OWL/RDFS axioms as documentary AI-signal — the consolidated doctrine

Scope amendment — 2026-09-05. This record remains proposed and preserves the earlier schema-derived implementation’s rationale and evidence. Its “documentary any-of” interpretation does not govern the SPDTF method. The selected source ODRs apply under ADR-0063’s 2026-09-05 amendment; see the modelling decision basis for the normative selection and implementation boundary.

Source ODR-0014’s 2026-08-28 correction supersedes the claim that disabling domain/range inference permits repeated RDFS values to mean disjunction. A runtime may decline materialisation; it cannot change published RDFS/OWL semantics. Source ODR-0014’s 2026-08-30 amendment, read with source ODR-0030/0036, requires complete, exclusive applicability documentation per property side: one universally true RDFS domain/range, Schema.org inclusion hints, or a governed waiver. SHACL expresses closed-world validation; none of these mechanisms substitutes for another. The historical body below is retained without certifying migration of its emitted triples, inference rules or validators.

Context and Problem Statement

OPDA’s entailment regime runs only a frozen 7-rule RDFS-Plus load-time closure (ODR-0025 §R1) and excludes rdfs:domain/rdfs:range and the rest of the “R2-excluded set” (owl:FunctionalProperty/InverseFunctionalProperty, owl:equivalentClass/equivalentProperty, owl:Restriction/unionOf/intersectionOf/complementOf/cardinality/oneOf/hasKey) from inference. Yet OPDA continues to author many of those constructs — rdfs:domain/range on ~750 datatype properties, scoped owl:disjointWith, property characteristics — as documentation and machine-legible signal of modeller intent. The positive principle that reconciles “author it but never infer from it” was stated only in fragments scattered across ODR-0025 §R2, ODR-0026 §R2, ODR-0027 §R5, and ODR-0029 §R1.

That fragmentation had a concrete cost. Council session-047 (implemented in ADR-0048) re-derived rdfs:domain/range from standard OWL/RDFS entailment semantics — the “everything-becomes-a-Person” trap — concluded the relationship-layer object properties (founds/mediates/playedBy/plays/hasAddress) must therefore drop documentary domain/range and pin types in SHACL only, and made those edges invisible to the class-graph model derivation. That reasoning is wrong under OPDA’s own regime: OPDA opted out of standard OWL/RDFS entailment, so reasoning about OPDA from those semantics yields false conclusions. The session-047/048 records were later corrected on exactly this point. The recurring nature of the error (two independent council passes fell into it) is the warrant for stating the doctrine once, as a single citable record, rather than leaving it implicit across four ODRs.

This ODR is the consolidating doctrinal home. It states the positive principle, the per-construct admission test, the standing corollary, and the one carve-out (FP/IFP), and cross-references the four source ODRs whose fragments it gathers. It changes no decision in those records — each retains its own normative rules; this record names the shared doctrine they all instantiate.

Decision Drivers

• Doctrine consistency. The regime already excludes domain/range and the R2 set from inference; authoring them is therefore safe by construction, and the ~750 datatype properties already do it. The principle should be stated, not re-litigated per construct.
• AI/LLM signal. OWL/RDFS axioms are machine-legible documentation of modeller intent — an LLM or external tool reading the TBox should see the author’s typing/disjointness/characteristic intent, even though OPDA’s own pipeline never reasons over it.
• Standards-legibility. Author W3C-legible axioms even though OPDA never reasons — a third-party DL tool may evaluate them independently, out of OPDA’s load-time scope.
• Proof obligation, not trust. Whatever is authored MUST NOT change the closure — the ADR-0035 zero-domain/range-triple consistency test is the formal proof that “model-but-don’t-evaluate” holds.
• A bounded admission test, not a blanket licence. Authoring-as-signal is admissible only where the construct’s published W3C semantics agree with OPDA’s enforced behaviour; where they disagree (notably IFP), authoring publishes a proposition OPDA holds false.

