What this graph is, and what it is not

It is a network of asserted relations between named things, each relation carrying the paragraph it came from and a quote from it. It is not a subsumption ontology, and most of what an ontology engineer would look for is missing on purpose. This page says what the design is, what it leaves out and why, It then runs the standard pitfall scan over the graph and reports what it finds, including the findings that are these decisions showing up as defects.

The address keeps the word ontology because the page runs an ontology pitfall scan; the graph itself is not one, and the paper says so.

Two tiers

The first tier is the relation as Jung's sentence phrased it. It is frozen: the string is part of the claim that was gated, and a model of a different family judged the paragraph against exactly that wording. Rewriting it into a controlled predicate would change what was verified.

The second tier is the family: a classification over the verified relation, drawn from a fixed vocabulary in a versioned document, assigned when two models reading that document independently name the same family, and ruled by a person when they do not. The second tier never reopens the first. Nothing on this site calls a family verified, and no family assignment changes a relation string.

The two words verified and assigned are deliberately kept apart. Different evidence, different word.

These are our families. They record how this project sorted relations Jung wrote. Not a structure discovered in his thought, and not a position the field has agreed on. The graph is a map of assertions rather than a reading, and that holds harder in the second tier than anywhere else on the site.

Where a relation carried a whole clause, a normalization layer records a cleaned verb phrase beside the frozen original, with the qualifiers pulled out into their own fields. 260 relations carry that layer. The family layer classifies the normalized form where it exists, which is why the map counts effective strings rather than surface ones.

The declarations

Two properties are declared, and only two: which relation strings are inverses of each other, and which are symmetric. Each declaration is transcribed from a line of the definitions document, cites that line, and is applied only when a reader asks for closure. The derived row is labelled, carries its source edge's references, and is never written into the graph.

7 active. 6 further pairs are recorded as candidates and are inert: the document has not ruled on them, and a declaration without a document line to cite would be this build's opinion rather than the ruleset's.

rulings behind them

Where the checks come from

The scan follows the OOPS! catalogue, the OntOlogy Pitfall Scanner, which numbers the modelling faults that recur across published ontologies and describes how each can be detected. A pitfall catalogue is not a standard to comply with; it is a list of the places ontologies are known to go wrong, so that an evaluation can be systematic instead of a matter of taste. tools/ontology_qa/quality.py implements a subset of it locally against this graph, and keeps the catalogue's numbering so a finding can be looked up in the literature. It is not the OOPS! service and does not run its full test set.

Poveda-Villalón, Gómez-Pérez and Suárez-Figueroa, "OOPS! (OntOlogy Pitfall Scanner!): An On-line Tool for Ontology Evaluation", International Journal on Semantic Web and Information Systems 10(2), 2014.

The catalogue was written for ontologies built out of classes and properties; this graph is built out of relations between named things. Where a number is borrowed for a fault of a different shape, the third column says so. A borrowed number presented as the catalogue's own finding would claim more agreement with the literature than there is. The standing frame holds over all of it: a finding maps to the catalogue where one applies, and several of the findings below are decisions recorded above rather than defects.

The pitfall scan

Run at every build by tools/ontology_qa/quality.py, which reads the graph and writes nothing. The four figures below are its own; the table under them is every finding it reported, with this project's stance beside each one. 7 of the 19 findings are decisions recorded above, 8 are answered by the declarations, and 4 are open defects carried in the record.

A stance is not a rebuttal of the finding. An open vocabulary is hard to query by predicate, which is why the family layer exists. That was a conscious design decision.

Checked against readers of Jung

In August 2026 we put a sample of these assignments to the secondary literature: 27 items, drawn by seeded stratification and weighted toward the contested families, checked in three pre-registered passes against eight works of Jungian scholarship: Jacobi's The Psychology of C. G. Jung, Samuels, Shorter and Plaut's A Critical Dictionary of Jungian Analysis, von Franz's Alchemy: An Introduction to the Symbolism and the Psychology, four volumes of Edinger, and Abraham's A Dictionary of Alchemical Imagery. The question was not whether Jung's claim is true; the verification chain had already asked that. It was whether the kind of relation we recorded is the kind a reader of the literature would say Jung asserts.

The result, stated exactly: the literature treats 16 of the 22 sampled assignments (the other 5 of the 27 items were queued arm-disagreements, scored separately); scholars' renderings agreed with our family on 8 of those 16, disagreed on 6, and split among themselves on 2. The other 6 pairs no consulted work treats at all. Jung's image-level assertions are often simply not discussed anywhere. The disagreements were not absorbed: they produced seven rule changes, folded into the ruleset as v0.7, the largest of which sends every amplified-by edge back for reclassification because two independent reference authors read those relations as content claims, not adduced parallels.

Two limits. Agreement here is not endorsement: no scholar reviewed this vocabulary, and where the sample agreed it mostly agreed on the everyday families. And the check also measured the literature itself: the same reference author writes "X is Mercurius" and "X is a symbol of Mercurius" for the same pair, sometimes on one page. On the hardest boundary this layer draws, the field the check consulted does not hold a stable line either. The full record (sample, per-item worksheets, both review passes, the disagreements in full) is committed at docs/experiments/external_validity_2026-08/ in the repository.

Deliberately absent

7 of the findings above are these decisions, showing up as defects because the catalogue has no way to tell a choice from an oversight.

• Transitivity. No family is transitive. Two terms joined through a third are joined through a third; the graph does not claim they are the same. The depth-bounded identity walk that reads such chains one edge at a time is a command-line tool in this release (tools/query/kg_query.py --walk) and is not exposed on the site.
• Domain and range as constraints. The signatures a relation is observed with are a measurement of the corpus. As a constraint they would reject an edge for disagreeing with the schema rather than with the paragraph.
• Disjoint types. A node can work in two registers at once, and a disjointness axiom would force a choice the material does not make.
• A subsumption backbone. There is no hierarchy to inherit through, because none was claimed. The type tag is a coarse label, not a class system.
• A closed relation vocabulary. The predicate is Jung's phrasing. Closing the set would mean editing claims, so the set stays open and the family layer carries the classification.