21h •21 hours ago • Visible to anyone on or off LinkedIn
Fascinating, and not something I was familiar with. My suspicion is that nRQL didn't succeed primarily because the physical infrastructure was just not ready for it. In many respects what you're seeing now with SHACL is that there is both enough infrastructure and sufficient use cases, especially in the AI space, that was once just not feasible has crossed over that hump. This has been the case with a number of W3C technologies - SVG, for instance.
🔍 SHACL (2017) gave RDF a standard way to validate knowledge graphs. But how much of it was conceptually new? I put that question to Claude (Anthropic's Opus 4.8) and had it run the whole investigation. Working from the original RacerPro manuals and the W3C SHACL specifications, Claude compared SHACL's expressivity against RacerPro's nRQL query language, then coded the entire W3C SHACL test suite, case by case, to keep the comparison honest. What it found: most of the OWL-centric, closed-world checks SHACL is used for were already expressible in nRQL back in 2004–2005: required properties, cardinalities, qualified values, datatype checks, even ontology-metadata constraints. The mechanism was negation-as-failure + projection: a negated projected subquery is exactly a NOT EXISTS / anti-join, which nRQL had years before SPARQL 1.1 (2013) standardized FILTER NOT EXISTS. By the numbers (Claude's coding of the test suite): ~73–84% of SHACL Core constraint tests have a direct nRQL counterpart. The one provably-native residual is only ~5%. And it's entirely unbounded regular property paths, which is a theorem rather than a hunch (transitive closure isn't first-order definable). So what WAS genuinely new in SHACL? A first-class (recursive) shape conformance semantics, native regular paths, RDF-term-level validation, and a standardized validation-report model, but not the closed-world constraint idea itself for which it is frequently cited. And Claude kept it calibrated: this is prior art and convergence, not documented influence. No claim that SHACL or SPARQL borrowed from nRQL. Just another interesting use case for Agentic AI - automatic, grounded, in-depth comparison of highly complex technical specifications. Full write-up + the reproducible coding: 👉https://lnkd.in/gC2uSzxa hashtag#SHACLhashtag#SemanticWebhashtag#KnowledgeGraphshashtag#RDFhashtag#SPARQLhashtag#DescriptionLogicshashtag#AI
Kurt Caglethanks. Well, yes, it's hard to tell. On one hand, nRQL + RacerPro - and the Racer Systems company behind it! - were *very* successful. It depends on your "success metrics". We granted hundreds of free academic research licenses. Many applications and citations of Racer. Dozens of PhD students (some professors by now) that used the system as a backbone for their research. Our commercial partner (Franz Inc.) is still a big player in the field. So, some successes - just not monetarily or historical. In the long run, the system will be forgotten. As always - the winners take it all, and rewrite history. I am only mildly bitter about this. But when I read here how revolutionary some of these supposed new technologies are, I sometime get the urge to write a few lines... In the end, I don't think it was the infrastructure - but academic network effects and not-invented-here bias. Our competitors rather copied and re-implemented our ideas than giving us credit. Personally, I didn't make enough money from the Racer business to afford a used car back in 2008, but so be it - it was fun anyway, and I am grateful for the experience and having worked with such brilliant and fun people (Ralf Möller, Volker Haarslev, Kay Hidde).
A few people asked: nice spec comparison — but does any of it still run? So I had Claude (Anthropic's Opus 4.8) go further: it downloaded the actual RacerPro 2.0 binary — the 2013 build we shipped at Racer Systems — started the server, and ran everything live. Twenty years on, it still works: it classifies OWL ontologies, answers nRQL queries, and reproduces every example exactly. • It reproduced the data-substrate example live — graph edges carrying their own key-value properties, queried directly. RacerPro had property graphs, fused with a DL reasoner, in 2005, before the term existed. • It wrote "What SHACL still can't do, 20 years on": classification, consistency, rules — and the exotic one, reasoning about its own queries (SHACL's versions are undecidable problems today). • It caught itself, disproving one of its own earlier claims — then wrote an nRQL tutorial + one-page cheat sheet, every example checked live. Agentic AI that runs the artifact and self-corrects against ground truth. And, honestly — it was good to watch the old thing run. 👉https://lnkd.in/gC2uSzxa hashtag#KnowledgeGraphshashtag#DescriptionLogicshashtag#SHACLhashtag#PropertyGraphshashtag#AgenticAI
![View Michael Wessel’s graphic link](data:image/svg+xml,<svg xmlns="http://www.w3.org/2000/svg" width="100" height="100"><rect fill-opacity="0"/></svg>){kind=avatar}
Michael Wessel • 2nd
But when I read here how revolutionary some of these supposed new technologies are, I sometime get the urge to write a few lines... In the end, I don't think it was the infrastructure - but academic network effects and not-invented-here bias. Our competitors rather copied and re-implemented our ideas than giving us credit. Personally, I didn't make enough money from the Racer business to afford a used car back in 2008, but so be it - it was fun anyway, and I am grateful for the experience and having worked with such brilliant and fun people (Ralf Möller, Volker Haarslev, Kay Hidde).
Michael Wessel • 2nd
So I had Claude (Anthropic's Opus 4.8) go further: it downloaded the actual RacerPro 2.0 binary — the 2013 build we shipped at Racer Systems — started the server, and ran everything live. Twenty years on, it still works: it classifies OWL ontologies, answers nRQL queries, and reproduces every example exactly.
• It reproduced the data-substrate example live — graph edges carrying their own key-value properties, queried directly. RacerPro had property graphs, fused with a DL reasoner, in 2005, before the term existed.
• It wrote "What SHACL still can't do, 20 years on": classification, consistency, rules — and the exotic one, reasoning about its own queries (SHACL's versions are undecidable problems today).
• It caught itself, disproving one of its own earlier claims — then wrote an nRQL tutorial + one-page cheat sheet, every example checked live.
Agentic AI that runs the artifact and self-corrects against ground truth. And, honestly — it was good to watch the old thing run.
👉 https://lnkd.in/gC2uSzxa
hashtag#KnowledgeGraphs hashtag#DescriptionLogics hashtag#SHACL hashtag#PropertyGraphs hashtag#AgenticAI
GitHub - lambdamikel/shacl-nrql-comparison: Two historical-semantic investigations: (1) what was conceptually new in SHACL vs RacerPro nRQL + the data/mirror substrate (~2005); (2) the property-graph lineage and RacerPro's early reasoning-coupled property graph.
Two historical-semantic investigations: (1) what was conceptually new in SHACL vs RacerPro nRQL + the data/mirror substrate (~2005); (2) the property-graph lineage and RacerPro's early reasonin...