferable design principles concerning supply-first discovery, machine-readable decision contexts, explicit information-demand representation, and graph-supported traceability.</p><p>The remaining work before ECIS 2027 is therefore substantive but no longer primarily developmental. The system has completed heterogeneous full-pipeline execution with verified manifests and relationship-level provenance. The principal research task is now to determine whether the generated semantic structures are valid, under which metadata conditions they are valid, where they fail, and what generalizable design knowledge can be derived from those results.</p><h3>17. Discussion</h3><p>The completed 65-portal execution changes the evidential position of the study. The artifact is no longer supported only by a single complete technical instantiation accompanied by broader preview graphs. The complete downstream pipeline has now operated across all 65 retained heterogeneous portal snapshots, processing 134,480 metadata records and producing the full set of candidate objectives, use cases, KPIs, Information Needs, Data Products, graph relationships, provenance references, and explicit gaps. This provides substantially stronger evidence of technical feasibility across heterogeneous metadata environments. It does not, however, establish the semantic validity of the generated structures.</p><p>The distinction between technical feasibility, semantic validity, and generalizability remains central. All 65 portal runs completed on their first attempt, all checksum manifests passed independent verification, and all 174,394 generated relationships contain artifact-level provenance references. These results demonstrate that the implemented architecture can execute consistently across the retained heterogeneous inputs and produce traceable graph artifacts. They do not establish that the 538 candidate objectives, 1,405 candidate use cases, 5,333 Information Needs, 661 candidate Data Products, or their relationships are semantically correct. That claim requires independent evaluation.</p><p>The scale of the source environment must also be interpreted carefully. The 65 catalogs reported 341,634 source datasets, while the completed pipeline analyzed 134,480 retained metadata records. Ten source inputs were capped. The study therefore demonstrates complete downstream processing of the retained snapshots, not complete ingestion of every dataset reported by every source catalog. Maintaining this distinction is necessary to avoid conflating the scale of the source ecosystems with the scale of the evaluated artifact execution.</p><p>The completed execution strengthens the case for treating metadata quality as part of the research phenomenon. Previous studies have documented substantial variation in metadata quality across open-data portals (Neumaier et al., 2016; Kubler et al., 2018). The current 65-portal population now provides an empirical setting in which this variation can be related to the validity of generated MRDMs. If lower semantic validity is systematically associated with sparse or weak metadata, the findings would identify an important boundary condition for metadata-driven supply-first discovery. If semantically rich metadata still produces weak decision models, the evidence would instead point toward limitations in inference, normalization, matching, or the MRDM design itself.</p><p>The 20,422 explicit gaps provide another important empirical object, but their interpretation requires caution. The Information Need layer deliberately permits generated demand to extend beyond current supply, consistent with prior research demonstrating the value of mapping information needs separately from available datasets (van den Homberg et al., 2018a, 2018b). A generated gap, however, does not by itself establish that required information does not exist. Gaps can result from unavailable supply, capped portal inputs, insufficient metadata, unsuccessful matching, or unsupported generated Information Needs. Their scientific value therefore depends on subsequent semantic evaluation rather than their raw number.</p><p>The Argentina result illustrates why negative outcomes should remain visible. The portal completed the pipeline but produced no candidate Data Product satisfying the implemented acceptance criteria. The absence was retained rather than repaired or suppressed. This behavior is consistent with the broader design principle underlying explicit gaps: the artifact should preserve unsupported or unsuccessful structures rather than force a complete decision model where the available evidence does not support one. The case provides a useful boundary condition for subsequent qualitative analysis.</p><p>The graph itself also remains an empirical proposition. Existing research demonstrates the value of knowledge graphs for dataset representation and decision-related knowledge (Färber &amp; Lamprecht, 2021; Hao et al., 2021; Wang et al., 2025). The completed MRDM graphs now provide sufficient scale to investigate whether graph integration produces comparable value in a supply-first setting. In particular, the study can examine whether Information Needs, candidate Data Products, Datasets, and Business Objectives recur meaningfully across decision contexts and whether this shared structure supports reuse, traceability, and systematic identification of information gaps. If little meaningful reuse occurs, the assumed advantage of graph-supported representation will require reconsideration.