narrativegoldmine: a Logseq graph that is also an OWL ontology

This repository holds a 7,874-class OWL ontology whose source of truth is 7,874 ordinary Logseq markdown pages, the Python pipeline that compiles them into RDF and a binary graph format, and the WebAssembly explorer that renders the result. It is a self-contained release: corpus, build, viewer and method in one tree, published at narrativegoldmine.com.

What this corpus is. Mostly AI-generated synthetic content, produced under human direction, by design. It is an ontology testbed (built to exercise a medium-scale Logseq→OWL pipeline and a GPU graph renderer), not an authoritative encyclopaedia. The provenance it carries (did:nostr attribution, prov:generatedAtTime, stable URNs) attests traceable generation under human direction, not human authorship. That framing is emitted by the pipeline itself and is asserted in dist/data/graph/stats.json under corpus.nature: "synthetic".

Numbers

Measured on the build of 2026-07-25 (pipelineVersion: ng-1.0.0), reproducible with the quickstart below.

Source pages	7,874 markdown files in ontology/pages/ (129 MB)
OWL classes	7,874 (0 individuals)
RDF triples	258,200 in Turtle (12.1 MB)
Declared edges	111,827 (9,481 subClassOf + 102,346 object-property)
Resolvable graph edges	98,776 after self-loop, duplicate and unresolved-target removal
Domains / categories	6 / 34
Multiple inheritance	1,401 classes with more than one parent; 454 spanning categories, 153 spanning domains
Validation	0 errors, 0 warnings, 1,401 info
Build time	~18 s, single-threaded, rdflib only

Classes per domain, before the 1,500-node tier cap (nodes + nodesTruncated for each domain-* scope in stats.json): Infrastructure 2,585 · Artificial Intelligence 1,899 · Blockchain 1,432 · Spatial Computing 1,216 · Robotics 600 · Distributed Collaboration 142. Only the first two exceed the cap.

The taxonomy is a lattice, not a tree, and that is deliberate. 1,401 of the 7,874 classes declare more than one subClassOf; 454 of them thereby sit in more than one of the 34 categories and 153 in more than one of the 6 domains. A Star Algorithm is under Search Algorithm, Informed Search and Graph Search at once. Multiple inheritance is legal in OWL 2 EL and here it carries meaning: it is how a concept that genuinely belongs to two branches is bridged rather than duplicated. The pipeline used to report each such class as a MULTI_PARENT validation warning: 957 of the 961 warnings the previous (2026-07-23) build published were this one code, the other four being INVALID_DOMAIN. That published a design property as a defect list. MULTI_PARENT is now info, and the overlap is published as data instead: dist/data/graph/bridges.json names all 542 classes that cross a category or domain boundary (65 cross both) with their full membership, and overview.json carries 90 weighted category-to-category bridge edges alongside the 34 backbone edges.

What makes it interesting: the hybrid page

Every page in ontology/pages/ is simultaneously two things, and neither degrades the other. Open one in Logseq and it behaves like any other note: outline bullets, [[wikilinks]], property:: values, aliases. Run the pipeline over the same file and it yields a formally typed OWL class with a definition, a taxonomy position and twelve kinds of typed relation.

The trick is that the machine layer lives in fenced ```json-ld blocks that Logseq renders as inert code, and the human layer lives in the outline below them, which the pipeline never reads. There is no lossy round-trip because there is no round-trip: each layer is authoritative for its own audience.

From ontology/pages/A Star Algorithm.md, the first block identifies the page, the second the class. Abridged: the real file carries 41 wikilinks, three parents, eight relation groups and a full-paragraph definition.