Considered Options

• Option A (chosen) — State one consolidated doctrine: author OWL/RDFS axioms as documentary AI-signal, never entailed; admit each R2-set construct only where its W3C semantics agree with OPDA’s enforced behaviour; carve out FP/IFP. A single citable record; the four source ODRs keep their rules and cite this as their shared principle.
• Option B — Leave the principle fragmented across ODR-0025 §R2 / 0026 §R2 / 0027 §R5 / 0029 §R1. Rejected: the fragmentation is exactly what let session-047/048 re-derive domain/range from standard OWL semantics twice; there is no single statement to cite against the error.
• Option C — State the doctrine as “author everything OWL, validate in SHACL” with no per-construct test. Rejected: a blanket licence would admit owl:InverseFunctionalProperty, whose published semantics assert the negation of ODR-0005’s bounded-context-identity ruling. The admission test (engineering-act = ontological-act) is load-bearing, not decoration.
• Option D — Drop the documentary axioms and express everything in SHACL only. Rejected: loses the documentary / AI-signal value and the standards-legibility of the TBox; the ~750 datatype properties already carry both the axiom and the SHACL dual without harm.

Decision Outcome

Chosen option: “Option A — state one consolidated doctrine”, because the principle is uniform across every R2-set construct, the cost of leaving it implicit is a demonstrated, recurring reasoning error, and a single record gives every future modeller and council one statement to cite.

The doctrine. OPDA authors OWL/RDFS axioms — rdfs:domain/range, property characteristics, owl:disjointWith, disjunctions — as documentary, AI-legible signal of modeller intent; they are NEVER entailed. The only inference is the frozen 7-rule RDFS-Plus load-time closure (ODR-0025 §R1). Everything else is authored for documentation / AI-signal AND validated via SHACL (ODR-0029 §R1/§R3), never reasoned.

Per-construct discipline (the admission test). Author an R2-set construct as documentary signal only where its published W3C semantics AGREE with OPDA’s enforced behaviour — “engineering-act = ontological-act.” Where the W3C reading of the axiom matches what OPDA’s SHACL actually does, the axiom is honest documentary signal (domain/range agrees with the sh:class/sh:or act → in). Where it disagrees, authoring it would publish a proposition OPDA holds false (IFP disagrees → out).

Standing corollary. Do NOT reason about OPDA from standard W3C OWL/RDFS entailment semantics. OPDA opted out of those semantics for everything outside the 7 frozen rules; a conclusion derived by applying standard OWL/RDFS inference to OPDA’s TBox (e.g. “multiple rdfs:domain ⇒ every subject is the intersection of all domains,” or “an IFP merges co-referents”) is unsound as a statement about OPDA, regardless of its correctness as a statement about standard OWL. This corollary is the explicit guard against the session-047/048 error.

The FP/IFP carve-out (session-050). owl:InverseFunctionalProperty and a general owl:FunctionalProperty documentary layer are NOT authored — the single place OPDA’s local constraint overrides authoring-as-signal, and it does so by the admission test itself. A published IFP asserts the negation of ODR-0005’s bounded-context-identity ruling: its W3C meaning (OWL 2 §2.3.5) promotes a contingent identifier to a constitutive global identity criterion (“shared value ⇒ owl:sameAs everywhere”), the inverse of OPDA’s ruling — an OntoClean rigidity-clash. Uniqueness lives in SHACL: dash:uniqueValueForClass, scoped within each identity-bearing sortal, never cross-sortal (a cross-sortal key is IFP in SHACL clothing — it would fire on legitimate UPRN co-reference across the Property Kinds). A narrow hand-curated owl:FunctionalProperty on a genuine world-fact singleton is admissible only if the modal marker is explicitly wanted — not a blanket pass.

Disjunction form. A multi-bearer (“any-of”) domain is expressed as repeated rdfs:domain read as a disjunction (the domainIncludes-in-RDFS idiom) under a CI-gated module-header convention note + per-property rdfs:comment, with the authoritative per-class disjunction in SHACL sh:or — NOT owl:unionOf (an excluded boolean class constructor; a union of disjoint sortals would reify a non-sortal carrying no identity criterion). This keeps the corpus boolean-constructor-free.