</p><p>The role of the LLM should similarly remain bounded. The completed execution required 950 model calls and 4,812,798 tokens, demonstrating that LLM-supported semantic inference can participate operationally in a heterogeneous pipeline at the reported scale. These figures do not provide evidence of semantic quality. Work on LLM-supported ontology and knowledge-graph engineering likewise distinguishes the ability to generate semantic structures from the need to evaluate their correctness and consistency (Garijo et al., 2025). The independent MRDM-level and relationship-level assessments are therefore essential to determining whether the generated semantic structures are defensible.</p><p>The role of human selection is equally important. A generated MRDM does not represent an organizational decision that should be taken. It represents a candidate decision context for inspection. The artifact is intended to expand and structure the space of potential uses available to data owners rather than substitute automated inference for organizational judgment. The 661 generated Data Products consequently remain candidates despite successful generation, graph integration, and provenance capture.</p><p>The wider theoretical implication concerns the relationship between data supply and information demand. Much existing discovery research improves the ability to locate data once a need, query, task, or example has been expressed (Castro Fernandez et al., 2018; Nargesian et al., 2018; Viswanathan et al., 2023). Information-needs research demonstrates how known demand can be mapped against available supply and where gaps between the two become visible (van den Homberg et al., 2018a, 2018b). Decision knowledge graphs demonstrate how articulated decision problems and associated knowledge can be represented computationally (Hao et al., 2021; Wang et al., 2025). Dataset knowledge graphs demonstrate how metadata describing existing datasets can be integrated semantically (Färber &amp; Lamprecht, 2021). The MRDM artifact investigates the connection between these directions by asking whether existing data descriptions can provide sufficient evidence from which candidate decision demand can first be constructed.</p><p>Across the literature examined, we did not identify an equivalent artifact combining the complete configuration studied here: starting from heterogeneous dataset metadata, deriving candidate decision contexts, expanding those contexts into explicit information needs that may exceed current supply, linking those needs to reusable data products and datasets, retaining unmet demand as explicit gaps, and integrating multiple candidate contexts through shared graph entities. The contribution claimed in this paper therefore concerns this supply-first configuration rather than any of its individual components. This positioning remains deliberately bounded by the literature examined and should not be interpreted as proof that no equivalent prior work exists.</p><p>The completed heterogeneous execution strengthens the research setting but does not resolve the central scientific question. The study has established that the artifact can construct large, provenance-linked MRDM graph populations from heterogeneous metadata sources. The next question is whether those structures are semantically defensible and useful. The scientific contribution will therefore depend less on the number of generated nodes, relationships, or candidate models than on what independent evaluation reveals about their validity, their dependence on metadata quality, their failure conditions, and the design principles that can be derived from those findings.</p><h3>18. Expected Contribution</h3><p>The expected contribution of the study operates at conceptual, design, and empirical levels. At the conceptual level, the study proposes the Machine-Readable Decision Model as a structured candidate decision context connecting potential use, decision supported, expected value, business objectives, information needs, reusable data products, available datasets, provenance, and explicit information gaps. The contribution does not lie in any of these concepts independently. Its purpose is to define how they can participate in a common machine-processable structure for supply-first discovery.</p><p>At the design level, the study contributes an architecture that begins with heterogeneous dataset metadata, separates source-specific ingestion from a normalized metadata representation, combines deterministic processing with LLM-supported semantic inference, constructs candidate MRDMs, and integrates those models through shared graph entities. The architecture also preserves the distinction between available data and inferred information demand, allowing unsupported information needs to remain represented as explicit gaps. This provides a mechanism for moving from a catalog-centered representation of what data exists toward a decision-centered representation of what the existing supply might support and what additional information would be required.</p><p>The graph-supported design provides a second design contribution. Rather than treating each generated use as an independent textual artifact, the graph allows information needs, data products, datasets, and objectives to be reused across candidate decision contexts. This creates the possibility of identifying recurring information demand, highly reusable data products, datasets supporting multiple potential uses, and systematic information gaps. Prior dataset and decision knowledge graphs establish the value of semantic graph representation in adjacent settings (Färber &amp; Lamprecht, 2021; Hao et al., 2021; Wang et al., 2025). The present study investigates whether similar value emerges when the graph is organized around supply-derived candidate decision contexts.