The identifiers underneath American academic standards are all dead, and the redirect still works

What ASN was, and why its death is not just a dead website

The Achievement Standards Network was the linked-data identifier layer for American academic standards. It minted a persistent, dereferenceable URI for every standards document and every individual standard statement, and it was the scheme the Learning Resource Metadata Initiative was designed around. When LRMI defined educationalAlignment, and when that pattern went into schema.org and from there into the metadata of open educational resources everywhere, the thing on the other end of the alignment was an ASN URI.

That means ASN identifiers are not confined to one dead service. They are sitting inside learning-object records, repository indexes, publisher catalogues and alignment tables built over fifteen years. A search of public code alone finds 44,084 distinct ASN identifiers across 52 repositories, cited 57,769 times: the Learning Registry, inBloom's tagging application, NCAR's digital library stack, the T3 Innovation Network's schema mapper, the CASS competency system, and DCMI's own worked examples of how to do this correctly.

The sharpest instance is 1EdTech. The organisation that publishes CASE, the specification the sector migrated to precisely because ASN was fading, ships a QTI version 3 example package whose curriculum-standard references are ASN URIs. All nine of them are dead. That is not a criticism of 1EdTech specifically; it is evidence of how deep the dependency runs, that even the successor's reference material still points at the predecessor.

We measured three populations. Identifiers still cited in public code: all 44,084 dereferenced, a complete census. Standards-document identifiers carried by the live standards corpus: all 3,057, a complete census. Statement identifiers in that corpus: a random sample of 20,000 drawn from 770,861, which at zero successes puts the 95 per cent upper bound on the resolving proportion below 0.02 per cent. A full census of the third population would have taken about eleven hours at a polite request rate, and sampling is stated rather than hidden.

The result in all three: nothing resolves. Not degraded, not slow, not partially migrated. The canonical form http://purl.org/ASN/resources/{id} redirects faithfully to the origin host and the origin host returns 404. Meanwhile http://purl.org/ASN/schema/core/ returns 200 with valid RDF/XML, served from an Amazon S3 bucket, and the education-level scheme still resolves too. Someone kept the vocabulary alive. Nobody kept the data.

This is the failure mode that persistent identifier schemes are supposed to prevent, arriving through the mechanism that was supposed to prevent it. A PURL is an indirection layer whose whole purpose is to outlive the hosting arrangement behind it. Here the indirection layer works perfectly and forwards every request to a 404, which is worse than an outage, because an outage is visible and this is not.

One standard, all the way down

Take AP Microeconomics learning objective POL-5.B, "Explain sources of income and wealth inequality." It is current, published by the College Board in the 2022 course and exam description, and in force. In the standards corpus it carries three names: a Common Standards Project identifier, the human coding scheme POL-5.B that a teacher would recognise, and the ASN identifier S21370147.

Only one of those is a globally unique, dereferenceable name, and it returns 404. The human coding scheme is not unique outside its own document, so any system joining two frameworks on it is joining on a label. The record also carries a rights holder, D2L Corporation, and a licence title of null, so the statement cannot be lawfully redistributed by anyone relying on the published metadata. Nothing in the source data flags any of this. The identifier is simply printed, and everything downstream assumes.

What the surviving standards can actually reject

An ontology earns its keep by ruling things out. If nothing you can write is inconsistent with it, it is a glossary. So we asked a mechanical question of each published artefact: given a specific wrong statement, does it notice?

Six mis-statements, each translated into the artefact's own vocabulary, each counted as detected only when an axiom actually present in that artefact is violated. Where an artefact has no term for what the mutation says, the result is recorded as inexpressible rather than as undetected, because an artefact cannot be blamed for failing to reject a sentence it cannot form. The Learning Standards Ontology detects six of six. The ASN schema detects one. The CEDS Ontology v14 detects none, and five of the six are inexpressible in it.

Underneath that sits an axiom census. CEDS v14 is 243,601 triples and carries 24 axioms of the kinds that let data contradict a model: domains, ranges, disjointness, cardinality, functionality. Six of those 24 are ranges on CEDS's own annotation properties such as textFormat and maxLength, and the other 18 are universal restrictions. Across all 2,336 of its domain properties it declares zero rdfs:domain, zero disjointness, and zero cardinality constraints. The ASN schema is 465 triples, has been unmaintained for years, and carries 121.

Two further results from the same parse. CEDS v14 declares zero owl:ObjectProperty, while the repository README states that the ontology provides "definitions and meaning about those relationships through Object Properties". The relationships are there, as 2,336 bare rdf:Property terms annotated with schema.org's domainIncludes and rangeIncludes, which schema.org defines as indicative rather than constraining. They are documentation, not logic. And 965 CEDS terms are typed as both an owl:Class and a skos:ConceptScheme, conflating a set of individuals with a container of concepts, while its 19,546 concepts carry zero skos:broader relations, so genuinely hierarchical vocabularies inside it, including 1,753 SCED course codes and 7,916 ISO 639-3 language codes, ship as flat lists.

None of this is a verdict on the people who built these things. CEDS calls its ontology a draft, maintains it actively, licenses it Apache-2.0, and runs a SHACL and JSON-LD workstream with NCES precisely because the shipped artefact does not constrain. The measurement says how large the gap currently is, which is the useful thing to know if you are about to validate a pipeline against it.

One correction, stated because it moved a number against us. Our first checker read multiple rdfs:range declarations conjunctively, which is what RDFS actually says. Under that reading ASN appeared to detect three mutations. It was not detecting them: asn:isChildOf declares two ranges plainly meaning "either", and the strict reading makes every ordinary ASN parent link violate its own schema. We switched to the charitable disjunctive reading, ASN's score fell from three to one, and the strict-reading defect is now reported separately as ten ASN properties declaring more than one domain or range. The lower number is the honest one.

