<?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[Spindox Labs - Medium]]></title>
        <description><![CDATA[From our R&amp;D lab working across European and National research programmes. Articles on data infrastructure, AI, interoperability, and the systems that connect them. - Medium]]></description>
        <link>https://medium.com/spindox-labs?source=rss----ab498e9e959d---4</link>
        <image>
            <url>https://cdn-images-1.medium.com/proxy/1*TGH72Nnw24QL3iV9IOm4VA.png</url>
            <title>Spindox Labs - Medium</title>
            <link>https://medium.com/spindox-labs?source=rss----ab498e9e959d---4</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Fri, 20 Nov 2026 09:13:20 GMT</lastBuildDate>
        <atom:link href="https://medium.com/feed/spindox-labs" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[What coding agents should remember, and where]]></title>
            <link>https://medium.com/spindox-labs/what-coding-agents-should-remember-and-where-1b668729ebbe?source=rss----ab498e9e959d---4</link>
            <guid isPermaLink="false">https://medium.com/p/1b668729ebbe</guid>
            <category><![CDATA[harness-engineering]]></category>
            <category><![CDATA[software-testing]]></category>
            <category><![CDATA[software-engineering]]></category>
            <category><![CDATA[open-source]]></category>
            <category><![CDATA[agents]]></category>
            <dc:creator><![CDATA[Luca Capra]]></dc:creator>
            <pubDate>Mon, 09 Nov 2026 10:31:01 GMT</pubDate>
            <atom:updated>2026-11-09T10:31:01.539Z</atom:updated>
            <content:encoded><![CDATA[<h3>Two months of building a harness around them: what seems to hold, and what we had to take back.</h3><p><em>This article has been drafted with AI support and manually reviewed by its author and the editorial team.</em></p><figure><img alt="Two trees side by side: the repository, which publishes, and the store beside it, which remembers. A short file in the first points at the second. (AI-generated image)" src="https://cdn-images-1.medium.com/max/1024/1*f-DEHbzHGlSd8xU5a8ktGg.png" /></figure><p>In a fresh session, a coding agent knows about a project only what it reads at the start, and it acts on what it reads. Until August we kept that memory in our repositories, as instruction files and notes. Since then we have moved it out, into a separate store with a few rules that keep it honest. This is a report on that setup, not a framework launch. We use it daily in a handful of repositories. We have not measured whether it makes agents better. In <a href="https://medium.com/spindox-labs/the-expertise-is-the-scarce-part-972108c42801">The expertise is the scarce part</a> we described how we use AI for writing; this is the software side.</p><h3>One file that only says where to look</h3><p>Each repository we run this way has one AGENTS.md, following the <a href="https://agents.md/">open convention</a> that mainstream coding assistants and open-source agents read. Ours is the same file in all of them. It tells the agent where the working material is and to read the rules there first. If the material is missing, the agent must ask and stop. It may not write notes into the repository instead.</p><p>Tools that look for a file with their own name get a one-line pointer. Our <a href="https://github.com/spindoxlabs/ds">data space repository</a> has a CLAUDE.md that says only &quot;Read AGENTS.md.&quot; We started with one instruction file per tool. Within weeks they had drifted apart.</p><h3>Working memory lives beside the code</h3><p>Agent notes do not belong in a published repository. Plans, investigations and open defects are written fast, for whoever is doing the work, and nobody checks them before they go public. The data space repository committed its agent material for a while, arguing that review and durability were worth the exposure. On 16 August it <a href="https://github.com/spindoxlabs/ds/blob/main/docs/decisions/ADR-0002-agent-material-is-not-committed.md">reversed that decision</a>. They were not.</p><p>The material now sits in a store, a directory beside the checkouts, found by convention and never by a committed path. Inside, every repository has a folder named after its organisation and its own name. That fixed layout lets a note about one service link a note about another by a relative path. An agent changing one repository can follow the link to what is known about its neighbour.</p><p>The layout has to be exact. One store, named a few characters off the convention, told every agent there was no store at all.</p><h3>Four kinds of note, sorted by what ends them</h3><ul><li><strong>Plans</strong> hold intent and the decisions taken. They end when the change is complete, and never record progress.</li><li><strong>Work records</strong> hold progress: what was done, how it was checked, what is still owed. They exist before the first change and end with the plan.</li><li><strong>Knowledge</strong> holds what is true of the code but not visible in it. Each note is named after the trap it describes, and deleted when it stops being true, because people trust outdated notes.</li><li><strong>Playbooks</strong> hold repeated, fragile procedures: syncing data, scaffolding a component, starting services in dependency order. One is written the second time a procedure is performed.</li></ul><p>We split plans from work so that a new session can pick up where the last one stopped. The data space repository <a href="https://github.com/spindoxlabs/ds/blob/main/docs/decisions/ADR-0003-plans-hold-intent-work-holds-progress.md">explains why</a>: every plan it held opened with a table of counts that was already wrong. A new session reads the rules, then the work record’s handover, then the plan’s decisions. One rule came out of it. A number a command can produce is never written down by hand.</p><p>A playbook earns its place in its trap section. A container stack can report success while a startup step inside it failed. A catalogue import can succeed with fewer entries because a neighbouring checkout was missing.</p><h3>Every requirement has a code, and a test must name it</h3><p>A requirement gets an identifier, a test that verifies it names the identifier, and a check fails when either is missing. Our <a href="https://github.com/celine-eu/celine-grid">grid resilience backend</a> does this in a dozen lines of its <a href="https://github.com/celine-eu/celine-grid/blob/main/.github/workflows/test.yaml">CI workflow</a>. The build fails on any requirement no test names and on any test tag naming a requirement nobody wrote. On its main branch that covers 53 requirements, and the step prints the count itself.</p><p>The claim goes in the test because a test runs, and its tag disappears when the test is deleted. A comment saying “this implements requirement 7” cannot be run. The rules tell an agent asked for something with no requirement behind it to ask, not to write one.</p><p>The data space repository already traced European data space blueprints in its own rulebook, so it <a href="https://github.com/spindoxlabs/ds/blob/main/docs/decisions/ADR-0001-conformance-stays-with-the-rulebook.md">kept that</a> and added no second set of codes. Its <a href="https://github.com/spindoxlabs/ds/blob/main/docs/rulebook/status.md">generated status page</a> for 5 October 2026 lists 166 rules. All 134 that claim enforcement are named by at least one of 1,166 test nodes. The page also states its limit: it measures that a test names a rule, not that the test passes. And 531 of 783 binding requirements are not assessed yet.</p><h3>What we have not shown</h3><p>We have no measurement of whether agents make fewer mistakes this way, and no comparison against working without it.</p><p>Two parts work less well than the rules say. Agents move lessons from work records into knowledge only some of the time. And one check stayed green for weeks because the file it tested had been deleted.</p><p><strong>Links</strong></p><ul><li>The data space repository and its decision records: <a href="https://github.com/spindoxlabs/ds">https://github.com/spindoxlabs/ds</a></li><li>The generated conformance status page: <a href="https://github.com/spindoxlabs/ds/blob/main/docs/rulebook/status.md">https://github.com/spindoxlabs/ds/blob/main/docs/rulebook/status.md</a></li><li>The grid resilience backend and its requirement check: <a href="https://github.com/celine-eu/celine-grid">https://github.com/celine-eu/celine-grid</a></li><li>The AGENTS.md convention: <a href="https://agents.md/">https://agents.md/</a></li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=1b668729ebbe" width="1" height="1" alt=""><hr><p><a href="https://medium.com/spindox-labs/what-coding-agents-should-remember-and-where-1b668729ebbe">What coding agents should remember, and where</a> was originally published in <a href="https://medium.com/spindox-labs">Spindox Labs</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Rules that travel with the data]]></title>
            <link>https://medium.com/spindox-labs/rules-that-travel-with-the-data-03460d4e881c?source=rss----ab498e9e959d---4</link>
            <guid isPermaLink="false">https://medium.com/p/03460d4e881c</guid>
            <category><![CDATA[dataspace]]></category>
            <category><![CDATA[open-source]]></category>
            <category><![CDATA[energy]]></category>
            <category><![CDATA[data-governance]]></category>
            <dc:creator><![CDATA[Luca Capra]]></dc:creator>
            <pubDate>Tue, 20 Oct 2026 10:01:03 GMT</pubDate>
            <atom:updated>2026-10-20T10:01:03.479Z</atom:updated>
            <content:encoded><![CDATA[<p><em>How consent and usage limits can follow energy data wherever it goes, and why a clear record of origin is what makes a shared number worth trusting.</em></p><p><em>This article has been drafted with AI support and manually reviewed by its author and the editorial team.</em></p><figure><img alt="Permission that stays with the data is the difference between a rule and a promise. (AI-generated image)" src="https://cdn-images-1.medium.com/max/1024/1*_yR26QlNqKTpe95ZWMEXDA.png" /></figure><p>In <a href="https://medium.com/spindox-labs/who-controls-access-to-the-data-your-energy-system-runs-on-44d4b4dbcb3c">Who controls access to the data your energy system runs on</a> we argued that energy data should be shared under its owner’s control. This piece is about keeping that control after the data leaves. A gate that checks credentials at the door stops working once a copy is on the other side, and energy data can reveal when a household is awake, away or charging a car.</p><p>Our alternative attaches the rules to the data. Every dataset carries its own rulebook: who owns it, what it may be used for, whether the person it concerns still agrees. Each time the data is served, that rulebook is consulted again.</p><h3>Consent that is still alive when the data is delivered</h3><p>Consent becomes a live condition, checked at each handover. If a person has withdrawn it, their records stop being returned. Even an exchange already in progress honours a withdrawal. What is served through the space is covered. A file someone has already downloaded stays a matter of law and contract.</p><p>People are asked about things they can judge: what the data would be used for, which organisation would receive it and in which of its roles, which category of parties may act on its behalf, how detailed the readings are and how long they may be kept. Which internal datasets sit behind that is the operator’s business. Our team has published an open reference implementation of this pattern for energy communities, with a consent portal where a member grants, refuses or withdraws permission one such bundle at a time.</p><figure><img alt="Consent withdrawn after the handover still reaches a transfer that has already started. (AI-generated image)" src="https://cdn-images-1.medium.com/max/1024/1*LkFgbX0yEKAtIScibssThw.png" /></figure><p>Each permission is bound to the facts the person read. Swapping the table underneath, or admitting a new service provider into a category already disclosed, changes none of those facts, so nobody is asked again. Changing the detail of the readings, the span of history, the retention or the purpose does change them. Such a change cannot slip in under the old wording. It goes out as a new version, and what each person already agreed to keeps meaning what it meant. Nothing blocks anyone; the new version waits in their own data view until they look. Asking constantly would train people to click through the questions that matter.</p><h3>Why the reason for asking decides the answer</h3><p>Permissions, allowed purposes, dataset sensitivity and consent conditions are written in configuration files next to the data and turned automatically into enforceable rules. Before deployment, a check refuses the whole set if a declared purpose is not in the shared vocabulary, or if a bundle names a receiving organisation the registry does not know.</p><p>Those rules then answer two questions. First, when a sharing agreement is negotiated: should this party receive this data at all? If a request does not fit the declared purpose or access level, it is refused there and no transfer begins. Second, at every query: does this person still agree, right now, for this use? Every query for personal data must state its purpose, and the system returns only the rows of people whose permission covers it. The same requester, under the same agreement and the same live transfer, gets different rows for a different purpose. Queries with no stated purpose get nothing. Permissions are never widened: a narrower use inside the one agreed is covered, a broader or neighbouring one is not.</p><p>When some people have not yet answered, the request is held open and resumes or lapses on their answer. When there is nobody to ask, it is refused.</p><h3>Why the record of origin is what earns trust</h3><p>Settlement figures and forecast inputs are only as trustworthy as their history. So each dataset also carries a record of where it came from and what was done to it. Whatever does the work writes that record as it goes, pipeline by pipeline, each declaring the licence and ownership of what it brings in.</p><p>Significant moments are logged the same way: a dataset published to the catalogue, an agreement negotiated, a transfer made, a query answered, a permission granted or withdrawn. Offline handovers are entered by the operator when they happen. Together they form a trail that can be walked in either direction. Backwards, a questioned figure’s history can be reconstructed step by step. Forwards, a gap found late in a source leads to everything that inherited it.