</p><p>At the empirical level, the study seeks evidence concerning MRDM validity, relationship validity, metadata sensitivity, graph reuse, provenance, and failure conditions. The objective is not merely to report an accuracy measure for a generation pipeline. Evaluation is intended to explain where the artifact works, where it fails, and which properties of source metadata and graph construction influence those outcomes. This is consistent with the Design Science Research objective of deriving transferable knowledge from the construction and evaluation of artifacts rather than treating the artifact itself as the final scientific contribution (Hevner et al., 2004; Peffers et al., 2007).</p><p>The expected design knowledge includes principles concerning the separation of supply from inferred demand, explicit representation of information needs, preservation of unsupported requirements as gaps, separation of source-specific ingestion from semantic discovery, relationship-level provenance, and the treatment of AI-generated semantic structures as candidates requiring independent evaluation. These principles remain provisional until supported, modified, or rejected through empirical evaluation.</p><p>The intended contribution is therefore not another data catalog, generic knowledge graph, data-product framework, or AI use-case generator. It is design knowledge about whether existing metadata can support systematic discovery of potential data use while preserving enough semantic structure and provenance for the resulting hypotheses to be inspected, compared, selected, reused, and independently evaluated.</p><h3>19. Conclusion</h3><p>Organizations do not always begin data-value creation with a clearly articulated business objective, information requirement, or use case. In many settings, the inverse problem exists: substantial data assets are already available, but their potential uses remain unclear. Existing research provides strong foundations for dataset discovery, information-needs mapping, data-product design, decision knowledge graphs, and dataset knowledge graphs. These streams commonly begin from an expressed query, requirement, decision problem, or design context and subsequently identify relevant information or data (van den Homberg et al., 2018a, 2018b; Castro Fernandez et al., 2018; Hao et al., 2021; Färber &amp; Lamprecht, 2021; Viswanathan et al., 2023; Wang et al., 2025). This study investigates the complementary direction in which existing data supply provides the starting evidence for discovering candidate demand.</p><p>The proposed approach treats dataset metadata as evidence from which candidate decision contexts can be generated. These contexts are represented as graph-supported Machine-Readable Decision Models rather than as textual use-case suggestions. An MRDM connects a potential use with the decision supported, expected value, business objectives, information needs, candidate data products, available datasets, provenance, and explicit information gaps. The Information Need layer is particularly important because it allows the model to represent information required by a candidate decision even when the current data supply does not satisfy that requirement.</p><p>The completed implementation now establishes heterogeneous technical feasibility at a substantially larger scale than the initial artifact instantiation. The complete downstream pipeline processed 134,480 retained dataset metadata records across 65 heterogeneous portal snapshots. It generated 538 candidate objectives, 1,405 candidate use cases, 1,091 KPIs, 5,333 Information Needs, 661 candidate Data Products, 130,514 graph nodes, 174,394 typed relationships, and 20,422 explicit gaps. All 65 portal runs completed on their first attempt, all 174,394 relationships contain artifact-level provenance references, and checksum manifests for all 65 runs passed independent verification. The execution required 950 model calls, consumed 4,812,798 tokens, and took approximately three hours and 27 minutes.</p><p>These results must be interpreted within their empirical boundary. The source catalogs reported 341,634 datasets, but ten portal inputs were capped. The study therefore does not claim that all reported source datasets were ingested or analyzed. The completed execution concerns the 134,480 retained and processed metadata records. Similarly, successful execution does not establish semantic validity. The generated objectives, use cases, Information Needs, Data Products, and graph relationships remain candidate semantic structures requiring independent assessment.</p><p>The completed run also demonstrates that absence and incomplete supply can remain observable outcomes. The graph population contains 20,422 explicit information gaps, while one portal completed processing without producing a candidate Data Product satisfying the implemented acceptance criteria. These outcomes are retained rather than repaired or suppressed. Their interpretation requires semantic evaluation because a gap can arise from absent data, capped input, insufficient metadata, unsuccessful matching, or an unsupported generated Information Need. Nevertheless, preserving these outcomes is central to the artifact’s design because it prevents existing supply from defining information demand by construction.