Record Ontology

curl -H "Accept: text/turtle" https://epistemic-ontology.net/record

Namespace

A hash namespace: dereferencing any term — say …/record#SelfVerifying — returns the whole ontology document, and your client resolves the fragment. The recommended prefix is rec:.

@prefix rec: <https://www.epistemic-ontology.net/record#> .

Every worked example is dereferenceable too, under a systematic IRI:

@prefix ex: <https://www.epistemic-ontology.net/record/examples/neptune-discovery#> .

The founding constraint

No class may be defined by a relation to an object that lies outside all records.

This is the rule the whole ontology is shaped by. Records are named by what the agent holds — form, warrant, givenness, relations to other records, pragmatic adequacy — never by a certified correspondence to a thing-in-itself. It is built for an agent (a person or an AI) and never steps outside one.

The one class, and the one non-record

There is deliberately no Form class. The form-in-itself is no more available to an agent than the thing-in-itself; a Form class would be the formal apple. Everything an agent holds is a Record.

Warrant — a triad, one value per limit

Warrant is an attribute, not a kind, so patchworks decompose: a real artifact is formal at some joints and conventional at others.

Two limits are excluded and never instantiated as classes. The third is not: the Agent is the one thing given to itself, and self-verifying warrant is exactly the warrant that reaches it.

Inference, support, metadata

The regress "what carries the carrier?" is not answered with another entity. It stops at a warrant. Credit Descartes for the existence, Hintikka for the warrant — and not the res cogitans: the cogito certifies that the agent is, never what it is.

Exercise or exorcise

hasWarrant rec:Formal is a claim, not a credential. A formally warranted record's formulation should be a description that can be run:

claim → exercise (run the description) → confirmed or exorcised

Never exercised means testimonially held — which is how most mathematics is in fact carried. A failed exercise expels the pretender, not the record: the document survives, records about it survive, but it can no longer transmit support, and everything resting on it cascades away. Kempe's 1879 "proof" of the four-colour theorem stood eleven years until Heawood exercised it; as of 1885, trusting it was justified.

The engine — dynamics over the statics

OWL DL is monotonic: it cannot retract. So all dynamics live in a computational layer over the static ontology — the engine/ package in the source repository, never axioms in the Turtle. The same rule strikes twice, and symmetrically: dynamics live over the statics because DL is monotonic; computation lives over the statics because DL is decidable.

• Revision log as the carrier of moments — an append-only stream of ground assertions, retractions and decisions. Log position is order-derived time, and any state is computable as of any moment. Derived records never enter the log; they regenerate by replay.
• Forks and corroboration — structural detection deliberately over-detects, so rivalry is declared, as a logged decision. Corroboration is temporal: only empirical evidence first asserted after a fork opened can collapse it. The losing branch is eclipsed, never deleted.
• Fidelity — not probability-of-truth (that would be the apple in numeric dress) but invariance under the web's own dynamics, computed from the ATMS label: a record's minimal environments. Scalars are purpose-relative projections.
• Identification — rivals proven equivalent make the fork dissolve rather than collapse, and their environments pool. Matrix versus wave mechanics (1926) is the type case.

Stance: between the OWL/DL and SKOS camps

DL-in-form, ecumenical-in-content. Every concrete choice here is the DL choice: defined classes, reasoner re-derivation, owl:disjointWith with a consistency check, rdfs:label rather than skos:prefLabel, and a deliberate refusal to import SKOS because it is OWL Full.

But the standing schism is demoted from a choice of framework to the rec:hasWarrant attribute: a DL hierarchy is a record warranted formally (subsumption); a thesaurus is one warranted empirically (community-fixed, defeasible — skos:broader is a curatorial edge, not a deductive one). A real scientific taxonomy is both. The ontology does not take a side; it represents the axis the camps fight over.

Validation

pip install -r requirements-dev.txt
python scripts/validate.py

A pure-Python OWL 2 RL reasoner (owlrl, no Java) checks four things:

1. Defined-class entailment — strip the asserted Inference type and confirm the definition re-derives it, with a negative control.
2. Cogito pattern — the pattern holds, and a self-verifying promise is correctly not conflated with it.
3. Consistency — no individual is both Record and Continuum.
4. Sub-property entailment — every metadataOf edge entails directedToward.

Status and open items

The conceptual source of truth is ROOT.md in the source repository; this page summarises what is settled. Three markers run through it — settled (encoded in the Turtle), prototyped (working code in engine/), unbuilt (recorded as horizon).

Known open items: composition transitivity is left off on purpose; the escalation and formation layer is recorded as horizon, not vocabulary; the knowledge-automaton — a groomed web whose formal interior regenerates by deduction — is an engine/asset question with no vocabulary proposed.

@misc{recordontology2026,
  title   = {The Record Ontology: A Domain-Neutral Model of Warranted Record-Structure},
  author  = {Hinchley, Ron},
  year    = {2026},
  version = {0.6.0},
  url     = {https://epistemic-ontology.net/record}
}