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*xCQrtHFo-_rFmi4x6SYP5A.png" /></figure><p>Only codes, pseudonymous identifiers and fingerprints go into the log, never a name, an address or a reading. Pseudonyms are a safeguard, not anonymity, but they keep the log from becoming a second copy of the data. Entries are chained, so one changed or removed after the fact shows up when the chain is checked, and after a set period even the pseudonyms are replaced. When a batch changes hands, the log records a fingerprint of the permissions in force, which can later be checked against the consent records to establish which state authorised the handover. Because rulebook and log are kept together, one trail answers both “where did this come from” and “what am I allowed to do with it”.</p><h3>Standing on public rules, not private ones</h3><p>European law supplies the ground. The <a href="https://digital-strategy.ec.europa.eu/en/policies/data-governance-act">Data Governance Act</a>, applicable since 2023, sets out how neutral intermediaries can help data move without appropriating it. The <a href="https://digital-strategy.ec.europa.eu/en/policies/data-act">Data Act</a>, applicable since September 2025, gives the user of a connected product, and a networked meter can qualify, a clearer right to reach its data and have it shared with a third party of their choosing. A <a href="https://energy.ec.europa.eu/publications/common-european-energy-data-space_en">Common European Energy Data Space</a> is being assembled on these rules, still pilot by pilot.</p><p>One of the sharper distinctions comes from the <a href="https://eur-lex.europa.eu/eli/reg/2016/679/oj">General Data Protection Regulation</a>. A party acting on the instructions of the organisation a person already agreed with is disclosed to them. A party that decides its own purposes is different: where the sharing rests on consent, it has to be asked afresh. Where processing rests on a contract, the arrangement is disclosed and no switch is offered, because there is no real choice to make. Grounding the rules in public law is what lets a receiving party recognise them as legitimate.</p><h3>What travels with the data</h3><p>Put together, no rule stays behind at the door. Consent is checked at every handover, and a withdrawal reaches even a transfer already running. Every query states its purpose and receives only the rows of people whose permission covers it. A change to anything a person read goes out as a new version, while what they already agreed to keeps its meaning. Behind it all, a log that holds no names lets any figure be walked back to its origin, or forward to everything that inherited it, with the permissions in force at each step.</p><p>That answers the three questions a party receiving energy data has: am I allowed to use this, does the person still agree, and can I defend the number if challenged. The answers rest on public European law, so the receiving party can recognise them as legitimate. The whole pattern is an openly licensed reference implementation, exercised end to end, and organisations can enroll with keys they hold themselves.</p><p>The data leaves and the conditions with it.</p><p><em>Open work referenced: the </em><a href="https://github.com/spindoxlabs/ds"><em>data space implementation</em></a><em> and its </em><a href="https://spindoxlabs.github.io/ds/"><em>documentation</em></a><em>, covering the consent, governance and provenance components discussed here, and the </em><a href="https://github.com/celine-eu"><em>open toolbox</em></a><em> it builds on.</em></p><p><em>Policy sources: the EU </em><a href="https://digital-strategy.ec.europa.eu/en/policies/data-governance-act"><em>Data Governance Act</em></a><em> and </em><a href="https://digital-strategy.ec.europa.eu/en/policies/data-act"><em>Data Act</em></a><em>; the </em><a href="https://energy.ec.europa.eu/publications/common-european-energy-data-space_en"><em>Common European Energy Data Space</em></a><em>; the </em><a href="https://eur-lex.europa.eu/eli/reg/2016/679/oj"><em>General Data Protection Regulation</em></a><em>.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=03460d4e881c" width="1" height="1" alt=""><hr><p><a href="https://medium.com/spindox-labs/rules-that-travel-with-the-data-03460d4e881c">Rules that travel with the data</a> was originally published in <a href="https://medium.com/spindox-labs">Spindox Labs</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[How data access rules are enforced where the data lives]]></title>
            <link>https://medium.com/spindox-labs/how-data-access-rules-are-enforced-where-the-data-lives-2037ce2f3b98?source=rss----ab498e9e959d---4</link>
            <guid isPermaLink="false">https://medium.com/p/2037ce2f3b98</guid>
            <category><![CDATA[open-source]]></category>
            <category><![CDATA[energy]]></category>
            <category><![CDATA[cybersecurity]]></category>
            <category><![CDATA[dataspace]]></category>
            <dc:creator><![CDATA[Luca Capra]]></dc:creator>
            <pubDate>Tue, 13 Oct 2026 10:01:03 GMT</pubDate>
            <atom:updated>2026-10-13T10:01:03.348Z</atom:updated>
            <content:encoded><![CDATA[<p><em>In the data space we built, whoever holds the data enforces the rules at the point where it is touched, and refuses when it cannot decide.</em></p><p><em>This article has been drafted with AI support and manually reviewed by its author and the editorial team.</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*iekv6rNdN1PTggMIdACpGg.png" /></figure><p>A household lets its energy community use its meter readings for billing, but not a third-party analytics firm. Something has to check that rule each time the data is requested, and it matters who runs that check. In <a href="https://medium.com/spindox-labs/who-controls-access-to-the-data-your-energy-system-runs-on-44d4b4dbcb3c">Who controls access to the data your energy system runs on</a> we argued that the owner of the data should decide. Here the question is who enforces that decision, and our answer is whoever holds the data, wherever it is touched.</p><h3>Why one gatekeeper does not fit a data space</h3><p>Inside a single organisation, a central permissions server works well enough, because everyone trusts the organisation running it. A data space joins households, community managers, grid operators and service providers, and none of them is the obvious owner of that server. A central gate can also become a bottleneck and a single point of failure. Trust is the deeper problem. If the grid operator’s rules are enforced on the community platform’s server, or a household’s on a service provider’s, sovereignty becomes a promise rather than a fact.</p><p>Security practice reached the same conclusion from another direction. <a href="https://csrc.nist.gov/pubs/sp/800/207/final">NIST Special Publication 800–207</a>, the US National Institute of Standards and Technology’s definition of zero-trust architecture, grants no implicit trust based on where a request comes from: every request is evaluated on its own. For data spaces, the <a href="https://blueprint.dssc.eu/">Data Spaces Support Centre Blueprint v3.0</a> requires policies to be enforced at publication, at discovery, during contract negotiation and during the sharing itself, and states that every participant must have the capability to enforce them. Catalogues and other services can still be shared. Our reading is that enforcement itself cannot sit with a single gatekeeper.</p><h3>Each service carries its own rules</h3><p>In the <a href="https://celine-eu.github.io/">CELINE toolbox</a>, access rules are written in Rego, the policy language of the <a href="https://www.openpolicyagent.org/">Open Policy Agent</a>, a graduated project of the Cloud Native Computing Foundation. Each service ships its own rules, versioned and reviewed with its code, so anyone auditing a service reads the rules it actually runs.</p><p>Those rules are evaluated inside the service. A shared library, imported by seven of the toolbox’s repositories, runs an open-source Rego interpreter written in Rust, which the project packages for Python. No policy server sits in the request path. Each decision costs no network hop, and a service can never be running but unable to authorise anyone because a policy server elsewhere is down.</p><p>What the library does when something is missing was decided case by case. A rule that does not say yes says no. A service whose policies did not load refuses to start, because an engine that loaded nothing and denied everything would look, from outside, exactly like a strict policy. For the same reason, the tests run the real rules and check that they refuse, since a broken setup can also produce a yes.</p><h3>Five places the data space checks</h3><p>The <a href="https://github.com/spindoxlabs/ds">data space platform</a> repeats the check at each stage the blueprint names, and adds one per query.</p><ul><li><strong>Publication.</strong> A dataset whose rules fail validation is not published.</li><li><strong>Discovery.</strong> A participant sees only the offers it is admitted to. Others stay hidden instead of being refused later.</li><li><strong>Negotiation.</strong> Membership, proven by a credential the data space signs, and the intended recipient are checked first. Purpose, consent and contract terms follow.</li><li><strong>During the transfer.</strong> Consent is checked again while data flows, and a withdrawal terminates a transfer already running.</li><li><strong>Every query.</strong> The data service, operated by the participant that holds the data, asks that participant’s own connector for a decision and for the filter that says which rows may leave. If it cannot apply that filter, it serves nothing.</li></ul><p>A check that cannot be decided is a denial. That rule is the platform’s own, stricter than the blueprint asks, and its thresholds are written down. Before a transfer starts, one unanswerable check refuses it, and the cost is a retry. Once the transfer is running, three unanswerable checks in a row are tolerated before it is terminated, because a single network blip should not destroy a live agreement, and the data service keeps refusing unanswerable queries in the meantime. A definite no stops the transfer at once.</p><p>Query decisions are cached for 30 seconds. The documentation calls that window a security parameter, since it is how long a revoked consent can still yield rows.</p><h3>Sovereignty you can check</h3><p>Put together, whoever holds the data enforces the rules, at the point where it is touched. Rules are code, reviewed and tested like any other change and readable by a regulator or a community member. The engine runs inside each service, answers with no round-trip to a remote authority, and denies when in doubt. On top of that, the data space checks at publication, discovery, negotiation, during the transfer and on every query, down to the individual row.</p><p>All of it is open reference code that can be inspected and run, and little of it is specific to energy. Energy communities providing household data is a sensitive enough scenario, and the participants varied enough, to need it first.</p><p>Enforcement on someone else’s server turns sovereignty into a promise. When it’s checked where the data lives, it becomes a fact.</p><p><em>Open work referenced: the </em><a href="https://github.com/celine-eu/celine-sdk"><em>shared library</em></a><em> that evaluates rules inside each service, the </em><a href="https://github.com/celine-eu/celine-regorus"><em>packaged policy interpreter</em></a><em>, the </em><a href="https://github.com/celine-eu/celine-policies"><em>identity and policy tooling</em></a><em>, and the </em><a href="https://github.com/spindoxlabs/ds"><em>data space platform</em></a><em> with its </em><a href="https://spindoxlabs.github.io/ds/rulebook/policies/"><em>rulebook on policy enforcement</em></a><em>.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=2037ce2f3b98" width="1" height="1" alt=""><hr><p><a href="https://medium.com/spindox-labs/how-data-access-rules-are-enforced-where-the-data-lives-2037ce2f3b98">How data access rules are enforced where the data lives</a> was originally published in <a href="https://medium.com/spindox-labs">Spindox Labs</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The expertise is the scarce part]]></title>
            <link>https://medium.com/spindox-labs/the-expertise-is-the-scarce-part-972108c42801?source=rss----ab498e9e959d---4</link>
            <guid isPermaLink="false">https://medium.com/p/972108c42801</guid>
            <category><![CDATA[engineering-culture]]></category>
            <category><![CDATA[technical-writing]]></category>
            <category><![CDATA[eu-ai-act]]></category>
            <category><![CDATA[content-provenance]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Luca Capra]]></dc:creator>
            <pubDate>Thu, 27 Aug 2026 09:31:04 GMT</pubDate>
            <atom:updated>2026-08-27T09:31:04.815Z</atom:updated>
