True but Irrelevant

The drawing was the talk.

The room in Thiruvananthapuram knew it was getting a systems talk. Rob had been with Elsevier since 1991. He had been a mathematician. He had been the Chief Content Architect. He had the résumé of someone who could have spoken comfortably about DTDs, XML pipelines, and enterprise content strategy for an hour without once mentioning a child with colored pencils.

Instead he told a story about his granddaughter Emily, who at age six told her teacher that she had seven grandmothers. The teacher said that was impossible. Emily did what every good data engineer hopes users will do: she brought evidence. Her drawing had numbered grandmothers, a family tree, and a policy choice hidden inside it. Great-grandmothers counted too.

"I have seven grandmothers." Rob's granddaughter, as retold from the transcript

The joke is that the child's drawing was not childish at all. It had a definition. It had a registry of entities. It had a unique identifier for each Oma. It even had a piece of defensive programming: when she did not know how to spell Hetty, she fell back to the generic label "Oma" instead of risking a bad fact.

That is the first large idea in the talk: data quality is not only about correctness. It is about provenance, definitions, and the courage to leave uncertainty visible when you do not know enough to tidy it away.

Elsevier's graph is bigger than a file.

Rob's next move was to zoom out from the drawing to the system he actually works on. Elsevier's knowledge graph is not a neat database table. It is a constantly re-argued map of research objects: 23.8 million full-text articles from ScienceDirect, 102.3 million Scopus records, author profiles, organization profiles, patents, standards, grant awards, and more.

The language he used was almost understated, which made it stronger: the platform harvests evidence, extracts occurrences, matches and clusters entities, and then forms a context view for each product. It is a machine for turning many partial truths into something usable without pretending the truth was ever simple.

"All these things are heavily probabilistic." The hidden philosophy of the whole pipeline

From markup to annotation.

Rob's central pivot was not a new file format. It was a change in posture. He said that the team had moved from markup to annotation. That means the document stays the document, while the claims about the document live beside it.

This is where the standards trail becomes important. TEI's standOff container explicitly supports linked data, contextual information, and stand-off annotations. The W3C Web Annotation Data Model does the same job on the web: a body, a target, and a relationship that can be shared across platforms. TEI standOff note W3C Web Annotation

Rob's line about wanting to make assertions about the LaTeX document rather than the XML document lands because it is funny and serious at the same time. He is not anti-XML. He is anti-collapsing everything into one fragile layer.

"Annotations are even more tagging than tagging." His sticky-note analogy, sharpened into a design principle

<!-- The source file carries the claim inside it -->
<article xml:id="art-17">
  <title>True but Irrelevant</title>
  <p>
    The copy editor changed <italic>page 38</italic>
    to fit the house style.
  </p>
</article>

<!-- The claim lives outside the source and points back to it -->
<standOff>
  <annotation xml:id="ann-38"
              target="#art-17"
              resp="#copy-editor"
              cert="0.98">
    <note>page 38</note>
  </annotation>
</standOff>

The difference between the two examples is almost comic until you realize how much institutional pain hides inside it. Inline markup is tidy until two people need to disagree about the same line. Stand-off annotation looks messier, but it lets one person say "author", another say "copy editor", and a third preserve the original text without flattening the argument.

The dog-food confession matters because it makes the talk self-inverting. The company that spent years teaching the world about XML discipline eventually found some of its own documentation easier to manage in Markdown. Not because Markdown is better for everything, but because real workflows punish overfit standards. The phrase on the slide was short: hard to write, hard to display.

Validate on departure, ignore on arrival.

This was Rob at his most operational. The producer does the tests. The receiver picks what it needs. The recipient does not get to pretend that incoming data is perfect, only to treat the source as if it were already fixed.

"Schema validation alone never has been sufficient." The reason the system needs provenance as much as it needs syntax

If that sounds obvious, it is only obvious after decades of broken content. The interesting point is the asymmetry: data should be validated before it leaves the producer, not "cleaned up" by a downstream system that no longer knows what changed.

Model the data, not the world.

Slide 9 was one of the most elegant in the deck because it reduced a large philosophy to three small corrections:

That is why Rob kept insisting on a common vocabulary and on the ability to represent absence. It is also why he kept returning to time, because content that changes without a record of change ceases to be content and becomes rumor.

Articles are not alone.

The platform view becomes more interesting once you realize that content changes outside the article itself: copyright notices move on separate workflows, authors revise, copy editors intervene, and the corpus keeps growing. Rob's answer was to optimize for revision resiliency, reproducible identifiers, and corpus-wide consistency rather than fetishizing a single pristine file.

"The whole database matters." Rob's reminder that the unit of value is the corpus, not the page

That is also where the phrase "precision and recall" stops being academic jargon. When the thing you are building has 100 million records, small error rates are no longer small. They are a business model, a trust problem, and a UX problem.

What comes after XML?

Rob did not answer that question with a moonshot. He answered it with a standard: Content Profiles / Linked Data (CP/LD). The important thing about CP/LD is not that it sounds modern. It is that it makes the separation he keeps talking about concrete: content, semantics, and display can be linked without collapsing into one giant schema. NISO CP/LD

That conclusion lines up with the standards literature. The RDF 1.1 Concepts and Abstract Syntax document defines RDF as a graph data model. JSON-LD 1.1 gives you a JSON-friendly way to serialize linked data. SHACL gives you a way to validate shapes. That stack is very close to Rob's "clear vocabulary" plus "language for validation" wish list. RDF / JSON-LD / SHACL

Provenance matters too. W3C's PROV-DM explicitly defines provenance as information about entities, activities, and people involved in producing a thing, which is almost exactly the intuition Rob was applying when he refused to let a downstream system silently overwrite source data. W3C PROV

The closest thing to a counterweight is not a contradiction, but a caution. NISO's own CP/LD material says the standard complements existing models instead of replacing them. That is the adult version of Rob's argument: the future is not a clean break from text; it is a layered system where the source stays visible and the claims about it become portable.

"The text must be extremely high quality." His answer to the AI question, and the reason corpus curation still matters

The audience's question about AI made the point sharper. If large language models are going to sit in front of the scholarly corpus, then the original text becomes more valuable, not less. Rob's second answer was almost managerial, but it was also strategic: the knowledge graphs are what let Elsevier do something distinctive with those models. The graph is not an ornament. It is the memory.

His multilingual answer followed the same pattern. He did not promise magical universal formatting. He said the problem is almost impossible if you try to generate text from tags for every permutation, and that the better move is to separate content from annotations so the textual form can stay textual.

My inference from those sources is simple: Rob is right that the next layer should preserve the source and attach meaning around it, but the broader standards ecosystem also suggests that annotation works best as a companion layer, not a replacement fantasy. That is the useful middle ground.