public:: true

# A Star Algorithm

```json-ld
{
  "@context": "https://narrativegoldmine.com/ns/v1",
  "@id": "urn:visionflow:page:a-star-algorithm",
  "@type": "Page",
  "vc:slug": "a-star-algorithm",
  "vc:public": true,
  "vc:schemaVersion": 2,
  "vc:outboundWikilinks": [
    {"@id": "urn:visionflow:linked:search-algorithm", "vc:label": "Search Algorithm"}
  ]
}
```

```json-ld
{
  "@context": "https://narrativegoldmine.com/ns/v2.jsonld",
  "@id": "urn:ngm:class:a-star-algorithm",
  "@type": "Class",
  "label": "A Star Algorithm",
  "definition": "The A* algorithm is an informed best-first graph-search method…",
  "domain": "artificial-intelligence",
  "maturity": "established",
  "subClassOf": [
    {"@id": "urn:ngm:class:search-algorithm", "label": "Search Algorithm"}
  ],
  "relations": {
    "hasPart": [{"@id": "urn:ngm:class:priority-queue", "label": "Priority Queue"}]
  }
}
```

- ### Definition
  - A* expands the open-list node with the lowest f(n) = g(n) + h(n)…

Three design consequences worth stating plainly:

• vc:public is the only publication gate, defaulting to false when absent (pipeline/jsonld_parser.py:205) and re-checked independently in each of the five emitting modules: Turtle, WebVOWL, page API, search index and graph tiers. Fourteen pages in the private source are withheld by it and are not in this repository.
• Wikilinks live in a weaker namespace than classes. urn:visionflow:linked:* is a surface link that may resolve to nothing; urn:ngm:class:* is an asserted class reference. Conflating them would fabricate taxonomy.
• Parse the JSON, never grep it. 1,004 pages emit their Page block as compact single-line JSON. A whitespace-sensitive grep '"vc:public": true' matches only 6,984 of 7,874 files — an 890-page silent drop, and the reason the class-count contract gate is in CI (.github/workflows/build.yml, gate 3: stats.json classes and ontology.json class[] length must both equal the pinned EXPECTED_CLASSES or the build fails; the pin reads 7874 and moves only alongside the corpus).

The context document that the class blocks dereference is static/ns/v2.jsonld (JSON-LD 1.1), which maps the bare terms onto rdfs, owl, skos and prov and declares the twelve canonical relations: hasPart, partOf, requires, enables, dependsOn, implements, uses, supports, contrastsWith, bridgesTo, relatedTo, standardizedBy.

How it fits together

flowchart LR
    subgraph source["Source"]
        md["ontology/pages/<br/>7874 Logseq .md<br/>+ embedded JSON-LD"]
        ctx["static/ns/v2.jsonld<br/>JSON-LD 1.1 context"]
    end

    subgraph pipe["pipeline/ — 7 stages, ~18 s"]
        p1["parse"] --> p2["validate"] --> p3["Turtle"]
        p3 --> p4["WebVOWL JSON"] --> p5["page API"]
        p5 --> p6["search index"] --> p7["NGG1 tiers"]
    end

    subgraph art["dist/ artefacts"]
        ttl["data/ontology.ttl<br/>258200 triples"]
        vowl["data/ontology.json<br/>WebVOWL"]
        bin["data/graph/*.bin<br/>NGG1 binary tiers<br/>+ bridges.json · overview.json"]
        api["api/pages/*.json<br/>api/search-index.json"]
    end

    subgraph cons["Consumers"]
        rdf["RDF tooling<br/>rdflib · Oxigraph · Protégé"]
        exp["explorer/<br/>React + Rust/WASM"]
        vc["VisionClaw<br/>Oxigraph + Whelk"]
    end

    md --> p1
    ctx -.defines.-> p1
    p3 --> ttl
    p4 --> vowl
    p7 --> bin
    p5 --> api
    p6 --> api
    ttl --> rdf
    ttl --> vc
    bin --> exp
    api --> exp
    vowl --> exp

