0020 Ontology: superposition at the center, borrowed vocabulary, per-scope indexes, genealogy (jokbo) and GEDCOM
The Korean version is authoritative. This page was translated by AI (Claude) and has not yet been checked by a person. 한국어 원문
This decides how to handle the ontology (a vocabulary system that sets out what kinds of things exist and how they connect). It is based on four Nodeone surveys (Korean institutions, Palantir, graph storage, AI memory, 2026-10-20) and their summary. It is reflected in 06-Relation network v0.6, 03-manifest v0.9 and 05-Glossary.
1. Superposition at the center
Section titled “1. Superposition at the center”One line is a bundle of several claims from different sources. They are not merged into one but kept side by side, and the view decides what comes first (06-Relation network sections 1 and 7, decision 0007). The international basis is CIDOC-CRM E13 (Attribute Assignment) and the factoid model. This direction does not change.
2. Vocabulary is borrowed
Section titled “2. Vocabulary is borrowed”- Relations and properties from Wikidata, events and objects from CIDOC-CRM and Linked Art, provenance from PROV-O, preservation events from PREMIS, rights and consent from ODRL and DPV.
- The edition handed to Korean institutions adds Dublin Core, SKOS, FOAF, schema.org and
owl:sameAson top. These are the common ground used by National Library of Korea LOD and the Korean studies digital humanities model. They come out in 08-Name file v2. - Korean institutions cannot take the “claim + source + status” shape. When handing over, we give a flat edition (derived) with “only confirmed claims, picked by one view”.
3. Newly defined terms (06-Relation network section 5)
Section titled “3. Newly defined terms (06-Relation network section 5)”- Needed for old Korean records:
clan_seatclan seat (bon-gwan),posthumous_nameposthumous name (siho),temple_nametemple name (myoho),title_grantedgranted title (bongjakho),social_statussocial status (sinbun),adopted_heiradopted heir (yangja),generationgeneration (se, 世), event kindexampassing the state examination (gwageo) (rolepasser), kin typekin_type(agnatic, maternal, affinal, adoptive). - Social association (gyoyu, keeping company) is not written as a claim. The view counts it from meeting and letter events.
- For a relative time (“last year”), the reference point goes in
when.anchor.
4. External identifiers (06-Relation network section 11.3)
Section titled “4. External identifiers (06-Relation network section 11.3)”- Added:
nlk(National Library of Korea author authority),sillok,sillok-person,itkc-people,heritage,jsg,kyu,korcis,emuseum. itkc-peopleis in beta, and under the Institute for the Translation of Korean Classics’ policy, the Institute is told when we link to it. Content is not copied.- Joseon-era places have almost no public identifiers, so our own
placenode becomes the owner.
5. Storage: ledger, per-scope indexes, export
Section titled “5. Storage: ledger, per-scope indexes, export”- The ledger (JSONL) is the original. The graph is an index that can be rebuilt (06-Relation network section 1).
- A separate index is built for each scope, and claims that cannot be seen are never put into it in the first place (06-Relation network section 9, item 4). If one index is filtered at query time, the fact that something exists can leak.
- The first index uses SQLite, and the RDF edition (TriG, JSON-LD) is made only on export. A graph database will be looked at again when several apps share one graph.
- arkive tests first with a ledger made with fake names (10,000 to 3 million claim lines).
6. Gaps filled on the AI side
Section titled “6. Gaps filled on the AI side”- When something is withdrawn or narrowed, the secondary indexes the AI made (vector indexes, summaries, caches) are rebuilt too, it is checked that it no longer shows up in answers, and this is recorded in the withdrawal line (06-Relation network section 9, item 6).
- When one source is withdrawn from a claim with several sources, the claim is made again from the remaining sources, and scope does not widen by itself (06-Relation network section 8).
- AI reading is kept apart from writing to the ledger. Text inside documents is never followed as an instruction (blocking memory poisoning).
- We make an hgp test suite: does scope leak, does what was withdrawn disappear from answers, does it say “I don’t know”. It comes out together with the HGP checker.
7. What we borrowed from Palantir and what we did not
Section titled “7. What we borrowed from Palantir and what we did not”- Borrowed: edits only through defined actions (already in 06-Relation network section 8), seeing only when both the scope and every “approval required” marker are met, a list of actions opened to AI (app key, decision 0011).
- Not borrowed: new values overwriting old ones, records kept only when turned on, administrators lifting markings, closed formats.
- The description of Palantir in 06-Relation network section 12 was corrected to match the facts (properties can hold several values, but there is no frame for attaching a source to each value).
8. Genealogy (jokbo) and GEDCOM
Section titled “8. Genealogy (jokbo) and GEDCOM”- A genealogy is a source record (original). One line of a genealogy (generation, name, courtesy name and art name, birth and death, spouse, grave, office) becomes several claims, and the source position is that genealogy’s face and line.
- When different editions of a genealogy (clan-wide genealogy, branch genealogy, lineage record) say different things, they are kept superposed as they are. The same applies when a genealogy, a collected works and the Annals say different things.
- Living people. Recent genealogies contain many living people. We follow 06-Relation network section 9, item 3 (possibly living persons are sealed from the start) as is. When exporting to GEDCOM, the same filtering is done first.
- GEDCOM 7 (FamilySearch, genealogy exchange format) is used as an import and export format (referenced).
- On import: the GEDCOM file itself is deposited as an original and gets a fingerprint. The facts inside it are machine-extracted candidate claims (
made_bymachine,candidate), and the source position is the line of the file. A person confirms them. - On export: GEDCOM can give the same person several events of the same kind, each with its own source, so it can carry over a good deal of superposition.
- The lunar calendar is not a calendar defined in GEDCOM. We define one extension calendar (a name starting with
_), and its name is decided during testing.
- On import: the GEDCOM file itself is deposited as an original and gets a fingerprint. The facts inside it are machine-extracted candidate claims (
- Testing with genealogy software and raising this to “aligned” will be done when there is a partner.
- The four surveys pointed the same way. Keeping claims superposed by source is almost absent from Korean institutions, from Palantir, and from AI memory tools. This is where HGP is ahead.
- No one in Korea can take our shape, so inside we keep our shape, and outward we publish in the light vocabularies they read.
- The graph database market is changing. Some tools shut down in 2025, and some switched to licenses that are not open source. If the ledger is the original, nothing is lost when tools change.
Other options
Section titled “Other options”| Option | Why not chosen |
|---|---|
| Building a large OWL ontology first | No Korean institution would read it, and automatic reasoning does not fit the principle of human confirmation |
| Making the graph database the original | If the tool disappears, everything goes with it. The tool does not protect superposition and records |
| Filtering permissions at query time in one index | The fact that something is hidden leaks |
| Taking genealogies only as GEDCOM | Source positions (face, line) and superposition are lost |
Where it has effect
Section titled “Where it has effect”- hgp: 06-Relation network v0.6, 03-manifest v0.9, 05-Glossary, 08-Name file v2 (after 03 is approved), test suite.
- arkive: per-scope indexes, testing with a fake ledger, new terms and external identifiers, a Linked Art export test, filtering living people when a genealogy is added as a source.
- secretary:
when.anchorwhen extracting expressions like “last year” as claims. Including secondary indexes in withdrawal. - Institute for the Translation of Korean Classics: tell them before publishing links to person numbers.
0 won.