            <content:encoded><![CDATA[<p><em>How we use generative models to get technical knowledge published, and what has to be built around them before the result is worth anyone’s time.</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*YqbHhGC9cWHcrIqYSAGZQw.png" /><figcaption>An orchestrated process: expertise enters, constraints shape the draft, a human decides what leaves.</figcaption></figure><p>Somewhere in your organisation there is an engineer who has spent three years on a problem almost nobody else understands. They know why the obvious approach fails, what it cost to find that out, and which received opinions in their field are wrong. None of it has ever been written down — not from reluctance, but because publishable prose is a different craft from the one they trained in, and the first paragraph is where most attempts die.</p><p>Generative models close that gap. Left to themselves they close it with filler: fluent, symmetrical text that readers now recognise and distrust within a sentence. We have spent the last four months building an editorial process on a wager — that what separates those two outcomes is not the model at all, but everything arranged around it.</p><h3>What the model is actually for</h3><p>Stating the division of labour honestly is the foundation, so we state it first. The model supplies structure, first sentences, and a willingness to produce a bad draft quickly: the thing most experts cannot find the time to make themselves. Knowledge is not on that list.</p><p>Every fact in a piece like this one came from somewhere a person had already been: a repository, a standards document, a measurement someone took. The draft is scaffolding around expertise that already existed, expressed as a technical document or a software artefact. When an author cannot say where a sentence came from, the sentence goes.</p><p>That rule does more work than any other. It is also why both “AI wrote this” and “a human wrote this” describe the situation badly, and why we would rather explain the arrangement than pick one.</p><h3>Constraints on what it may say</h3><p>Two constraints do most of the visible work, and both are boring in a useful way.</p><p>The first is a ceiling on claims. Every article is assigned a maturity level before drafting starts (from work that is running and published, down to work that is conceptual) and the language is held to it. A ceiling is not a target. An article about a reference implementation may not describe a production rollout, however much better that would look. The author who wants to claim more has to make it real first.</p><p>The second is a rule about sources. Public, primary, named in full and linked at first mention. Vague appeals to “the literature” or “the European framework for this” are not plain language; they are unverifiable claims wearing it as a disguise. If a source cannot be named and linked, the claim needs a better source.</p><p>Neither is clever. Both are things a model violates quietly unless something holds it to them.</p><h3>The line private material never crosses</h3><p>Most of what makes our work worth reading lives in places that cannot be published: internal repositories, unpublished deliverables, consortium documents. Pretending otherwise would be convenient. It is also how organisations leak.</p><p>So the rule runs one way. Private material may flow <em>inward</em>: an agent may read internal code to work out what is worth writing about, and what it learns may shape the argument. Nothing identifiable may flow outward. No file path, no class or module name, no configuration key, no hostname, no internal metric, no project codename that outsiders have never heard. Where a private detail matters, it is abstracted to the public-safe concept it represents. Where it cannot be, it is dropped and flagged as a gap needing a public source.</p><p>One question is asked of every sentence before it leaves a draft: could a reader trace this back to something unpublished? A cheap test, and it catches a surprising amount — the leaks that matter are rarely dramatic. They are a variable name in an innocent example.</p><p>This article is subject to that rule too, which is why it describes a method and shows you no repository. The tooling behind all of this is not public. Saying so is more honest than the alternative. But let us know if there is interest, and we will consider open sourcing!</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*y85WmY1GByKn7WO0KW-3ww.png" /><figcaption>The information-flow rule in one picture: private material may inform the argument, and nothing traceable to it may reach the page.</figcaption></figure><h3>Fluent is not the same as readable</h3><p>Here is where we were most wrong at the start.</p><p>We assumed the giveaways would be the familiar ones — “It is important to note”, “In conclusion”, sentences opening with “Moreover”, adjectives like “crucial” and “seamless” arriving in threes. Then we measured our own back catalogue — the pieces already published here, which you can go and read — against that list, and found essentially none of it. The guidelines had already banned that register, and the drafts were clean.</p><p>The real signature was structural, and once seen it could not be unseen. Every section of every article ended on a small aphorism — a neat, quotable closing line. Individually each was defensible. Read in sequence they made the argument feel performed rather than reasoned, and a reader who cannot say why an article annoys them is usually reacting to a rhythm like that. Two other habits sat alongside it: every section a single wall of text, and roughly two sentences in five opening the same three ways.</p><p>None of that is caught by a checklist of banned phrases, and none of it showed up in the statistical measures we tried — sentence-length variance, vocabulary range, the spread of both across a piece. On those our writing scored comfortably human while being unmistakably machine-drafted. Detection research is wider than what we tested.</p><p>So the fix is not a ban list but budgets, and a report showing where they went. Budgets, because a rule driven to zero just gets an author swapping synonyms while the structure stays put — and the structure was the problem. The report never blocks publication. It only shows where to look.</p><h3>The human is the gate, and the law agrees</h3><p>Since 2 August 2026, <a href="https://eur-lex.europa.eu/eli/reg/2024/1689/oj">Article 50 of the EU AI Act</a> has applied. Text generated by AI and published to inform the public on a matter of public interest has to be disclosed as artificially generated. Unless it has undergone human review or editorial control, and a person or organisation holds editorial responsibility for it. Both conditions, together.</p><p>The <a href="https://digital-strategy.ec.europa.eu/en/policies/guidelines-transparency-ai-generated-content">Commission’s guidance</a> rewards reading precisely. It names exactly one thing that does not count: superficial or formal checks, spell-checking and grammar. The rest is inference, ours included. It defines human review as deliberate examination by <em>natural persons</em> with relevant knowledge and judgement — so an editorial policy sitting in a repository is not review, a wave-through is not editorial control, and an agent checking a draft is neither (ouch!). The law stops there — it does not say how to build your pipeline. Our conclusion: the automated checks in our process are advisory reports and never gates, and none may write into the file where a human records their review. A design choice the rule invites, not one it imposes.</p><p>A separate obligation (and we got confused with this one) falls on whoever provides the AI system: under Article 50(2) they must mark generated output so that it is machine-detectable. That is the provider of the <em>system</em>, not whoever trained the model — not us either way, though its results reach us. Our illustrations arrive carrying a signed provenance manifest: it names the generating service, tags them with the <a href="https://cv.iptc.org/newscodes/digitalsourcetype/">IPTC term for AI-generated media</a>, and records an invisible watermark. Every generated image in our archive carries one, though publishing tools may strip them.</p><p>We checked. That is also how we learned how little it settles: a screenshot or a re-encode loses it. Watermarking is a backstop, and the Commission says a deployer cannot lean on the provider’s mark to discharge its own duty. No substitute, then, for a line at the foot of the article saying plainly what the reader is looking at. Ours says the text was drafted with AI assistance and reviewed by its named author, and that the illustrations are AI-generated — the exception covers text, not images.</p><h3>What this does not prove, and what we do next</h3><p>Very little of the above is settled. Six articles have gone out this way, under two bylines: enough to say the process runs, not enough to say it works. We have no reader study, no measure of whether these pieces landed better than the same authors would have managed unaided, and no counterfactual. The voice rules are weeks old and postdate most of what is already published.</p><p>One limit looks structural, and we do not expect to remove it. The process can record that a human reviewed a draft, and it can measure how much text changed between versions. Neither shows that judgement was exercised. An author might rewrite every sentence without checking a claim; another might verify every claim and change almost nothing. Nothing we generate tells those two apart, which is why the responsibility sits with the named author and not with the process.</p><p>The rest we expect to improve by running it. This is a method under evaluation: we are watching where drafts drift, which constraints get broken most often, and what reviewers change, and we will tighten whatever turns out to matter. Describing it in public is part of that — a process you have published is harder to let slide.</p><p>Then there is the user base. Not everyone can point a coding agent at a repository and load up the harness: some upload the plain markdown with instruction in a project context over web UIs, some use exotic setups and editor to do the job. There is much to explore there too, targeting a simplified interface and connected set of tools to the sources. Nice things to come, yet!</p><p>What we would claim today is narrower than where we think this ends up. The knowledge in a technical organisation is worth more than the writing skill it currently waits on, and the gap between them is closable — provided the closing is governed hard enough that the result deserves the reader’s attention.</p><p><strong>Links</strong></p><ul><li>The articles this process has produced, and the back catalogue measured above: <a href="https://medium.com/spindox-labs">https://medium.com/spindox-labs</a></li><li>The EU AI Act, Regulation (EU) 2024/1689: <a href="https://eur-lex.europa.eu/eli/reg/2024/1689/oj">https://eur-lex.europa.eu/eli/reg/2024/1689/oj</a></li><li>Commission guidelines on transparency of AI-generated content: <a href="https://digital-strategy.ec.europa.eu/en/policies/guidelines-transparency-ai-generated-content">https://digital-strategy.ec.europa.eu/en/policies/guidelines-transparency-ai-generated-content</a></li><li>Code of Practice on Transparency of AI-Generated Content: <a href="https://digital-strategy.ec.europa.eu/en/policies/code-practice-ai-generated-content">https://digital-strategy.ec.europa.eu/en/policies/code-practice-ai-generated-content</a></li><li>C2PA, the content provenance standard: <a href="https://c2pa.org/">https://c2pa.org/</a></li><li>IPTC digital source type vocabulary: <a href="https://cv.iptc.org/newscodes/digitalsourcetype/">https://cv.iptc.org/newscodes/digitalsourcetype/</a></li><li>SynthID text watermarking, open-sourced by Google DeepMind: <a href="https://github.com/google-deepmind/synthid-text">https://github.com/google-deepmind/synthid-text</a></li></ul><p><em>This article was drafted with AI assistance and reviewed, fact-checked and edited by its named author before publication. Its illustrations are AI-generated.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=972108c42801" width="1" height="1" alt=""><hr><p><a href="https://medium.com/spindox-labs/the-expertise-is-the-scarce-part-972108c42801">The expertise is the scarce part</a> was originally published in <a href="https://medium.com/spindox-labs">Spindox Labs</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[What Europe’s energy data space blueprints recommend you build]]></title>
            <link>https://medium.com/spindox-labs/what-europes-energy-data-space-blueprints-recommend-you-build-4ca184a191d3?source=rss----ab498e9e959d---4</link>
            <guid isPermaLink="false">https://medium.com/p/4ca184a191d3</guid>
            <category><![CDATA[dataspace]]></category>
            <category><![CDATA[open-source]]></category>
            <category><![CDATA[energy]]></category>
            <category><![CDATA[interoperability]]></category>
            <dc:creator><![CDATA[Luca Capra]]></dc:creator>
            <pubDate>Wed, 05 Aug 2026 10:01:03 GMT</pubDate>
            <atom:updated>2026-08-05T10:01:03.336Z</atom:updated>
            <cc:license>http://creativecommons.org/licenses/by/4.0/</cc:license>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*58vfebwrR4UREoSJLZOwFA.png" /><figcaption>The generic blueprint gives you the vocabulary. The energy one settles the questions it leaves open.</figcaption></figure><p>If you accept the argument — and European legislation increasingly requires data to be accessible, portable, and shared under defined conditions — that energy data should be shared under the owner’s control rather than locked in silos or surrendered wholesale, then the next question is immediate and practical: what should you build?</p><p>Europe has answered that question twice. Once for data spaces in general, once for energy specifically. Understanding how those two answers fit together is the difference between building something that connects to the rest of the continent and building something that does not.</p><h3>Two blueprints, one hierarchy</h3><p>The first answer is the cross-sector blueprint published by the <a href="https://blueprint.dssc.eu/">Data Spaces Support Centre</a>. It breaks any data space — energy, health, agriculture, mobility — into a set of technical building blocks: how participants establish trust, how they prove who they are, how they attach usage rules to data, how they describe and discover datasets, how they exchange data under a contract, and how they track where data came from and what happened to it. Version 3.0 was released in early 2025 as the concluding act of the Support Centre’s first phase. It is the shared vocabulary that every European data space now speaks.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*k6xPDFUZMdd7eCw6hMvhMQ.png" /><figcaption>Build to the sector-specific blueprint; describe what you built in the language of the generic one.</figcaption></figure><p>The second answer is the energy-specific blueprint, published in mid-2024 by the <a href="https://intnet.eu/images/resources/Blueprint_CEEDS_v2.pdf">interoperability network for the energy transition</a>. It takes the generic building blocks and settles the questions the cross-sector blueprint deliberately leaves open: which roles exist in the energy sector — grid operators, distribution operators, flexibility providers, prosumers, energy communities — which vocabularies describe energy data, which regulations govern access, and which existing systems the data space has to connect to.</p><p>The energy blueprint is a specialisation of the cross-sector one, not a replacement. Build to the generic building blocks for the foundation; layer the energy-specific requirements on top to speak the sector’s language.</p><p>This matters because the two blueprints were written at different times. The energy blueprint was written against version 1.0 of the cross-sector reference. The cross-sector blueprint has since evolved to version 3.0, which is considerably more detailed on trust, identity, policy enforcement, and provenance. The practical consequence is that an implementation aligned with the latest cross-sector version covers more recent generic ground, while the energy blueprint continues to supply sector-specific requirements that need separate attention. The gap is not a conflict — it is additive.