Loading

Repository layout

Path	What it is
ontology/pages/	The corpus. 7,874 Logseq markdown pages, each with a Page block, a Class block and often a link-resolutions annotation.
pipeline/	The build. 9 modules plus a one-line __init__.py, 2,472 lines in total, one dependency (rdflib>=7.0.0).
pipeline/tests/	9 pytest cases, including a byte-exact 183-byte NGG1 golden fixture parsed by an independently written struct reader.
static/ns/v2.jsonld	The published JSON-LD 1.1 context the class blocks reference.
explorer/modern/	React 19 + Vite 6 + React Three Fiber SPA. One instanced mesh for all nodes, one line-segment mesh for all edges.
explorer/rust-wasm/	Crate webvowl-wasm 0.3.4, NGG1 reader (src/ngg1.rs) and CSR force simulation (src/layout/csr_sim.rs), #![deny(unsafe_code)].
explorer/FORMAT-NGG1.md	The frozen binary contract. Three implementations (Python writer, Rust reader, TypeScript reader/writer) agree byte-for-byte against it.
explorer/CAPABILITIES.md	Honest capability register for the explorer: what is verified, and with what evidence.
dist/data/	Committed dataset artefacts: Turtle, WebVOWL JSON, NGG1 tiers, stats.json, bridges.json, overview.json.
dist/api/	Website surface: 15,701 files, gitignored, rebuilt in seconds.
docs/	The method in prose: methodology/the-hybrid-approach.md, architecture/pipeline.md, architecture/explorer.md, ci-cd/build-and-gates.md, reference/jsonld-schema.md, ecosystem.md.

Quickstart

git clone https://github.com/DreamLab-AI/knowledgeGraph.git
cd knowledgeGraph

python3 -m venv .venv
source .venv/bin/activate
pip install "rdflib>=7.0.0"

python -m pipeline.build ontology/pages dist

Expected output, from a clean run:

[1/7] Parsing ontology/pages...
       7874 pages (7874 OntologyClass, 7874 public)
[2/7] Validating...
       0 errors, 0 warnings
[3/7] Generating Turtle...
       258200 triples → dist/data/ontology.ttl
...
Pipeline complete in 18.3s

Individual stages, if you want only one artefact:

python -m pipeline.validate ontology/pages [--json]         # advisory report
python -m pipeline.jsonld_to_turtle ontology/pages out.ttl  # Turtle only
pip install pytest && python -m pytest pipeline/tests -q    # 9 passed

The explorer needs a Rust toolchain and Node:

cd explorer/rust-wasm && wasm-pack build --release --target web --out-dir pkg
cd ../modern && npm install && npm install ../rust-wasm/pkg && npm run build

What you get, and how to consume it

dist/data/ontology.ttl: 258,200 triples, 12.1 MB. Load it in anything that speaks RDF. IRIs are https://narrativegoldmine.com/class/<slug>, rewritten from the urn:ngm:class:<slug> form in the source. The ontology targets OWL 2 EL: owl:inverseOf and owl:SymmetricProperty are deliberately omitted because neither is in the profile. requires and dependsOn are transitive; requires is a sub-property of dependsOn; uses, supports and implements are sub-properties of vc:utilises. Every requires and hasPart edge onto a declared class also emits an owl:Restriction (someValuesFrom) as an extra rdfs:subClassOf, and the six domain roots are asserted pairwise disjoint via a single owl:AllDisjointClasses.

from rdflib import Graph
g = Graph().parse("dist/data/ontology.ttl", format="turtle")
print(len(g))  # 258200

dist/data/ontology.json is 38.5 MB of WebVOWL-format JSON (class / classAttribute / property / propertyAttribute mirror arrays), consumable by WebVOWL-lineage viewers.

dist/data/graph/*.bin holds NGG1, a little-endian CSR binary designed so a browser can fetch a tier and hand it straight to WebAssembly with no JSON parse. 32-byte header, 24-byte node records, CSR adjacency, u8 edge types, and a UTF-8 string table where strings[n*2] is node n's label and strings[n*2+1] its IRI. full.bin is 7,874 nodes and 98,776 edges in 1,339,983 bytes. Six domain tiers cap at 1,500 nodes and 8 object-property edges per source node (the subClassOf backbone is never capped, and node degree is always the full-graph value so scale and label ranking stay correct). The format is specified in explorer/FORMAT-NGG1.md §1–§7.