</p><p>The next empirical stage therefore moves from technical feasibility to semantic validity. Independent reviewers will assess complete MRDMs and sampled graph relationships against their source metadata and provenance. Metadata-sensitivity analysis will examine whether model validity varies with the richness and quality of the available descriptions, building on prior evidence that open-data metadata quality varies substantially across portals (Neumaier et al., 2016; Kubler et al., 2018). Structural analysis will examine whether shared Information Needs, Data Products, Datasets, and Business Objectives provide meaningful reuse and traceability across candidate decision contexts.</p><p>The central proposition of the study is consequently both narrower and more testable than a claim that AI can generate use cases from data. The question is whether metadata describing existing data supply contains sufficient evidence to construct structured candidate decision contexts that remain traceable to their source, expose the information required beyond current supply, and can be independently evaluated and selected by humans. The completed 65-portal execution demonstrates that such structures can be generated at heterogeneous multi-portal scale. The remaining scientific task is to establish whether they are semantically defensible, under which metadata conditions they are defensible, where the approach fails, and what transferable design principles follow from those findings.</p><h3>References</h3><p>Castro Fernandez, R., Abedjan, Z., Koko, F., Yuan, G., Madden, S., &amp; Stonebraker, M. (2018). AURUM: A data discovery system. In Proceedings of the 34th IEEE International Conference on Data Engineering (ICDE). IEEE.</p><p>Färber, M., &amp; Lamprecht, D. (2021). The data set knowledge graph: Creating a linked open data source for data sets. Quantitative Science Studies, 2(4), 1324–1355. <a href="https://doi.org/10.1162/qss_a_00161">https://doi.org/10.1162/qss_a_00161</a></p><p>Garijo, D., Poveda-Villalón, M., Amador-Domínguez, E., Wang, Z., García-Castro, R., &amp; Corcho, O. (2025). LLMs for ontology engineering: A landscape of tasks and benchmarking challenges. Proceedings of the ISWC Special Session on Harmonising Generative AI and Semantic Web Technologies, CEUR Workshop Proceedings.</p><p>Hao, J., Zhao, L., Milisavljevic-Syed, J., &amp; Ming, Z. (2021). Integrating and navigating engineering design decision-related knowledge using decision knowledge graph. Advanced Engineering Informatics, 50, 101366. <a href="https://doi.org/10.1016/j.aei.2021.101366">https://doi.org/10.1016/j.aei.2021.101366</a></p><p>Hasan, M. R., Finkel, B., &amp; Legner, C. (2026). The Data Product Canvas: Designing data products for sustained value from enterprise data. Information Systems Journal, 36(1), 144–158. <a href="https://doi.org/10.1111/isj.12603">https://doi.org/10.1111/isj.12603</a></p><p>Hevner, A. R., March, S. T., Park, J., &amp; Ram, S. (2004). Design science in information systems research. MIS Quarterly, 28(1), 75–105. <a href="https://doi.org/10.2307/25148625?utm_source=chatgpt.com">https://doi.org/10.2307/25148625</a></p><p>Horkoff, J., Aydemir, F. B., Cardoso, E., Li, T., Maté, A., Paja, E., Salnitri, M., Piras, L., Mylopoulos, J., &amp; Giorgini, P. (2019). Goal-oriented requirements engineering: An extended systematic mapping study. Requirements Engineering, 24(2), 133–160. <a href="https://doi.org/10.1007/s00766-017-0280-z">https://doi.org/10.1007/s00766-017-0280-z</a></p><p>Kubler, S., Robert, J., Neumaier, S., Umbrich, J., &amp; Le Traon, Y. (2018). Comparison of metadata quality in open data portals using the Analytic Hierarchy Process. Government Information Quarterly, 35(1), 13–29.</p><p>Nargesian, F., Zhu, E., Miller, R. J., Pu, K. Q., &amp; Arocena, P. C. (2018). Table union search on open data. Proceedings of the VLDB Endowment, 11(7), 813–825.</p><p>Neumaier, S., Umbrich, J., &amp; Polleres, A. (2016). Automated quality assessment of metadata across open data portals. Journal of Data and Information Quality, 8(1), Article 2.</p><p>Peffers, K., Tuunanen, T., Rothenberger, M. A., &amp; Chatterjee, S. (2007). A design science research methodology for information systems research. Journal of Management Information Systems, 24(3), 45–77. <a href="https://doi.org/10.2753/MIS0742-1222240302">https://doi.org/10.2753/MIS0742-1222240302</a></p><p>van den Homberg, M. J. C., Monné, R., &amp; Spruit, M. R. (2018a). Bridging the information gap: Mapping data sets on information needs in the preparedness and response phase. In S. Hostettler, S. Najih Besson, &amp; J.-C. Bolay (Eds.), Technologies for Development: From Innovation to Social Impact (pp. 213–225). Springer. <a href="https://doi.org/10.1007/978-3-319-91068-0_18">https://doi.org/10.1007/978-3-319-91068-0_18</a></p><p>van den Homberg, M. J. C., Monné, R., &amp; Spruit, M. R. (2018b). Bridging the information gap of disaster responders by optimizing data selection using cost and quality. Computers &amp; Geosciences, 120, 60–72. <a href="https://doi.org/10.1016/j.cageo.2018.06.002">https://doi.org/10.1016/j.cageo.2018.06.002</a></p><p>Viswanathan, V., Gao, L., Wu, T., Liu, P., &amp; Neubig, G. (2023). DataFinder: Scientific dataset recommendation from natural language descriptions. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics, Volume 1: Long Papers (pp. 10288–10303). Association for Computational Linguistics. <a href="https://doi.org/10.18653/v1/2023.acl-long.573">https://doi.org/10.18653/v1/2023.acl-long.573</a></p><p>Wang, R., Sun, Y., Peng, T., Hua, Y., Wang, G., &amp; Yan, Y. (2025). A knowledge graph-aided decision guidance method for product conceptual design. Journal of Engineering Design, 36(2), 216–255. <a href="https://doi.org/10.1080/09544828.2024.2368405">https://doi.org/10.1080/09544828.2024.2368405</a></p><p>Wilkinson, M. D., Dumontier, M., Aalbersberg, I. J., Appleton, G., Axton, M., Baak, A., Blomberg, N., Boiten, J.-W., da Silva Santos, L. B., Bourne, P. E., Bouwman, J., Brookes, A. J., Clark, T., Crosas, M., Dillo, I., Dumon, O., Edmunds, S., Evelo, C. T., Finkers, R., et al. (2016). The FAIR Guiding Principles for scientific data management and stewardship. Scientific Data, 3, 160018. <a href="https://doi.org/10.1038/sdata.2016.18?utm_source=chatgpt.com">https://doi.org/10.1038/sdata.2016.18</a></p><h3>Standards and Technical Sources</h3><p>CKAN Association. (2026). CKAN API guide. CKAN Documentation.</p><p>Open Data Products. (2025–2026). Open Data Products standards family. Linux Foundation AI &amp; Data Foundation.</p><p>World Wide Web Consortium (W3C). (2024). Data Catalog Vocabulary (DCAT), Version 3. W3C Recommendation.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=537a79c2cb51" width="1" height="1" alt=""><hr><p><a href="https://medium.com/demand-to-data-knowledge-graphs/from-dataset-metadata-to-decision-models-537a79c2cb51">From Dataset Metadata to Decision Models</a> was originally published in <a href="https://medium.com/demand-to-data-knowledge-graphs">The Information Continuity Lab</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Losing Context Costs Money. Keeping It Does Too.]]></title>
            <link>https://medium.com/demand-to-data-knowledge-graphs/losing-context-costs-money-keeping-it-does-too-b7188ce4ea12?source=rss----bcba237793cc---4</link>
            <guid isPermaLink="false">https://medium.com/p/b7188ce4ea12</guid>
            <category><![CDATA[continuity-economics]]></category>
            <dc:creator><![CDATA[Dr. Jarkko Moilanen]]></dc:creator>
            <pubDate>Fri, 02 Oct 2026 15:19:02 GMT</pubDate>
            <atom:updated>2026-10-02T15:19:02.349Z</atom:updated>
            <content:encoded><![CDATA[<p>There is an economic assumption hiding underneath much of my research. Losing context creates work. Someone has to find the data again. Someone has to reconstruct the requirements. Someone has to explain terminology between teams. Someone discovers halfway through implementation that an important information source is missing. Someone builds a data product that another team already built. Someone fixes an assumption that should have been resolved before coding started. None of these activities looks dramatic individually. Together, they consume time and money.</p><p>This is what I describe as the economic side of information discontinuity. But there is an important counterargument. Maintaining information continuity also costs money.</p><p>Knowledge graphs do not maintain themselves. Information needs have to be represented. Relationships need provenance. Data products need governance. Context needs to be maintained as things change.</p><p>So the fifth phase of my PhD research is not based on the assumption that more structure is automatically better.</p><p>The question is economic: When does the value of preserving context exceed the cost of maintaining it?</p><p>The earlier research phases provide the mechanisms required to study that question. The Demand-to-Data Knowledge Graph provides traceability. Cross-Portal Use Case Patterns test reuse and generalization. Information-demand measurement exposes gaps and reusable assets. The Context-to-Code experiments show what happens during implementation.</p><p>The fifth study brings those observations together. The research proposal identifies measures such as time from business objective to relevant data, number of datasets inspected, requirements effort, data-product reuse, duplicated work, information gaps detected before implementation, human corrections, repair cycles, time to working prototype and computational cost.</p><p>Those measures allow a different conversation about data and AI. Instead of: “Did we build the platform?” or: “Did we deploy the AI agent?” we can ask: “What did it cost to move from an identified business need to a working solution?” And then: “Where did that cost occur? Search? Interpretation? Missing information? Requirements reconstruction? Implementation repair? Duplicated data products? Model usage? Human review?”</p><p>If information continuity reduces some of those costs while adding others, we can begin to identify where the balance becomes favorable. That is the research question. Not whether information continuity sounds sensible. Not whether knowledge graphs are useful. Not whether AI agents are impressive. Whether preserving context changes the economics of moving from business intent to implemented value.</p><p>This research stream will follow that evidence as it develops.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=b7188ce4ea12" width="1" height="1" alt=""><hr><p><a href="https://medium.com/demand-to-data-knowledge-graphs/losing-context-costs-money-keeping-it-does-too-b7188ce4ea12">Losing Context Costs Money. Keeping It Does Too.</a> was originally published in <a href="https://medium.com/demand-to-data-knowledge-graphs">The Information Continuity Lab</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Does Better Context Produce Better AI-Built Software?]]></title>