</p><h3>What the building blocks actually ask for</h3><p>Strip away the formal language and what you need to build is surprisingly concrete.</p><p>Trust and identity. Participants need identities and credentials that can be verified according to the data space’s trust framework. The blueprint does not prescribe a single mechanism — federated identity, certificates, decentralised credentials, or hybrid models are all valid — but the result must be the same: when a participant claims to be who they say they are, others can check.</p><p>Usage rules that travel with the data. When a dataset owner says “this may be used for grid planning but not for marketing,” that rule should be expressed in a machine-readable form and linked to the offering or agreement. It can then be interpreted and, where technically feasible, enforced during negotiation, transfer, and subsequent processing — though once data has been received, enforcement may also rest on contractual obligations or audit controls. This is the mechanism that makes “shared but controlled” more than a slogan.</p><p>Descriptions and discovery. Datasets need standard descriptions: what they contain, who owns them, how they may be used, when they were last updated. Those descriptions are published through catalogues — which may be centralised, federated, or distributed depending on the governance model — so that participants can search across them to find what they need.</p><p>Exchange under contract. Data moves after both sides agree to terms. The provider offers; the consumer requests; the two sides negotiate until a machine-readable agreement is established under the data space’s governance framework; only then is access granted and the data transferred. Where supported by the implementation, consent or authorisation revocation can invalidate future access and interrupt active transfers — though data already received remains subject to the applicable agreement and legal obligations.</p><p>Provenance. Relevant lifecycle events — a dataset published, a contract negotiated, a transfer completed, a consent revoked — should be recorded according to the data space’s audit and accountability requirements. Not for bureaucracy, but because a regulator, an auditor, or a data owner who wants to know what happened to their data should be able to find out.</p><p>These are the cross-sector building blocks. They apply whether you are exchanging automotive supply-chain data or household energy consumption readings. The energy blueprint accepts all of them and adds its own layer on top: the sector’s standard vocabularies for grid models and smart appliances, a role model that maps the actors who actually operate in the energy system, real-time and near-real-time considerations for use cases that depend on grid operations or flexibility, and governance that accounts for energy-specific regulation.</p><h3>Where we built to this</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*kRvhdczkU5ga7WY1C7AwJg.png" /><figcaption>Each component maps to a named building block, which is what makes the stack describable in the same terms as everyone else’s.</figcaption></figure><p>We set out to build an implementation that maps explicitly to the cross-sector building blocks and is benchmarked against the energy blueprint. The platform is open source, Apache-licensed, and <a href="https://github.com/spindoxlabs/ds">published on GitHub</a>.</p><p>Each piece of the platform corresponds to a named building block. Participants carry verifiable digital identities. Data governance rules are written in a structured file and compiled into machine-readable offers that specify who may access what, under which conditions. Contract negotiation and access-controlled data transfer run through an open connector framework. A shared catalogue describes available datasets in a standard format. A provenance service logs the full lifecycle of every exchange. And a consent portal lets data subjects grant and revoke permission, with revocations reaching transfers that are already in progress.</p><p>We then ran an internal line-by-line comparison between the cross-sector and energy blueprints — the methodology and evidence are <a href="https://github.com/spindoxlabs/ds/tree/main/docs">in the repository’s documentation</a> — and assessed where the platform stands. The assessment indicates broad alignment with the cross-sector building blocks, and an energy-specific gap that is additive, not corrective. What needs to be added is the energy domain layer. What does not need to happen is redoing the foundation.</p><p>Complementary application-level components are developed within the <a href="https://github.com/celine-eu">CELINE project</a>, an EU-funded initiative building an open-source toolbox for energy communities. The CELINE repositories, with documentation at <a href="https://celine-eu.github.io/">celine-eu.github.io</a>, cover ingestion, dashboards, forecasting, flexibility, and an openly licensed vocabulary for community energy — the application layer that sits on top of the data space infrastructure.</p><h3>What is not done yet</h3><p>Honesty about scope is the price of credibility, so here is what is still open.</p><p>There is no tagged release. The platform is working code — you can clone it, deploy it, and run the full exchange flow — but we have not cut a version number, and the interfaces will move.</p><p>The energy-domain vocabulary layer is not built. The platform’s vocabulary component is designed to be extended with sector-specific models, but it does not yet host the standard vocabularies that grid operators already use or the shared vocabulary for smart appliances and energy devices. That is the single highest-impact gap we identified in the benchmark.</p><p>Marketplace and compensation mechanisms discussed in the energy blueprint — including possible token-based and barter models for data exchange — are not implemented. The platform handles catalogue, contract, and transfer, but not pricing or settlement.</p><p>Legacy integration is designed for but not yet demonstrated. The energy sector runs on operational systems that predate data spaces by decades. Connecting to them is primarily an engineering challenge, but it may also require architectural decisions around adapters, semantics, and security.</p><p>And the energy blueprint itself was written against an earlier version of the cross-sector reference. A synchronised version that accounts for the evolution to version 3.0 does not yet exist. Future revisions may bring the two into closer alignment, but today anyone building has to hold both documents and reconcile them.</p><h3>The practical hierarchy</h3><p>If you are building for the energy sector, here is the working order.</p><p>Build to the cross-sector building blocks. They are the foundation, the shared language, and the part that connects your data space to every other one in Europe.</p><p>Layer the energy blueprint’s domain requirements on top. It tells you which roles, which vocabularies, which regulations, and which integration points the energy sector specifically needs.</p><p>Use open, maintained implementations for the building blocks you can. Open-source implementations exist for exchange, identity, and catalogue — their suitability should be assessed against your project’s requirements and production readiness.</p><p>And be honest about what is not yet settled. The energy vocabulary layer, the marketplace model, and the synchronisation between the two blueprints are all open questions. Building on the cross-sector foundation means those questions do not require starting over — they require adding a layer.</p><p>The <a href="https://blueprint.dssc.eu/">cross-sector blueprint</a> and the <a href="https://intnet.eu/images/resources/Blueprint_CEEDS_v2.pdf">energy blueprint</a> are both public documents. Our implementation is at <a href="https://github.com/spindoxlabs/ds">github.com/spindoxlabs/ds</a>, the community toolbox is at <a href="https://github.com/celine-eu">github.com/celine-eu</a>, and its documentation is at <a href="https://celine-eu.github.io/">celine-eu.github.io</a>.</p><p><em>This article was drafted with AI assistance and reviewed, fact-checked and edited by its named author before publication.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=4ca184a191d3" width="1" height="1" alt=""><hr><p><a href="https://medium.com/spindox-labs/what-europes-energy-data-space-blueprints-recommend-you-build-4ca184a191d3">What Europe’s energy data space blueprints recommend you build</a> was originally published in <a href="https://medium.com/spindox-labs">Spindox Labs</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Foundation models come to time series: what actually changed]]></title>
            <link>https://medium.com/spindox-labs/foundation-models-come-to-time-series-what-actually-changed-26e25d6a2fb7?source=rss----ab498e9e959d---4</link>
            <guid isPermaLink="false">https://medium.com/p/26e25d6a2fb7</guid>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[time-series-forecasting]]></category>
            <category><![CDATA[open-source]]></category>
            <category><![CDATA[energy]]></category>
            <dc:creator><![CDATA[Mattia Pavan]]></dc:creator>
            <pubDate>Tue, 04 Aug 2026 10:01:04 GMT</pubDate>
            <atom:updated>2026-08-04T10:01:04.338Z</atom:updated>
            <content:encoded><![CDATA[<p>Pretrained forecasters promise predictions for meters they have never seen. A field guide before we put them to work.</p><figure><img alt="The wake is the history, the widening cone is the forecast: a model pretrained on everyone’s voyages, navigating waters it has never seen." src="https://cdn-images-1.medium.com/max/1024/1*f5KLPkSaRppbSH3vZ5n8RQ.png" /><figcaption>The wake is the history, the widening cone is the forecast: a model pretrained on everyone’s voyages, navigating waters it has never seen.</figcaption></figure><p>For years, the recipe for forecasting electricity demand was stable: collect history for the thing you care about, engineer features, train a model on that history, and start again for the next dataset. Large language models broke that recipe for text: one model, pretrained once, serving many tasks. Time series is now having the same moment. A new family of pretrained forecasters claims something that would have sounded implausible a few years ago: point them at a meter they have never seen, give them a window of recent readings, and they return a forecast with no training at all. For energy communities, where many small meters each carry too little history to justify a hand-built model, the promise is obvious. Before this series presents our own forecasting engine and the lessons of running these models on community data, it is worth establishing what they actually are, and what, as of mid-2026, they can and cannot be trusted to do.</p><h3><strong>One model, pretrained on everyone’s history</strong></h3><p>The models in question are usually called time series foundation models, and the recipe is borrowed directly from language modelling: pretrain a single network on an enormous and deliberately diverse corpus of time series (electricity load, traffic, sales, weather, web metrics) and then ask it to forecast new series without any further training, in what its authors call zero-shot forecasting. Four openly released families stand out here, and not by accident: they are the ones our own engine runs as backends. <em>Chronos</em> (https://github.com/amazon-science/chronos-forecasting), from Amazon Science, treats forecasting as language modelling over discretised values, and has evolved through the faster Chronos-Bolt into <em>Chronos-2</em> (https://arxiv.org/abs/2510.15821), which extends zero-shot use to multivariate and covariate-informed tasks. <em>TimesFM </em>(https://github.com/google-research/timesfm), from Google Research, is a decoder-only model described in a <em>paper</em> presented at ICML 2024 (https://arxiv.org/abs/2310.10688). <em>Moirai </em>(https://github.com/SalesforceAIResearch/uni2ts), from Salesforce AI Research, was pretrained on a corpus its authors report at over twenty-seven billion observations across nine domains. And <em>Tiny Time Mixers </em>(https://github.com/ibm-granite/granite-tsfm), from IBM, takes the opposite bet: models starting at around one million parameters, small enough to run comfortably on a CPU. One licensing note matters in practice: Chronos, TimesFM and Tiny Time Mixers publish their model weights under the permissive Apache 2.0 licence, while Moirai’s weights are released for non-commercial use, with only its code under Apache 2.0. The claim shared by all four is the interesting part: forecasting competence can be learned once, from everyone’s history, and transferred to your data. Zero-shot is the headline, but not the only mode: every one of these models can also be fine-tuned, nudged with a modest amount of your own data so the generalist becomes a specialist. Tiny Time Mixers is explicitly designed for that few-shot adaptation, and the main public benchmark distinguishes zero-shot submissions from models that saw more, because the two are different claims.</p><h3><strong>The same trick everywhere: cut the series into patches</strong></h3><p>Under the hood, these models had to solve a problem language models never faced: numbers are not words. The original Chronos solved it by making them words: scaling each series into a common range, quantising values into a fixed vocabulary of bins, and training a language-model architecture on the resulting tokens with an ordinary cross-entropy loss. It worked, but one token per time step makes long histories expensive. The answer these families converged on is patching: cut the series into short windows of consecutive values (a day of hourly readings, say) and embed each window as a single token. Chronos-Bolt chunks its context into patches and forecasts multiple steps directly, which its authors report makes it dramatically faster than the original. TimesFM feeds patches to a decoder-only transformer. The original Moirai, a masked encoder, varies the patch size with the frequency of the data, so a monthly series and a minutely one are both carved at a sensible grain; its successor, <em>Moirai 2</em> (https://arxiv.org/abs/2511.11698), moved to a decoder-only design with a single patch configuration. Tiny Time Mixers dispenses with attention altogether, applying adaptive patching inside an all-MLP mixer architecture. Around the patches sits a second, quieter piece of machinery: input normalisation, the rescaling that lets a household meter and a national load curve, or a flat overnight baseline and a solar spike, pass through the same weights. The families handle it differently, some building the rescaling in and others expecting pre-scaled inputs at the door, and the choice matters more than it looks for energy data, which is spiky, intermittent and spread across wildly different magnitudes.