dist/data/graph/bridges.json is the overlap the binary cannot carry. The NGG1 node record holds a single u16 category (FORMAT-NGG1 §3), so a tier keeps only the nearest category ancestor of each node; the full membership lives here — 542 entries, each with the class IRI, label, every category and domain index it belongs to, and the parent labels that put it there. Indices match overview.json, whose 40 nodes (6 domains + 34 categories) and 124 edges (34 backbone, 90 weighted bridges) are the category-level summary graph. The bridge edges also feed the force layout, so the baked positions match the topology the explorer draws.

dist/api/: 7,874 class JSON files plus a _domain-index.json, a flat search index (6.8 MB), and a markdown mirror of each page body. 7,823 of 7,874 pages have a body to mirror.

dist/data/graph/stats.json carries every headline number above, machine-readable, including the bridging block (multiParent 1,401, crossCategory 454, crossDomain 153).

Licensing

Three components, three licences, deliberately. Where the American spelling appears below it is a literal filename on disk; the prose uses "licence" for the noun.

Component	Licence	Why
pipeline/, CI, integration glue	AGPL-3.0-or-later	Wholly original work; copyleft chosen so pipeline improvements stay open when run as a service.
explorer/	MIT	A derivative of WebVOWL, © 2014–2019 Vincent Link, Steffen Lohmann, Eduard Marbach, Stefan Negru, Vitalis Wiens (explorer/license.txt), also published as DreamLab-AI/WasmVOWL. AGPL-ing it would be hollow while identical code sits MIT one repository away.
ontology/ corpus and dist/data/	ODbL-1.0	A database, licensed as one. Rights basis: UK CDPA 1988 s.9(3), computer-generated works vest in the person who made the arrangements.

Full texts: LICENSE (AGPL-3.0), LICENSE-EXPLORER (MIT, carrying the WebVOWL copyright verbatim), LICENSE-DATA (ODbL-1.0). The per-path mapping and the reasoning behind the split are in LICENSING.md; separate commercial terms are in COMMERCIAL.md.

Ecosystem

Verified 2026-07-25. Only repositories with real content and a real relationship to this one are listed.

Repository	Relationship	Licence
VisionClaw	Second, independent consumer of the same corpus. Its GitHubSyncService writes SHACL-gated triples into an embedded Oxigraph store and reasons over them with Whelk. Does not build this site.	AGPL-3.0-only
VisionFlow	Ecosystem canon: vision report, ADRs, PRDs, website.	No root licence file
WasmVOWL	Upstream of explorer/; itself a fork of VisualDataWeb/WebVOWL. Default branch is master.	MIT
agentbox	Agent runtime used to build and maintain this pipeline.	AGPL-3.0
Metaverse-Ontology	Independent precedent: Logseq markdown to OWL Functional Syntax via a Rust extractor, rather than embedded JSON-LD via Python.	No licence file

VisionClaw's pipeline documentation: system overview · graph schema · REST API. Note that VisionFlow's docs/architecture/licensing.md still lists VisionClaw as MPL 2.0; that is stale. VisionClaw relicensed to AGPL-3.0-only at ADR-032 M1 when it began linking the solid-pod-rs crates, and its NOTICE records why.

Status

Works, verified by running it. The seven-stage build completes in about 18 s with 0 errors. python -m pytest pipeline/tests -q reports 9 passed, including the byte-exact NGG1 golden. The Rust crate carries 154 unit tests and 49 integration tests.

The published site is one build behind. dist/ in this tree is the 2026-07-25 build described above. https://narrativegoldmine.com/data/graph/stats.json still returns the 2026-07-23 build (classes: 7457, uncategorised: 4498, no bridging block); pipelineVersion is ng-1.0.0 in both, so it does not distinguish them — read datasetDate and classes. The corrected build lands on the site at the next publish.

Partial. The explorer's SharedArrayBuffer position transport is disabled in shipped builds: an unsynchronised read/write race let the force simulation amplify half-written frames to ~1e20 and blank the view in production only. Transferable ping-pong is the sole transport until a double-buffered, Atomics-gated flip lands. MAX_EDGES = 4000 is declared to mirror the client contract but is not enforced by the writer; domain tiers legitimately ship up to 12,093 edges because the client sub-selects. Four benchmark suites cover the inherited WebVOWL stack and none covers the live NGG1 path.