The corpus fails in both directions at once

Identifier systems fail in two opposite ways, and this corpus does both simultaneously.

One name, several things. 433 ASN identifiers are attached to statements whose text is materially different. The clearest case we verified by hand against the live API: S100EC5D is attached to five statement records inside one document, Arizona's Academic Content Standards for Science. Four of them read "Design and conduct controlled investigations." The fifth reads "State that respiration involves the action of enzymes in cells."

One thing, many names. "Demonstrate command of the conventions of standard English capitalization, punctuation, and spelling" appears verbatim under 718 distinct ASN identifiers. Nothing in the published data says those are the same expectation. A system asked whether two states teach the same thing has to decide by comparing strings, which is exactly the problem the identifier layer existed to solve.

And the alignment layer that should adjudicate is mostly not alignment. In the live Georgia CASE package we examined, 2,453 of 2,483 associations, 98.8 per cent, are isChildOf: document structure, a table of contents. There are 30 isRelatedTo and zero cross-framework associations of any kind. isRelatedTo is itself worth naming, because no observation counts against it. It is the most common cross-framework predicate in practice and it is unfalsifiable, and a claim that cannot fail cannot inform a decision.

Ten findings, and what each one costs

All figures computed on 16 August 2026. Only the CEDS measurement is pinned and byte-for-byte reproducible, from the V14.0.0.0 tag. The standards corpus, the CASE server and the citation population are live systems, so a later run gives different totals while the method reproduces exactly.

What the ontology does about it

Resolvability is recorded data, not an assumption. An identifier assertion carries its scheme, the source that asserted it, and a dated resolution observation holding the HTTP status, the media type and the final URL after redirects. That last field is what distinguishes a live indirection layer over dead data from a healthy scheme, and it is the distinction that made this whole study possible. No incumbent artefact in this space has anywhere to put that fact, which is why the failure went unmeasured for years.

Scope and custody are declared. The SKOS registry records for each of seven schemes whether it is document-, statement- or jurisdiction-scoped, whether it is globally unique, whether it promises dereferenceability, and whether anyone is currently honouring that promise. ASN is recorded as dereferenceable and orphaned, and the gap between those two is the subject of this page. The human coding scheme is recorded as not globally unique and minted per document, which is the fact that stops a team joining two frameworks on it.

Alignments record what would refute them. An alignment assertion carries a falsifiability grade and may carry explicit refutation criteria; one with neither is classified automatically as an unfalsifiable alignment. An exact match asserted between a statement and its own ancestor is a contradiction the shapes reject outright, because containment and equality cannot both hold. That is a rule no incumbent standard here can express, which is why injected mis-mappings survive validation against them.

The test suite enforces on this ontology the properties whose absence it measures in others: every property declares exactly one domain and one range, the ontology declares disjointness, and no term is typed as both a class and a concept scheme. A pull request breaking any of those fails CI, deliberately. It would be embarrassing to publish this study and repeat the defects.

The identifier discipline here was first proved on the United States fund universe in our Investment Fund Ontology and then on European insurance in the Insurance Register Ontology. Education is its third instantiation, and the transfer without redesign is itself the finding: regulated register fabrics fail the same way whatever they name.

Why this matters if you work in education data

Content alignment claims. Publishers and assessment providers sell alignment to standards. If the identifier those claims resolve against is dead, the claim is unverifiable by anyone outside the company that made it, and internal drift is undetectable. That is a commercial exposure, not just a metadata problem.

Cross-state and cross-framework comparison. Any product that promises "this works for your state's standards too" is doing framework-to-framework mapping. With 433 colliding identifiers, 718 names for one common expectation, and 98.8 per cent of the available associations being table-of-contents structure, that mapping is being done by string similarity whether or not anyone says so out loud.

Migration off a dead scheme. Moving from ASN to CASE or CTDL-ASN means minting new identifiers. Unless the old identifier is carried forward explicitly, every historical alignment silently loses its anchor. The measurement to run before that migration is simple: how many of your stored identifiers still resolve, and what will you record for the ones that do not.

Grounding for automated systems. Every serious education data programme in 2026 wants language models and agents reading its curriculum graph. A model grounded in a graph where one identifier names both a statement about controlled investigations and a statement about cellular respiration will answer confidently and wrongly, and the confidence is the dangerous part. Validation is not a step before the interesting work; it is what makes the interesting work safe to ship.

Hiring signals. The sector has noticed. HMH is currently recruiting an Ontology Engineer for the K-12 domain, remote, at $100,000 to $115,000, asking for RDF, OWL, SHACL, SPARQL and experience with foundational ontologies including BFO, DOLCE and the Common Core Ontologies. That is a serious specification, and it exists because the artefacts this page measures do not yet do the job.

Common questions

Open, reproducible, and free to use

The ontology, the identifier-scheme registry, the three-layer SHACL suite, the citation harvester, the resolution census, the corpus harvester, the graph builder and the query library are public. Ontology and documentation under CC BY 4.0, code under MIT. The build report lists what could not be obtained as carefully as what could, including the corrections we made against ourselves during the build, because a study that only reports its successes is marketing. Harvested source data is deliberately not redistributed: some of it states no licence, which is one of the findings.

Related work: our machine-validated ontologies study measures when an ontology can reject a wrong statement, the Skills England occupational maps apply the same discipline to the UK skills system, and the industrial ontology crosswalks measure how often standards crosswalks can detect a wrong mapping at all.