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*tnAH4EfzrmtwySCN0Mkirw.png" /><figcaption>The landscape, not a leaderboard: the four ways today’s time series foundation models are built (values as tokens, patch decoders, masked encoders and compact mixers) share one map and one destination, a forecast with honest uncertainty.</figcaption></figure><h3><strong>Forecasts that say how sure they are</strong></h3><p>Most of these models can answer with more than a single number. The original Chronos samples many future trajectories and reads probabilities off the spread. The original Moirai predicts the parameters of a mixture distribution, from which any interval can be drawn, while Moirai 2 switched to direct quantile forecasts. TimesFM 2.5 adds an optional quantile head to its point forecasts, and the recent Chronos generations return quantile forecasts natively: the model states, for each future hour, the level it expects the truth to fall under ten percent of the time, half the time, ninety percent of the time. Tiny Time Mixers is the exception that proves the rule: it is a point forecaster, and any statement of uncertainty has to be wrapped around it from outside. This is not an academic nicety. A community deciding when to charge a battery or shift a load is not asking “what is the most likely value” but “what is the range I should plan for”, and a forecast that carries its own uncertainty is the difference between a guess and a decision aid. The honest caveat is that a model can state intervals and still be miscalibrated on your data: a band that claims ninety percent coverage has to be checked against reality, series by series, before anyone should plan around it. How that checking is done, and how intervals can be re-calibrated when they are off, is a story for the next article in this series.</p><p><strong>## Benchmarks, leaderboards and the scaling question</strong></p><p>How good are they? The most useful public answer comes from <em>GIFT-Eval </em>(https://github.com/SalesforceAIResearch/gift-eval), a benchmark from Salesforce AI Research built specifically for general-purpose forecasters: ninety-eight dataset configurations across seven domains and frequencies from seconds to years, scored on both point accuracy and the quality of probabilistic forecasts, with a <em>public leaderboard</em> (https://huggingface.co/spaces/Salesforce/GIFT-Eval) anyone can inspect. Two features of that leaderboard deserve more attention than the rankings themselves. First, it carries an explicit test-data-leakage flag per model, because the awkward question for any pretrained forecaster is whether it has already seen the benchmark’s test series inside its pretraining corpus; the benchmark ships a separate pretraining corpus that does not overlap its test split, precisely to make untainted training possible. Second, and this is our own reading of the public leaderboard rather than a claim the benchmark makes: scale has not settled the field the way it did in language. Compact models often sit remarkably close to models many times their size, and no single family dominates every metric. Whether time series forecasting scales like language modelling is, as of mid-2026, a genuinely open research question. For a practitioner, both observations point the same way: treat every headline number with care, and re-evaluate on your own data.</p><h3><strong>The counterpoint: the old way has not conceded</strong></h3><p>Against all of this stands a much less glamorous fact: for tabular, feature-rich forecasting problems, gradient-boosted decision trees remain the baseline any newcomer must be measured against. An open library like <em>LightGBM</em> (https://github.com/microsoft/LightGBM), fed with calendar features and weather data, trains on the actual meter being forecast (its holidays, its habits, its rooftop panels) on modest hardware, and can be retrained as often as the data changes. A pretrained model knows the world’s time series; a tree trained on your meter knows your meter. Which knowledge wins is an empirical question that depends on how much history the meter has and how typical its behaviour is, and it is exactly the kind of question that should be settled by evaluation, not by fashion. That is also why any serious comparison keeps a naive baseline on the floor: energy data is so strongly patterned by daily and weekly rhythm that “the next two days will look like the same hours last week” sets a floor that is famously awkward to clear, and a result reported without that floor says nothing at all.</p><h3><strong>What this means for an energy community</strong></h3><p>Time series forecasting now has genuine foundation models: openly released, probabilistic in most configurations, and able to produce a forecast for a meter they have never seen. It also still has gradient-boosted trees that train in seconds and know the data intimately, a benchmark culture honest enough to flag its own contamination risks, and an unresolved argument about whether scale helps. The practical conclusion we draw is deliberately unheroic: this is not a moment to bet an energy community’s infrastructure on any single model family; it is a moment to make models replaceable, so that today’s choice can be revisited the day the leaderboard moves. The claims in this article trace to public papers, repositories and leaderboards, linked below, and where a judgement is our own reading we have said so; we have deliberately not ranked these models, because rankings belong to evaluations you can reproduce. The next article in this series shows the engine we built on exactly that principle (one engine, many models), and the one after it reports what we learned when community smart-meter data met all of them.</p><p><strong>Links</strong></p><p>- Chronos, Chronos-Bolt and Chronos-2 (Amazon Science): <a href="https://github.com/amazon-science/chronos-forecasting,">https://github.com/amazon-science/chronos-forecasting,</a> papers: <a href="https://arxiv.org/abs/2403.07815,">https://arxiv.org/abs/2403.07815,</a> <a href="https://arxiv.org/abs/2510.15821">https://arxiv.org/abs/2510.15821</a></p><p>- TimesFM (Google Research): <a href="https://github.com/google-research/timesfm,">https://github.com/google-research/timesfm,</a> paper: <a href="https://arxiv.org/abs/2310.10688">https://arxiv.org/abs/2310.10688</a></p><p>- Moirai / Uni2TS (Salesforce AI Research): <a href="https://github.com/SalesforceAIResearch/uni2ts,">https://github.com/SalesforceAIResearch/uni2ts,</a> papers: <a href="https://arxiv.org/abs/2402.02592,">https://arxiv.org/abs/2402.02592,</a> <a href="https://arxiv.org/abs/2511.11698">https://arxiv.org/abs/2511.11698</a> (Moirai 2)</p><p>- Tiny Time Mixers (IBM): <a href="https://github.com/ibm-granite/granite-tsfm,">https://github.com/ibm-granite/granite-tsfm,</a> paper: <a href="https://arxiv.org/abs/2401.03955">https://arxiv.org/abs/2401.03955</a></p><p>- GIFT-Eval benchmark and leaderboard: <a href="https://github.com/SalesforceAIResearch/gift-eval,">https://github.com/SalesforceAIResearch/gift-eval,</a> <a href="https://huggingface.co/spaces/Salesforce/GIFT-Eval">https://huggingface.co/spaces/Salesforce/GIFT-Eval</a></p><p>- LightGBM: <a href="https://github.com/microsoft/LightGBM">https://github.com/microsoft/LightGBM</a></p><p>- The CELINE forecasting engine: <a href="https://github.com/celine-eu/celine-forecasting">https://github.com/celine-eu/celine-forecasting</a></p><p><strong>— -</strong></p><p><em>*This article was drafted with AI assistance and reviewed, fact-checked and edited by its named author before publication.*</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=26e25d6a2fb7" width="1" height="1" alt=""><hr><p><a href="https://medium.com/spindox-labs/foundation-models-come-to-time-series-what-actually-changed-26e25d6a2fb7">Foundation models come to time series: what actually changed</a> was originally published in <a href="https://medium.com/spindox-labs">Spindox Labs</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[One engine, many models: forecasting energy without betting on a single approach]]></title>
            <link>https://medium.com/spindox-labs/one-engine-many-models-forecasting-energy-without-betting-on-a-single-approach-0de97f8cb51e?source=rss----ab498e9e959d---4</link>
            <guid isPermaLink="false">https://medium.com/p/0de97f8cb51e</guid>
            <category><![CDATA[time-series-forecasting]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[open-source]]></category>
            <category><![CDATA[energy]]></category>
            <dc:creator><![CDATA[Mattia Pavan]]></dc:creator>
            <pubDate>Mon, 03 Aug 2026 07:22:16 GMT</pubDate>
            <atom:updated>2026-08-03T07:22:14.963Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Npzrvtr8zz3gnH2n81QtFQ.png" /><figcaption>One engine, many models: the forecasting method is the durable asset; the models are replaceable parts.</figcaption></figure><p>How much electricity will your neighbourhood draw tomorrow evening? How much will the rooftop panels feed back at noon? Every plan an energy community makes (when to charge a battery, when to shift a load, how to share what it produces) quietly rests on an answer to those two questions. The uncomfortable truth about answering them with machine learning is that models age quickly: the public leaderboards that rank forecasting models churn season after season. So instead of betting on a model, we built and published the thing that outlives them all: an open forecasting engine where the data contract, the evaluation protocol and the uncertainty estimates are the durable assets, and the models themselves are replaceable parts.</p><h3><strong>Forty-eight hours ahead, one meter at a time</strong></h3><p>The forecasting engine in the CELINE toolbox answers a concrete question: for each smart meter in a community, what will the next forty-eight hours look like, hour by hour? It works from ordinary quarter-hour smart-meter readings, aggregated to hourly energy values, and it forecasts two things per meter: the energy drawn from the grid and the energy fed back into it, the second one being how rooftop solar shows up in the data. Both matter, because a community’s room for manoeuvre lives in the gap between them. Weather is a major driver of both consumption and rooftop production, so the engine builds hourly weather features from [Open-Meteo](https://open-meteo.com/), a free and open weather service: historical values come from ERA5, the reanalysis dataset of the Copernicus programme, and the days ahead come from the forecast service, stitched into one continuous series. Two days ahead, one meter at a time, weather in hand: that is the raw material every downstream decision is made of.</p><h3><strong>A data contract instead of a dataset</strong></h3><p>The repository ships no meter data. What it publishes instead is the shape of the data: which columns a meter file must have, what the units mean, and that every timestamp must sit on a quarter-hour grid in a well-defined timezone. This contract is enforced the moment data is loaded: common column-name variants are mapped automatically, and violations fail early with a clear message rather than surfacing as a cryptic error deep inside model training. The same discipline extends to deciding which meters can be forecast at all. A device must show enough history (around six weeks) and enough coverage, with at least forty percent of its expected readings present, before a model is trained for it. A meter that does not clear those bars is not silently dropped: the engine produces a per-device report explaining why it was excluded. That is a small design choice with a large consequence: anyone can run the pipeline on their own community’s data, from files or straight from their own database, with the meter readings staying on their own infrastructure. The contract travels; the data does not: the same principle we have argued for throughout this series on energy data sovereignty.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*7D_jWuVDopY4OgskhYOqKA.png" /><figcaption>The shape of the engine: a public data contract feeds a model-agnostic core, and every model backend (from gradient-boosted trees to time-series foundation models) plugs into the same interface and is scored by the same evaluation protocol.</figcaption></figure><h3><strong>One engine, many models</strong></h3><p>The architecture draws a hard line between what depends on the model and what does not. On one side sits the engine: data cleaning, weather features, naive baselines, prediction intervals, bias correction, evaluation and experiment tracking. On the other side sit the models, each one a self-contained strategy behind a shared interface. The default is gradient-boosted trees, via the open LightGBM library, a workhorse that trains quickly on modest hardware. Alongside it, openly released foundation models for time series (Chronos, TimesFM, Moirai, IBM’s Tiny Time Mixers) plug into the very same slot behind optional dependencies. Switching approach is a command-line flag, not a rewrite. The pretrained backends run zero-shot exactly as released, and most of them can also be fine-tuned on the community’s own data through the same engine, with strategies that vary by backend, from lightweight adapters sized for a single consumer GPU to full fine-tuning. The adaptation obeys the same rule as everything else: it only ever sees the data a model is allowed to train on, so evaluation and calibration stay honest. The engine also decides how widely a model is trained: one model per individual meter, or one pooled model per group of similar devices, trained across the fleet and then validated against each meter individually. The point of all this plumbing is strategic, not aesthetic. Time-series forecasting is moving fast, and any system hard-wired to one model family will be obsolete on that family’s schedule. An engine that treats models as replaceable parts can adopt whatever comes next, and can prove, rather than assume, that the new thing is better.