Known gaps in the data. All of these are measured, not estimated. Three of the gaps the previous release listed here have gone, and the reasons are worth naming: the category walk was one hop deep and mislabelled 4,033 classes as uncategorised; MULTI_PARENT was reported as a warning when it is a design property; and a repair pass created 417 pages for concepts that two or more existing pages already referenced, which also cleared the four invalid domains. What follows is what is left.

• 1,401 validation notices, all info, all MULTI_PARENT. 0 errors, 0 warnings. The notice is a statement about the shape of the taxonomy, not a fault; see the lattice note above.
• 3 of 7,874 classes are uncategorised. Category membership is now resolved by a breadth-first walk of subClassOf/instanceOf ancestry, nearest category ancestor winning, parents visited in declared order so the NGG1 tiers stay byte-identical across runs (pipeline/emit_graph_tiers.py:480). The deepest real path needs 7 hops; the walk stops at 12 (MAX_DEPTH). The three with no category root anywhere in their ancestry are electric-vehicle, ethan-mollick and urban-planning.
• The binary format carries one category per node. The NGG1 node record is <u32 id, f32 x, f32 y, u16 domain, u16 category, u8 flags, 3× pad, u32 degree>, so a tier keeps only the nearest category and the other memberships of the 454 cross-category classes are invisible to any consumer reading .bin alone. They are in bridges.json. This is a limitation of the frozen format, not of the data.
• Unresolved reference targets. 111,827 declared edges resolve to 98,776; the 13,051-edge difference is self-loops, duplicates, and references to classes that do not exist. Most of the missing targets are named exactly once in the whole corpus, and materialising a page for each singleton is deferred rather than done.
• 789 classes emit vc:qualityScore "0.0". The context defines qualityScore; the parser reads quality first and falls back to vc:qualityScore, a key that appears in 0 of the 7,874 pages. 742 of the 789 carry a bare qualityScore the parser never looks at; the other 47 carry no score at all. A further 1,259 carry both quality and qualityScore, and the two disagree on 1,013 of them.
• stats.json reports 7,870 pages against 7,874 classes. Four pairs of files share a page IRI (bitcoin, comfy-ui, ethereum, foundation-models). The page count is deliberately deduplicated by IRI rather than reported as the class count under a "pages" label.
• 0 individuals. The corpus is terminology only: every entity is a class, and nothing is asserted as an instance of one.
• 97 wikilinks across 2 files (Data Parallelism.md, Data Preprocessing.md) are bare strings rather than {@id, vc:label} objects and therefore produce no backlinks.
• 51 pages have an empty body and so get no markdown mirror in dist/api/markdown/.
• 95 classes carry a urn:ngm:class: slug that differs from their page's vc:slug. A substring rewrite in the corpus source turned bitcoin-cash into bitcoin-proof-of-work-protocol-cash and apple-vision-pro into apple-inc-technology-corporation-vision-pro. For those 95 the class IRI in ontology.ttl and the filename in dist/api/pages/ do not join.
• Backlinks are derived by slug-suffix matching, not IRI resolution, so a wikilink whose slug differs from the target page's vc:slug silently produces nothing.
• Vocabulary drift is tolerated, not corrected. maturity takes 9 observed values where 4 were intended; domain takes 16 raw strings that a 10-entry alias table collapses onto 6. 45 non-canonical relation predicates appear in the corpus and are dropped by the parser.

Provenance

Class-level attribution resolves to four DIDs: did:nostr:ontology-mesh (3,541), did:nostr:lcr-swarm (1,565), did:nostr:jjohare (1,079) and did:nostr:enrichment-swarm (226), with the generating pass named in provenance.inferenceRule: GapMaterialisation, RelationEnrichment, R5DomainRootFallback, R1Explicit, ManualEnrichment and others. Those rule names are self-describing about machine origin, which is the point. Coverage is partial: 6,411 of 7,874 classes carry a provenance.attributedTo; the remaining 1,463 carry none. The record is an audit trail of generation, not a claim of authorship.