            <link>https://medium.com/demand-to-data-knowledge-graphs/does-better-context-produce-better-ai-built-software-8960c86b1d18?source=rss----bcba237793cc---4</link>
            <guid isPermaLink="false">https://medium.com/p/8960c86b1d18</guid>
            <category><![CDATA[context-to-code]]></category>
            <dc:creator><![CDATA[Dr. Jarkko Moilanen]]></dc:creator>
            <pubDate>Fri, 02 Oct 2026 15:12:18 GMT</pubDate>
            <atom:updated>2026-10-02T15:12:18.880Z</atom:updated>
            <content:encoded><![CDATA[<p>Finding the right data is not the end of the problem. Someone still has to build something with it. And every handover creates another opportunity to lose context.</p><p>A business objective becomes a use case. The use case becomes requirements. Requirements become technical decisions. Data gets selected. Someone writes specifications. Increasingly, those specifications become prompts given to AI coding assistants. By the time implementation begins, how much of the original intent is still present?</p><p>This is the fourth phase of my PhD research. I already have a multi-agent system that takes structured use-case and data context and turns it into an implementation-ready package for coding assistants.</p><p>The system is useful operationally, but that is not the research question. The research question is whether preserving more structured context upstream produces measurably different results downstream.</p><p>That gives me a relatively clean experiment. One coding assistant receives a conventional textual use case. Another receives structured use-case and data context. A third receives an implementation package produced by an AI-agent team analysing the same structured context. Then they build.</p><p>Now we have outputs that can be compared.</p><ul><li>Did the implementation satisfy the original requirements?</li><li>Did it select the correct datasets?</li><li>How many unsupported assumptions appeared?</li><li>How many defects required correction?</li><li>How much human intervention was required?</li><li>How many repair cycles occurred?</li><li>How long did implementation take?</li><li>What did the model usage cost?</li></ul><p>And perhaps most importantly, can the final implementation still be traced back to the original business objective and information needs? These are the types of measures proposed for the fourth study. The purpose is not to prove that multi-agent systems are better.</p><p>I am deliberately separating the research question from the architecture. The agents are a mechanism for preserving, analysing and transforming context. A different implementation might eventually perform the same role.</p><p>The underlying question is more durable: Does preserved context improve AI-assisted implementation? That question is becoming more important as coding assistants become normal development tools.</p><p>Giving an AI agent access to data is easy. Giving it enough context to understand why the data matters, what information the user needs, what evidence supports the solution and what outcome the implementation is supposed to produce is a different problem.</p><p>This research stream will document those experiments from context through agents and coding assistants to working prototypes. Some of the most useful findings will likely come from where the context still fails.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=8960c86b1d18" width="1" height="1" alt=""><hr><p><a href="https://medium.com/demand-to-data-knowledge-graphs/does-better-context-produce-better-ai-built-software-8960c86b1d18">Does Better Context Produce Better AI-Built Software?</a> was originally published in <a href="https://medium.com/demand-to-data-knowledge-graphs">The Information Continuity Lab</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Information Demand as a Measure of Data Ecosystems]]></title>
            <link>https://medium.com/demand-to-data-knowledge-graphs/information-demand-as-a-measure-of-data-ecosystems-b5a5abe3c080?source=rss----bcba237793cc---4</link>
            <guid isPermaLink="false">https://medium.com/p/b5a5abe3c080</guid>
            <category><![CDATA[information-demand]]></category>
            <dc:creator><![CDATA[Dr. Jarkko Moilanen]]></dc:creator>
            <pubDate>Fri, 02 Oct 2026 14:55:23 GMT</pubDate>
            <atom:updated>2026-10-02T14:55:23.547Z</atom:updated>
            <content:encoded><![CDATA[<h4>What If We Measured Data by What People Need?</h4><p>We have spent years measuring data supply. How many datasets are published? Does the metadata meet a standard? Is the data machine-readable? Does the portal provide APIs? Are licenses clear?</p><p>These are useful questions. They tell us something about the condition of a data ecosystem. They do not answer another question: <em>Does the available data satisfy meaningful information demand?</em></p><p>This distinction is the starting point for the third phase of my PhD research. The first two phases create something that does not normally exist at scale: structured information about demand for data.