</p><h3><strong>An evaluation you cannot cheat</strong></h3><p>Proving it is the part most forecasting claims skip. The engine’s evaluation protocol is built around two rules. First, a naive baseline is always in the room: the seasonal-naive forecast, which simply predicts that the next two days will look like the same hours one week earlier. It is always evaluated, whether or not anyone asks for it, because energy data is so strongly patterned by daily and weekly rhythms that this simple guess sets a demanding floor, and any model that cannot beat it has no business in production. Skill is measured as improvement over that floor. Second, evaluation is leakage-free by construction. The backtest rolls the forecast origin through history: at each origin, models are retrained using only the data available up to that moment, then scored on the following forty-eight hours, exactly as they would have been used in reality. Every candidate (every model, every training scope) is scored on identical device, target and origin cells, so comparisons are apples to apples by design rather than by discipline. What this article deliberately does not give you is the results. Those belong to the next article in this series, which reports what we learned from putting every backend through this protocol, and the benchmark suite itself is published in the repository, where anyone can run it on their own data.</p><h3><strong>A forecast that says how sure it is</strong></h3><p>A single predicted number is a poor basis for a decision; what a community actually needs to know is the range the truth is likely to fall in. The engine treats that range as a first-class output. Quantile models produce an initial band, and Conformalized Quantile Regression, a method by [Romano, Patterson and Candès](https://arxiv.org/abs/1905.03222), then adjusts that band on held-out calibration data, so that an interval’s claimed coverage is checked and corrected against data the model never trained on, rather than taken on trust. Coverage is tracked as a first-class metric alongside accuracy, and the same calibration machinery can wrap model backends that only produce point forecasts, so a forecast leaving the engine can carry a calibrated band whichever model produced it. The honesty runs in both directions. Where the data supports confidence, the bands can stay tight; where it is thin or noisy, honest calibration produces wider ones, and meters with too little history are flagged out rather than force-fitted. A forecast that admits what it does not know is worth more than one that is confidently wrong.</p><h3><strong>The method is the product</strong></h3><p>What we have published is not a model; it is a method, running as code. A data contract anyone can shape their own data to. An engine that separates the durable machinery from the replaceable models. A baseline that keeps every claim honest, an evaluation protocol built not to peek at the future, and uncertainty estimates that are calibrated rather than decorative. It is a working reference implementation, released under the Apache 2.0 licence, not yet a documented production rollout, and we are deliberate about that distinction. The results are deliberately absent too: methodology first, lessons next. The follow-up article reports what we learned from putting every backend through this benchmark: which classes of model delivered, where accuracy degrades, and where the honest limits sit.</p><p><strong>Links</strong></p><p>- The forecasting engine: <a href="https://github.com/celine-eu/celine-forecasting">https://github.com/celine-eu/celine-forecasting</a></p><p>- The CELINE open-source organisation: <a href="https://github.com/celine-eu">https://github.com/celine-eu</a></p><p>- Open-Meteo, the open weather service used for features: <a href="https://open-meteo.com/">https://open-meteo.com/</a></p><p>- Conformalized Quantile Regression (Romano, Patterson &amp; Candès, 2019): <a href="https://arxiv.org/abs/1905.03222">https://arxiv.org/abs/1905.03222</a></p><p><em>*This article was drafted with AI assistance and reviewed, fact-checked and edited by its named author before publication.*</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=0de97f8cb51e" width="1" height="1" alt=""><hr><p><a href="https://medium.com/spindox-labs/one-engine-many-models-forecasting-energy-without-betting-on-a-single-approach-0de97f8cb51e">One engine, many models: forecasting energy without betting on a single approach</a> was originally published in <a href="https://medium.com/spindox-labs">Spindox Labs</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Why energy systems still struggle to talk to each other]]></title>
            <link>https://medium.com/spindox-labs/why-energy-systems-still-struggle-to-talk-to-each-other-2ddb3c58a555?source=rss----ab498e9e959d---4</link>
            <guid isPermaLink="false">https://medium.com/p/2ddb3c58a555</guid>
            <category><![CDATA[energy-transition]]></category>
            <category><![CDATA[interoperability]]></category>
            <category><![CDATA[dataspace]]></category>
            <category><![CDATA[energy]]></category>
            <dc:creator><![CDATA[Luca Capra]]></dc:creator>
            <pubDate>Sun, 02 Aug 2026 10:01:03 GMT</pubDate>
            <atom:updated>2026-08-02T10:01:03.030Z</atom:updated>
            <cc:license>http://creativecommons.org/licenses/by/4.0/</cc:license>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*N0qc3LoI63FPu8wgqgoAJw.png" /><figcaption>The wires connect. The meaning does not always survive the trip.</figcaption></figure><p>A modern energy community is a conversation between systems that were never introduced. Solar panels on a roof, a battery in a garage, a car charging overnight, a smart meter reporting every quarter of an hour, a grid operator watching a substation, a national body checking that shared energy adds up and that incentives are paid correctly. For any of this to work, all of these systems need to understand each other. Too often, they do not.</p><p>This is the quiet problem underneath the energy transition. We picture the hard parts as physical: more panels, bigger batteries, stronger cables. But once the hardware is in place, what fails is less visible. A number leaves one system meaning one thing and arrives meaning something slightly different, or nothing at all. The engineering word for getting this right is interoperability, and it is worth understanding why it is so stubborn.</p><h3>It is not one problem, it is four</h3><p>The <a href="https://interoperable-europe.ec.europa.eu/collection/nifo-national-interoperability-framework-observatory/european-interoperability-framework-detail">European Interoperability Framework</a>, the Commission’s reference model for making systems work together, makes the point that “talking to each other” is not a single thing: it separates legal, organisational, semantic and technical interoperability. In energy practice the technical layer is worth splitting in two, because moving bytes and agreeing on how they are laid out fail in different ways. That gives four levels, and a system can pass on one while failing on the next.</p><p>The first level is the plumbing. Can two systems physically exchange bytes over a network, securely, without one of them hanging up on the other? The sector has shared protocols for moving data safely from machine to machine, and this is usually not the hardest layer to solve, though legacy devices and proprietary interfaces still cause grief in the field.</p><p>The second level is shape. When the data arrives, is it laid out in a structure the receiver can open and read? Common interchange formats exist, and richer context can travel as linked data alongside the values. This level is comparatively mature, and where it breaks it is usually cheap to fix.</p><p>The third level is meaning. Two systems can exchange a perfectly formed file and still misunderstand each other completely. One system’s “energy withdrawn” is another’s “energy consumed net of local production.” One records a reading at the start of an interval, another at the end. Nothing in the plumbing or the file shape catches this. The numbers line up; the meaning does not.</p><p>This is where interoperability most often fails in energy, and the landscape of standards shows why. The map grid operators use for the physical network — the <a href="https://webstore.iec.ch/en/publication/62698">Common Information Model</a>, standardised by the IEC as IEC 61970 and IEC 61968 — describes substations, feeders and load flows. <a href="https://saref.etsi.org/">SAREF and its energy extension SAREF4ENER</a>, open standards from ETSI, describe meters and demand response from the building’s point of view. <a href="https://www.dlms.com/">DLMS/COSEM</a>, from the DLMS User Association, speaks a compact, device-level language about registers and profiles. IEC 62325 models schedules and bids between market parties. The <a href="https://openchargealliance.org/protocols/open-charge-point-protocol/">Open Charge Point Protocol</a>, from the Open Charge Alliance, has its own vocabulary again. Each serves a real community of practice. None was designed to be understood by the others.</p><p>The fourth level is rules and responsibility. Even when two systems agree on meaning, the organisations behind them must agree on who may see what, under which consent and legal basis, for which purpose, for how long, and who is accountable when something goes wrong — plus auditability, revocation, and change control over the shared definitions themselves. The <a href="https://www.cencenelec.eu/media/CEN-CENELEC/AreasOfWork/CEN-CENELEC_Topics/Smart%20Grids%20and%20Meters/Smart%20Grids/4_sgcg_methodology_sgamusermanual.pdf">Smart Grid Architecture Model</a>, the European standards bodies’ reference architecture for the sector, organises this work into five interoperability layers; the <a href="https://intnet.eu/">int:net</a> whitebook proposes a sixth, “Framework”, because the political and regulatory dimension does not fit in the five that exist. Governance is a design problem, not paperwork around one.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Nl1p5ucjPTNgMLNdEokjCQ.png" /><figcaption>Four layers. The lower two are comparatively mature; better plumbing will not fix the upper two.</figcaption></figure><p>The pattern is consistent. The lower two levels are comparatively mature and usually the easier ones to solve. The harder problems are shared meaning and shared rules.</p><h3>The market solves half the problem, twice</h3><p>The same split shows up in the products. Take one national market — Italy, where energy communities have a defined regulatory framework and a working incentive scheme — and look at the platforms a community can buy today next to the open frameworks built on the European data space blueprints. Almost everything leans toward one of two profiles.</p><p>The first is the administrative platform: very good at the paperwork of running a community under national rules, registering members, calculating the shared energy, producing the reports that release the incentives. These tools are mature and genuinely useful. What they are not built to do is act as one node in a wider exchange of data. Across the Italian platforms reviewed for this article, none was built around open, shared vocabularies for what its numbers mean, and only a few documented an API at all. That is not an oversight — nothing in the national rules asks for it — but their data cannot be consumed from outside without bespoke translation.</p><p>The second is the architectural framework: open, standards-aligned building blocks for exactly the sovereign, rule-governed exchange the first profile lacks — identity, consent, provenance, and the governance of who may use data for what. Some follow the <a href="https://blueprint.dssc.eu/">Data Spaces Support Centre blueprint</a> closely, mapping each capability to a named building block. What they are not is a finished product for running a community under a specific country’s regulations. They are the foundations and the beams, not the house.</p><p>Nothing reviewed here combined the two. One national market is not the whole of Europe, and the picture will differ where regulation pushes harder on data access. But the split looks structural rather than Italian: products follow the obligations their buyers actually have, and no obligation yet asks a community platform to be understood by systems outside it. That gap is the whole unsolved middle of the problem.</p><h3>The tempting wrong answer</h3><p>Faced with a room full of systems that disagree, the instinct is to declare a single standard and make everyone adopt it. It sounds decisive, and it almost never works, because the systems that need to cooperate were built for different worlds and carry years of investment. The grid operator’s model of the network has decades of tooling behind it. A community platform models something quite different: members, their share of a shared array, their consent, their bill.</p><p>The <a href="https://doi.org/10.5281/zenodo.12609569">Common European Energy Data Space blueprint</a>, published by int:net in 2024, makes the same point structurally: it lists ten domain-specific standards as complementary, not competing. Telling all of them to abandon what they have is expensive, slow, and almost always resisted — for the practical reason that nobody can afford it and the political one that whoever controls the standard controls everyone.</p><h3>Another approach, a shared vocabulary with translators</h3><p>The pattern recent standards-aligned work converges on is almost the opposite. Keep the standards people already use, add a thin shared vocabulary that connects them, and build translators between the layers.</p><p>“Thin” is a design constraint, not a shortcut: a deliberately small connecting layer holding only the cross-domain concepts and the mappings between them, with domain-specific detail left in the source standards. In practice it imports terms others have already defined — devices from the smart-appliance vocabulary, observations from the sensor vocabulary, buildings, communities — and adds only the glue needed to link them. Instead of redefining what a meter is, it says: a meter is the thing the smart-appliance standard already defined, and here is how it relates to the community it belongs to. This is the approach we are taking with the <a href="https://w3id.org/celine-eu">CELINE core ontology</a>, published at a permanent identifier under an open licence; the <a href="https://github.com/celine-eu/ontologies">source and the mappings</a> are open for anyone to check or reuse.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*kv5aHPQFUjt-xws1dpyr7w.png" /><figcaption>A thin shared vocabulary in the middle, translators on each spoke — and every existing standard left where it is.</figcaption></figure><p>Between them, small purpose-built translators move a reading from one world into another. A grid operator’s view of a feeder can meet a community’s view of its members’ consumption without either side giving up its own model. Each translator is narrow and testable, and the shared vocabulary is the contract both sides build against. Translators do not make semantic disagreement disappear — lossy mappings, mismatched granularity, conflicting units and time conventions are all still there. What they change is that the disagreement becomes explicit, testable, versioned and governable instead of silent. A translator between two maps is more durable than a demand that everyone throw their maps away.</p><h3>Why openness matters</h3><p>A connecting vocabulary only earns its keep if others can read it, trust it and build on it. A private, perfect model that no one else can see solves nothing, because the whole point is to be understood by systems you do not control. That is why the reference points here are public documents and public code: the framework that names the interoperability layers, the blueprints that list the domain standards, and open implementations that can be read as running code rather than described in a slide deck. It is also why a shared vocabulary needs a permanent, versioned, openly licensed address, so anyone can adopt and extend it instead of inventing another private model.</p><p>The gap between where the sector is and where it needs to be remains large, and no amount of better plumbing will close it, because the plumbing is not what is broken. The practical task is not to pick a winning standard. It is to publish the shared concepts, the mappings between existing standards, the validation rules that keep them honest, and the governance process that decides how they change, and then to build the translators one connection at a time.</p><p><em>Open work referenced in this piece:</em></p><ul><li><em>The connecting vocabulary for energy communities, at </em><a href="https://w3id.org/celine-eu"><em>w3id.org/celine-eu</em></a></li></ul><p><em>This article was drafted with AI assistance and reviewed, fact-checked and edited by its named author before publication.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=2ddb3c58a555" width="1" height="1" alt=""><hr><p><a href="https://medium.com/spindox-labs/why-energy-systems-still-struggle-to-talk-to-each-other-2ddb3c58a555">Why energy systems still struggle to talk to each other</a> was originally published in <a href="https://medium.com/spindox-labs">Spindox Labs</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The reference standards behind Europe’s energy data space]]></title>
            <link>https://medium.com/spindox-labs/the-reference-standards-behind-europes-energy-data-space-654eacf9afcc?source=rss----ab498e9e959d---4</link>
            <guid isPermaLink="false">https://medium.com/p/654eacf9afcc</guid>
            <category><![CDATA[open-source]]></category>
            <category><![CDATA[energy]]></category>
            <category><![CDATA[interoperability]]></category>
            <category><![CDATA[dataspace]]></category>
            <category><![CDATA[semanticweb]]></category>
            <dc:creator><![CDATA[Luca Capra]]></dc:creator>
            <pubDate>Thu, 30 Jul 2026 08:21:35 GMT</pubDate>
            <atom:updated>2026-07-30T08:21:34.312Z</atom:updated>
            <cc:license>http://creativecommons.org/licenses/by/4.0/</cc:license>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ImgFKoaqIjGPRQrm91iD7w.png" /><figcaption>Several reference standards, each already agreed for its own interface. The work is connecting them, not replacing them.</figcaption></figure><p>In the previous piece we argued that energy systems fail to talk to each other not at the level of wires or file formats but at the level of meaning: two systems can exchange a perfectly formed file and still misunderstand each other, because the words inside mean different things to each. The answer, we said, is not one universal model but to keep the standards people already use and connect them through a thin shared vocabulary, with translators in between.</p><p>That argument was deliberately abstract. This one is not. Europe’s <a href="https://intnet.eu/images/Blueprint_CEEDS_v3.0.pdf">Common European Energy Data Space blueprint</a> does not leave the vocabulary question unanswered. It identifies a set of prominent reference standards for the points where energy data has to cross between systems — the grid, the meter, the market, the device, the charger. It stops short of naming one exclusive model for every layer: the standards overlap, implementations vary, and reconciling them is part of the work. But it gives implementers a credible starting set.</p><h3>The operator’s side: the grid and the equipment in it</h3><p>Start where the power flows. Operators that exchange network models need a shared one — substations, transformers, feeders, switches and their connections — or one operator’s model cannot be read by another. The reference here is the <a href="https://www.entsoe.eu/data/cim/">Common Information Model</a>, a family of information standards maintained by the IEC that spans energy-management systems, distribution integration and electricity-market exchange. It is a very large model, describing the grid from a continental transmission corridor down to a single rural feeder, and its value lies as much in the relationships as in the components. Interoperability comes not from claiming support for “CIM” in the abstract but from agreeing on a particular profile, version and validation regime. ENTSO-E, which coordinates Europe’s transmission operators, develops the European profile for cross-border and pan-European grid-model exchange — the <a href="https://www.entsoe.eu/data/cim/cim-for-grid-models-exchange/">grid model exchange standard</a> — with a <a href="https://www.entsoe.eu/data/cim/cim-conformity-and-interoperability/">conformity assessment scheme</a> that lets conformance be tested.</p><p>One layer deeper sits the equipment. When an operator automates a substation or controls a resource connected at medium voltage, a major reference is <a href="https://iec61850.dvl.iec.ch/">IEC 61850</a>, which standardises both the object models of the equipment — what a breaker or a protection relay reports — and the services used to communicate them, with an extension for advanced control of distributed energy resources. It is a major reference, not the only protocol in the field: the IEC publishes its own guidelines for gatewaying IEC 61850 information to <a href="https://webstore.iec.ch/en/publication/25511">IEC 60870–5–101/104</a> and to <a href="https://webstore.iec.ch/en/publication/26906">Modbus</a>. You may never touch it directly, but knowing it exists avoids reinventing those terms.</p><h3>The measured layer: meters and markets</h3><p>Between grid and household sits the meter, where consumption and generation are measured — the foundation for billing, settlement, shared-energy accounting and flexibility verification. One standard the blueprint points to here is <a href="https://www.dlms.com/">COSEM</a>, the meter-side object model normally used through the wider DLMS/COSEM suite maintained by the DLMS User Association. It standardises how a meter describes its registers and readings, while the protocol layer governs retrieval. Together they are why a smart meter from one vendor can, in principle, be read by another’s head-end system — but only in principle: two products can implement the standard and still differ in the options they choose. The blueprint lists COSEM among the standards its vocabulary infrastructure relies on, not as a settled national metering choice.</p><p>What is measured then has to be traded. Flexibility — shifting consumption or generation in time, in response to a grid signal — needs a common way to describe what is offered, for what period and price, and whether it was delivered. For the market-facing part, the blueprint points to <a href="https://www.entsoe.eu/publications/electronic-data-interchange-edi-library/">IEC 62325</a> and its European market profile, which structures the schedules, bids and market documents passed between operators, aggregators and platforms; it belongs to the same family as the grid model. It does not cover the whole flexibility chain, though: the blueprint names <a href="https://webstore.iec.ch/en/publication/26267">IEC 62746</a> — formally the interface between a customer energy management system and the power management system — for what it calls service provided to technical aggregator communication, and points back to the DER-control standards above for device-level control. Market document, aggregator coordination and device command are distinct interfaces.</p><h3>Behind the meter: devices and vehicles</h3><p>At the other end sits a crowded set of devices: solar inverters, battery storage, heat pumps, chargers, thermostats. Here proprietary models coexist with a handful of building-, home-automation and energy-management standards, and the result is the interoperability failure the previous piece described — a heat pump from one maker and a battery from another may share no model at all, even in the same house.</p><p>Closing that gap is what <a href="https://saref.etsi.org/">SAREF</a> was created for — the Smart Applications Reference ontology published by ETSI, a shared vocabulary for what a device is, does, measures and accepts as commands. It is deliberately generic and extended for particular domains; the energy-flexibility extension, <a href="https://saref.etsi.org/saref4ener/">SAREF4ENER</a>, specialises it for energy and flexibility-related concepts. Its value is not that it replaces a manufacturer’s internal model, or the protocol suites such as EEBUS or KNX that devices actually implement, but that it gives two systems the same words when they need to interoperate — so an aggregator and a household battery can describe what the battery can offer. It is a semantic layer to map to, not a protocol to adopt.</p><p>For the public charging point, the blueprint points to a dedicated protocol. <a href="https://openchargealliance.org/protocols/">OCPP</a>, from the Open Charge Alliance, defines the interface between a charging station and its central management system: status, sessions, authorisation, metering data, and in recent versions smart-charging profiles. It is only one part of the EV data chain: vehicle-to-charger communication, roaming between networks and grid-facing flexibility each rely on their own standards.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*sUQots2st5jcoaanGpjflw.png" /><figcaption>Reference standards mapped by interface, not stacked by layer. Notice where their territories meet and overlap — that is where mappings and profiles do the work.</figcaption></figure><h3>Connecting them without merging them</h3><p>Seen together, these standards do not stack into tidy layers. They cover different interfaces within the same physical system, and their scopes overlap — several share a family, and a device ontology and a charging protocol can describe the same equipment from different sides. What the blueprint offers is not a mandate but reference standards per interface, with a call to harmonise them. Aligning to the relevant ones reduces how much two parties must negotiate about what “energy consumed” means; it does not remove that work, since they still have to agree on profiles, versions, optional fields and mappings.</p><p>The pattern we draw from this — our reading, not a prescription in the blueprint — is federated: established models stay in use, and a thin layer of shared vocabulary, profiles and mappings makes them interoperable, rather than a grand unification that replaces what exists. The <a href="https://w3id.org/celine-eu">CELINE core ontology</a> is one project-level example — not the blueprint’s official vocabulary, but an implementation of the same idea. Published as a draft specification at a permanent, versioned address, it <a href="https://celine-eu.github.io/ontologies/index-en.html">defines its own lightweight classes</a> for the grid nodes and devices a community platform handles, and rather than leave them standing alone it imports a selected set of established vocabularies — the energy-community, sensor and smart-appliance ones — and ties its own classes back to those, and to the distribution profile of the grid model, through a mix of subclassing, direct reference and looser close-match links rather than by absorbing them. On top it adds what a community platform needs — from how a community and its members are structured to the datasets, KPIs and flexibility commitments it works with, and the connections back to the operator’s map.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*VMq9YULEyY1mWVO6C1hbwg.png" /><figcaption>A federated pattern: established standards stay in place at the edges. A community platform keeps a small set of its own lightweight classes in the middle, each mapped to the standard it borrows meaning from — anchoring to those standards, not replacing them.</figcaption></figure><h3>What this covers, and what it does not</h3><p>The <a href="https://www.celineproject.eu/">CELINE</a> ontology is a published draft specification, not a sector-wide standard: it connects the standards where a community platform needs them to meet, and leaves the rest to those that own that ground.</p><p>What matters is the principle. These models, ontologies and protocols were each built for their own interface and community of practice, and each aims at interoperability within its own domain. Integration across them was rarely the design goal: the int:net <a href="https://arxiv.org/abs/2309.08284">interoperability roadmap</a>, from the community behind the blueprint, still calls European smart-grid interoperability “heterogeneous and fragmented”. The blueprint is a recent European attempt to place these standards in a common frame and build a shared vocabulary infrastructure around them, harmonising rather than reinventing; its own argument for that is blunt — external systems cannot know how your data relates unless they are given a machine-readable expression of it. That does not mean the list is closed — the blueprint anticipates further convergence and new profiles. It means that when you choose how to describe a measurement, a device or a grid node, checking whether an agreed vocabulary already covers it comes first — the blueprint’s own formulation is that the data space “relies on the harmonization and usage of prominent standards-based data models and ontologies”. Our reading of that: reuse what exists, define profiles where it falls short, connect the pieces, and publish the connecting layer openly.