</p><p>Use cases contain information needs. Those information needs can be connected to data products and supporting datasets. When the required data is unavailable, the gap remains visible. Once enough of these structures exist, they become an empirical object in their own right. We can stop looking only at what organizations publish and start comparing supply with observed demand. A simple first measure might be:</p><p>Demand Coverage = Supported Information Needs / Observed Information Needs</p><p>The actual research will need more sophisticated measures. Not every information need has equal importance. A technically available dataset might be too old, too incomplete or too poorly structured to satisfy the need. Some needs require several independent sources. But even the basic change in perspective is significant.</p><p>Imagine two portals.</p><ul><li>Portal A publishes 10,000 datasets.</li><li>Portal B publishes 2,000.</li></ul><p>Traditional supply-oriented measures make Portal A look much larger.</p><p>Now suppose the use cases we observe contain 500 recurring information needs. Portal A provides adequate evidence for 180. Portal B provides adequate evidence for 320. Which ecosystem better supports value creation?</p><p>Dataset count no longer answers the question. The research proposal therefore treats accumulated information needs as a way to investigate recurring demand, systematic gaps, high-reuse data categories and differences between data ecosystems. This could also tell us something about what should be published next.</p><p>If the same information need repeatedly appears across independent use cases and jurisdictions but adequate supporting data is consistently absent, that gap becomes more than an individual user’s failed search. It becomes evidence of unmet information demand.</p><p>Likewise, a dataset category that supports dozens of different use cases might deserve more attention than another category with little observed reuse.</p><p>This is where the economics side of my research begins to become visible. Data has supply. Data also has demand. We have built extensive machinery for measuring the first. I want to investigate what becomes visible when we start measuring the second.</p><p>This research stream will follow that work, including the measures, recurring information needs, gaps, reuse patterns and comparisons between data ecosystems.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=b5a5abe3c080" width="1" height="1" alt=""><hr><p><a href="https://medium.com/demand-to-data-knowledge-graphs/information-demand-as-a-measure-of-data-ecosystems-b5a5abe3c080">Information Demand as a Measure of Data Ecosystems</a> was originally published in <a href="https://medium.com/demand-to-data-knowledge-graphs">The Information Continuity Lab</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Stream 2–01: Cross-Portal Use Case Patterns]]></title>
            <link>https://medium.com/demand-to-data-knowledge-graphs/cross-portal-use-case-patterns-426b58c1bab1?source=rss----bcba237793cc---4</link>
            <guid isPermaLink="false">https://medium.com/p/426b58c1bab1</guid>
            <category><![CDATA[cross-portal-patterns]]></category>
            <dc:creator><![CDATA[Dr. Jarkko Moilanen]]></dc:creator>
            <pubDate>Fri, 02 Oct 2026 14:43:51 GMT</pubDate>
            <atom:updated>2026-10-04T12:57:59.887Z</atom:updated>
            <content:encoded><![CDATA[<h4>A Use Case Does Not Care Which Portal Owns the Data</h4><p>Data portals create an artificial boundary around data. Real problems rarely respect it. Consider an urban heat-risk use case. To understand where intervention is required, we might need temperature observations, land cover, building information, population vulnerability and health indicators. Those datasets are unlikely to belong to one publisher.</p><p>One organization might publish weather observations. Another maintains population statistics. A municipality holds building information. Health indicators come from somewhere else. Yet the information need remains coherent.</p><p>This observation led to something interesting in the implementation behind my PhD research: Cross-Portal Use Case Patterns. A Cross-Portal Use Case Pattern separates the information demand from the particular data environment used to satisfy it. The urban heat example might therefore have a relatively stable structure:</p><p>Urban Heat Risk</p><p>→ temperature information<br>→ vulnerable population information<br>→ building information<br>→ land-cover information<br>→ health information</p><p>In one city, three portals might satisfy those needs. In another, five portals might be required. Somewhere else, two of the information needs might have no adequate data at all.</p><p>The datasets change. The question is whether the demand pattern does. That is the second phase of my research.</p><p>The study will investigate whether patterns identified from one collection of data environments transfer to environments the model has not previously seen. The research proposal describes a straightforward experimental approach: derive patterns using one set of portals, hold other portals out, and then test how successfully the same information needs can be instantiated using those unseen data sources. This makes generalization measurable.</p><p>We can examine information-need coverage. We can assess whether the selected datasets make sense. Independent experts can review the relationships. We can identify which parts of a pattern transfer successfully and which are specific to the original environment.</p><p>Failures are particularly interesting. If a pattern repeatedly works across cities, countries or organizations, that suggests the underlying information demand has some degree of stability.