</p><p><em>References for the standards and models discussed:</em></p><ul><li><em>Common European Energy Data Space blueprint: </em><a href="https://intnet.eu/images/Blueprint_CEEDS_v3.0.pdf"><em>CEEDS Blueprint v3.0</em></a></li><li><em>Common Information Model (CIM): </em><a href="https://www.entsoe.eu/data/cim/"><em>ENTSO-E CIM pages</em></a><em> and </em><a href="https://www.entsoe.eu/data/cim/cim-for-grid-models-exchange/"><em>CGMES grid model exchange</em></a></li><li><em>SAREF and SAREF4ENER: </em><a href="https://saref.etsi.org/"><em>ETSI SAREF portal</em></a><em> and </em><a href="https://saref.etsi.org/saref4ener/"><em>SAREF for Energy Flexibility</em></a></li><li><em>Metering (COSEM / DLMS): </em><a href="https://www.dlms.com/"><em>DLMS User Association</em></a></li><li><em>IEC 62325 market profile: </em><a href="https://www.entsoe.eu/publications/electronic-data-interchange-edi-library/"><em>ENTSO-E electronic data interchange library</em></a></li><li><em>IEC 61850 substation communication: </em><a href="https://iec61850.dvl.iec.ch/"><em>IEC 61850 knowledge base</em></a><em>, with the IEC’s mapping guidelines to </em><a href="https://webstore.iec.ch/en/publication/25511"><em>IEC 60870–5–101/104</em></a><em> and </em><a href="https://webstore.iec.ch/en/publication/26906"><em>Modbus</em></a></li><li><em>IEC 62746 customer-to-power-management interface: </em><a href="https://webstore.iec.ch/en/publication/26267"><em>IEC 62746–10–1</em></a></li><li><em>OCPP charging protocol: </em><a href="https://openchargealliance.org/protocols/"><em>Open Charge Alliance</em></a></li><li><em>Interoperability landscape: </em><a href="https://arxiv.org/abs/2309.08284"><em>Towards an Interoperability Roadmap for the Energy Transition</em></a><em> (int:net, ComForEn 2023)</em></li><li><em>CELINE core ontology: </em><a href="https://w3id.org/celine-eu"><em>w3id.org/celine-eu</em></a><em> and </em><a href="https://celine-eu.github.io/ontologies/index-en.html"><em>documentation</em></a></li><li><em>Open work: </em><a href="https://github.com/spindoxlabs/ds"><em>github.com/spindoxlabs/ds</em></a><em> and </em><a href="https://github.com/celine-eu"><em>github.com/celine-eu</em></a></li></ul><p><em>This article was drafted with AI assistance and reviewed, fact-checked and edited by its named author before publication.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=654eacf9afcc" width="1" height="1" alt=""><hr><p><a href="https://medium.com/spindox-labs/the-reference-standards-behind-europes-energy-data-space-654eacf9afcc">The reference standards behind Europe’s energy data space</a> was originally published in <a href="https://medium.com/spindox-labs">Spindox Labs</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Who controls access to the data your energy system runs on]]></title>
            <link>https://medium.com/spindox-labs/who-controls-access-to-the-data-your-energy-system-runs-on-44d4b4dbcb3c?source=rss----ab498e9e959d---4</link>
            <guid isPermaLink="false">https://medium.com/p/44d4b4dbcb3c</guid>
            <category><![CDATA[energy-transition]]></category>
            <category><![CDATA[data-privacy]]></category>
            <category><![CDATA[energy]]></category>
            <category><![CDATA[dataspace]]></category>
            <dc:creator><![CDATA[Luca Capra]]></dc:creator>
            <pubDate>Wed, 22 Jul 2026 13:16:58 GMT</pubDate>
            <atom:updated>2026-07-22T13:16:57.218Z</atom:updated>
            <cc:license>http://creativecommons.org/licenses/by/4.0/</cc:license>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*hsbAkUHDB8dFJovoa_zCXw.png" /><figcaption>The question is not whether energy data moves, but on whose terms.</figcaption></figure><p>If you operate part of a distribution grid, run or advise an energy community, or build services on top of energy data, you already live with the tension in that question. And if you are simply someone whose home is filling up with connected devices, it is your question too, because the answer shapes what everyone else in that list is allowed to do with the data those devices produce.</p><p>Every solar panel, heat pump, battery, smart meter, and charging point is producing a quiet, constant stream of data. How much a household generates at noon. How much it draws at seven in the evening. When a fleet of cars charges. When a local feeder is under strain. That data is the raw material for almost everything the energy transition needs next: coordinating flexibility across thousands of homes, spotting a problem on the grid before it spreads, letting a community see and settle what it actually shared. The volume and variety are increasing rapidly, creating governance and interoperability challenges for many organisations.</p><p>So here is the question that sounds simple and turns out to be the whole game: who gets to access it, who gets to use it, and on what terms? In practice, two unsatisfactory patterns have remained common, even though alternatives have always existed.</p><h3>The two bad options</h3><p>The first is that the data stays locked away. It sits inside a manufacturer’s app, a utility’s billing system, or a single vendor’s platform, in a format nobody else can read, cut off from the operator next door and often from the system in the next room. This is the world of silos. It feels safe, because nothing leaves. It is also the world in which data silos can prevent or delay an energy community from obtaining a complete and timely view of its performance, in which fragmented data reduces an operator’s ability to anticipate faults and coordinate preventive action, and in which no better service can be built on top, because the raw material is trapped.</p><p>The second option is that the data is handed over wholesale. Someone clicks “accept,” and a company takes a broad, permanent, hard to revoke licence to everything the devices produce. The data flows, which is useful, but the people whose devices generated it have stopped having any meaningful say in how it is used.</p><p>Neither pattern is good enough for the systems we are actually building, where power is shared between neighbours, where flexibility has to be coordinated across many parties at once, and where trust between operators, communities, and households is not a nice-to-have but the thing the whole arrangement rests on.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*skOVzBK852o77QJ8W1HIFw.png" /><figcaption>Three ways to arrange the same data. Only the third leaves the owner in control of what happens after the handover.</figcaption></figure><h3>A better model: shared, under defined rights</h3><p>There is a better model, and it now has both a policy framework behind it and working software to point to. The idea is to treat energy data as something that can be shared without being surrendered. Users, data holders, and data subjects retain defined rights over access and permitted use, and sharing happens under transparent, enforceable conditions.</p><p>The technical community calls the plumbing for this a “data space.” You do not need the term. What matters is the shape of it: rather than requiring all participants to surrender data to one dominant platform, a data space allows data to remain at source or be shared through centralised, federated, or distributed services, depending on the use case and governance model. Sharing becomes a deliberate, scoped act governed by clear rules rather than a one-time surrender.</p><p>Two design principles do the real work.</p><p>The first is sovereignty. The people and organisations that hold rights over the data have a say in what happens to it. Usage conditions can be expressed in machine-readable policies linked to the offering or agreement, supporting enforcement and audit where the technical and governance framework permits.</p><p>The second is transparency. Data offerings should be described through standardised metadata covering their content, provider, provenance, access conditions, and permitted uses. Clear metadata makes data easier to interpret and assess, while trust additionally depends on provenance, quality, identity, security, and accountability.</p><h3>Europe has set a clear direction</h3><p>European policy has established a clearer framework for access to and sharing of data, through two complementary moves.</p><p>The first move was legal. The <a href="https://digital-strategy.ec.europa.eu/en/policies/data-act">Data Act</a> gives users of connected products and related services stronger rights to access certain data generated through their use and, in defined circumstances, to have that data shared with third parties. The <a href="https://digital-strategy.ec.europa.eu/en/policies/data-governance-act">Data Governance Act</a> complements this by establishing mechanisms intended to increase trust in data intermediation and voluntary data sharing. Personal data remains subject to the GDPR, while electricity metering and consumption data are also governed by sector-specific legislation. Together, these instruments strengthen rights, obligations, and trusted mechanisms for data access and sharing — though they do not create a single property rule determining who “owns” all data.</p><p>The second move was practical. Rights you cannot exercise are not worth much, so Europe is also supporting the development of a shared framework that should let energy data move between countries, operators, and communities in a safe and standard way. This is the <a href="https://energy.ec.europa.eu/publications/common-european-energy-data-space_en">Common European Energy Data Space</a> — not one completed central platform, but an evolving framework and set of initiatives intended to improve interoperability and trusted data exchange. It is being developed through EU-supported initiatives involving public authorities, industry, research organisations, and other energy-sector stakeholders, with <a href="https://digital-strategy.ec.europa.eu/en/policies/data-spaces">design commitments</a> that participation should be based on transparent and non-discriminatory conditions, that the infrastructure is secure and privacy-preserving, and that access rules are fair. <a href="https://energy.ec.europa.eu/publications/implementing-regulation-interoperability-requirements-and-non-discriminatory-and-transparent_en">An implementing regulation adopted in 2023</a> establishes interoperability requirements and procedures for access to electricity metering and consumption data — an important sector-specific measure, though it does not cover every category of energy data. The emerging approach favours interoperability and federation between existing national, regional, and sectoral infrastructures rather than replacement by one central European system.</p><p>The law gives users stronger rights to access and share certain data, and the emerging infrastructure provides mechanisms through which those rights can be exercised securely. Rights without infrastructure are a frustration. Infrastructure without rights is a free-for-all. You need both, and both are now taking shape.</p><h3>Why this is the first thing to get right</h3><p>It is tempting to jump straight to the exciting parts of a smart energy system: the forecasting, the automation, the assistant that answers questions about the grid. All of that is coming, and all of it is worth building.</p><p>But none of it is trustworthy if the foundation is wrong. If the data underneath is either trapped in a silo or signed away wholesale, then every clever thing built on top inherits that flaw. Get access rights, governance, and accountability right first, and everything else has something solid to stand on.</p><h3>What this means, and what we owe you next</h3><p>Who controls access to the data your energy system runs on? There is no single answer based on ownership alone. EU law increasingly defines specific rights for connected-product users, data subjects, data holders, authorised third parties, and regulated energy actors. The Data Act strengthens users’ ability to access and share certain connected-product data. The GDPR protects individuals where energy data is personal. Electricity legislation establishes additional procedures for metering and consumption data. The Common European Energy Data Space aims to make these forms of sharing more interoperable and trustworthy in practice.</p><p>The remaining challenge is implementation. Policies, consent mechanisms, identity systems, metadata, contracts, access controls, and audit records must work together. Open implementations can make those claims testable, which is why our own work is public rather than described: an <a href="https://celine-eu.github.io/">open toolbox for energy communities</a> developed within the EU-funded CELINE project, and a <a href="https://github.com/spindoxlabs/ds">data space implementation</a> developed with reference to the blueprints from initiatives such as <a href="https://blueprint.dssc.eu/">DSSC</a> and <a href="https://intnet.eu/images/resources/Blueprint_CEEDS_v2.pdf">CEEDS</a>, with the current alignment assessment documented in the repository. It is early, there is no tagged release yet, and we would rather say that than imply otherwise.</p><p>The rest of this series takes the foundation apart piece by piece: what the European blueprints recommend you build, why systems still fail to understand each other, and how usage rules can be made to travel with the data instead of being left at the door. A companion piece hands you the shortlist directly — the open implementations a team can build on today, and the test we applied before putting anything on it. The work now is making the answer real, in the open, one running system at a time.</p><p><em>Policy sources: the EU </em><a href="https://digital-strategy.ec.europa.eu/en/policies/data-act"><em>Data Act</em></a><em> and </em><a href="https://digital-strategy.ec.europa.eu/en/policies/data-governance-act"><em>Data Governance Act</em></a><em>; the </em><a href="https://energy.ec.europa.eu/publications/common-european-energy-data-space_en"><em>Common European Energy Data Space</em></a><em> and the wider </em><a href="https://digital-strategy.ec.europa.eu/en/policies/data-spaces"><em>Common European data spaces</em></a><em> initiative, both rooted in the </em><a href="https://digital-strategy.ec.europa.eu/en/policies/strategy-data"><em>European strategy for data</em></a><em>.</em></p><p><em>Open work referenced: the </em><a href="https://github.com/celine-eu"><em>CELINE toolbox</em></a><em> for energy communities, with </em><a href="https://celine-eu.github.io/"><em>documentation</em></a><em>, and the </em><a href="https://github.com/spindoxlabs/ds"><em>data space implementation</em></a><em>.</em></p><p><em>This article was drafted with AI assistance and reviewed, fact-checked and edited by its named author before publication.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=44d4b4dbcb3c" width="1" height="1" alt=""><hr><p><a href="https://medium.com/spindox-labs/who-controls-access-to-the-data-your-energy-system-runs-on-44d4b4dbcb3c">Who controls access to the data your energy system runs on</a> was originally published in <a href="https://medium.com/spindox-labs">Spindox Labs</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
    </channel>
</rss>