</p><p>If it consistently fails when transferred, that tells us something else. Perhaps the use case was more context-specific than expected. Perhaps terminology differs. Perhaps the underlying information need changes. Or perhaps the new environment simply lacks the necessary data.</p><p>At sufficient scale, this could lead to reusable libraries of information-demand patterns that exist independently of individual catalogs. A new data environment could then be examined against known demand patterns rather than evaluated only by counting its datasets. That is a much larger idea than cross-portal search. The research has to establish whether it works first.</p><p>This research stream will follow those experiments, including patterns that transfer, patterns that fail, information gaps and what the results tell us about the relationship between stable demand and changing data supply.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=426b58c1bab1" width="1" height="1" alt=""><hr><p><a href="https://medium.com/demand-to-data-knowledge-graphs/cross-portal-use-case-patterns-426b58c1bab1">Stream 2–01: Cross-Portal Use Case Patterns</a> was originally published in <a href="https://medium.com/demand-to-data-knowledge-graphs">The Information Continuity Lab</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Demand-to-Data Knowledge Graphs]]></title>
            <link>https://medium.com/demand-to-data-knowledge-graphs/demand-to-data-knowledge-graphs-04f8a924e0f2?source=rss----bcba237793cc---4</link>
            <guid isPermaLink="false">https://medium.com/p/04f8a924e0f2</guid>
            <category><![CDATA[demand-to-data]]></category>
            <dc:creator><![CDATA[Dr. Jarkko Moilanen]]></dc:creator>
            <pubDate>Fri, 02 Oct 2026 14:13:44 GMT</pubDate>
            <atom:updated>2026-10-04T13:00:56.139Z</atom:updated>
            <content:encoded><![CDATA[<h4>Stream 1–01: From a Business Objective to the Data Required to Act</h4><p>Organizations have become good at cataloging data. We know what datasets exist, who owns them, what fields they contain and, when things go well, something about their quality and lineage. Start from the other direction and the situation becomes much less clear.</p><p>Take a business objective such as reducing congestion, improving preventative healthcare or increasing tourism expenditure. What information would you need to pursue that objective? Which use cases depend on that information? Which reusable data products should provide it? Which datasets contain the underlying evidence?</p><p>Most organizations do not have one machine-readable structure connecting those questions. This is where the first part of my PhD research starts. I am studying a structure I currently call the Demand-to-Data Knowledge Graph:</p><p>Business Objective → Use Case → Information Need → Data Product → Dataset</p><p>None of these concepts is new by itself. Goals and requirements have a long research history. Information needs have been studied extensively. Data products are now becoming an established field of research and practice. Knowledge graphs provide mature mechanisms for representing relationships.</p><p>The research question sits between them. <em>Can these concepts form a useful and valid traceability structure from business demand to actual data supply?</em></p><p>Information Need has a particularly important role in the model. A use case describes what we intend to accomplish. An information need describes what must be known, measured, observed or inferred for that use case to work. A data product then provides a reusable means of satisfying that need, supported by one or more datasets.</p><p>This also works backwards. If I start with a dataset, I should be able to trace which data products depend on it, which information needs those products satisfy, which use cases require that information and which business objectives those use cases support.</p><p>That changes the role of a data catalog. Instead of answering only: “What data do we have?” we should also be able to ask: “Why do we need it?”</p><p>And there is another useful outcome. Sometimes the answer is that the required data does not exist. In this model, that missing connection stays visible as an information gap instead of disappearing because no dataset was found.</p><p>I already have an implementation of this model. That gives me an unusual starting point for the research. I am not beginning by drawing a conceptual diagram and then trying to build it. The artifact exists. The next job is to test it scientifically.</p><p>That means examining the graph structure, evaluating individual relationships independently, finding where those relationships fail and determining whether the model provides useful traceability across heterogeneous data environments. This research stream will document that process.</p><p>Some posts will discuss the semantic model. Others will look at prior research, evaluation methods, information gaps and results from the empirical work. The objective is not to prove that the model I built is correct. The objective is to find out whether it is.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=04f8a924e0f2" width="1" height="1" alt=""><hr><p><a href="https://medium.com/demand-to-data-knowledge-graphs/demand-to-data-knowledge-graphs-04f8a924e0f2">Demand-to-Data Knowledge Graphs</a> was originally published in <a href="https://medium.com/demand-to-data-knowledge-graphs">The Information Continuity Lab</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
    </channel>
</rss>