Consequences

• Good, because the doctrine is stated once and applies uniformly; future modellers and councils cite this record instead of re-deriving the principle per construct.
• Good, because the standing corollary is an explicit, citable guard against the recurring “reason from standard OWL semantics” error that cost two council passes.
• Neutral, because nothing changes at inference time: the closure still adds ZERO domain/range triples (the ADR-0035 test is the proof obligation), and the four source ODRs keep every decision and rule unchanged — this record consolidates, it does not supersede.
• Bad (accepted), because authoring documentary axioms is a standing maintenance commitment, and the admission test plus the FP/IFP carve-out must be enforced by CI (the excluded-construct meta-shape and the object-property-coverage gate).

Confirmation

• The ADR-0035 zero-domain/range-triple consistency test passes: the load-time closure materialises no domain/range/equivalence/sameAs triple — the formal proof that authored documentary axioms are never entailed.
• The excluded-construct CI meta-shape fails on any authored owl:InverseFunctionalProperty (and on owl:unionOf), enforcing the carve-out and the disjunction form.
• The relationship-layer ci-object-property-coverage gate requires the documentary domain/range + matching SHACL dual for disjunction predicates.

Rules

R1 — Author OWL/RDFS axioms as documentary AI-signal; never entail them

OWL/RDFS axioms outside the 7 frozen rules — rdfs:domain, rdfs:range, property characteristics, owl:disjointWith, disjunctions — are first-class modelling vocabulary authored wherever they carry information (human documentation, machine-legible signal of modeller intent, input for external DL tooling). The load-time closure (ODR-0025 §R1) evaluates none of them. They are authored for documentation/AI-signal AND validated via SHACL (ODR-0029 §R1/§R3), never reasoned. This is the consolidated statement of the “model-but-don’t-evaluate” principle previously fragmented across ODR-0025 §R2, ODR-0026 §R2, ODR-0027 §R5, and ODR-0029 §R1.

R2 — The per-construct admission test (engineering-act = ontological-act)

Author an R2-set construct as documentary signal only where its published W3C semantics agree with OPDA’s enforced (SHACL) behaviour. The test is applied per construct, not blanket:

• rdfs:domain/rdfs:range — in. The closed-world SHACL reading (“are the subjects of <pred> of class C?”, ODR-0029 §R3) agrees with the documentary intent; the generative W3C reading is never run, so the disagreement never materialises (proven by the ADR-0035 test).
• owl:disjointWith — in, scoped. Authored only between rigid sortals with their own, incompatible identity criteria where a real misclassification hazard occurs; the authored pairwise axiom feeds a real consistency gate. (Scope and the three-part bar are governed by ADR-0049 / session-050; this record states only the admission principle.)
• owl:InverseFunctionalProperty / general owl:FunctionalProperty — out (R4). The W3C semantics disagree with OPDA’s enforced behaviour.

The asymmetry is principled, not ad-hoc: the same test admits domain/range (agrees) and excludes IFP (disagrees).

R3 — Standing corollary: do not reason about OPDA from standard W3C OWL/RDFS entailment

A conclusion reached by applying standard OWL/RDFS inference to OPDA’s authored TBox is unsound as a statement about OPDA. OPDA’s only entailment is the 7 frozen rules; every other construct is authored-but-unevaluated. Specifically: authored multiple rdfs:domain do not intersect-type their subjects; an authored characteristic does not fire its W3C inference; an authored disjunction does not classify. The correct reading of any such axiom in OPDA is “documentary signal, validated in SHACL.” This corollary exists because two independent council passes (session-047, session-048) derived rdfs:domain/range behaviour from standard OWL semantics and reached a wrong disposition that had to be corrected.

R4 — The FP/IFP carve-out

Do NOT author owl:InverseFunctionalProperty (out entirely) or a general owl:FunctionalProperty documentary layer. The home for uniqueness is SHACL: sh:maxCount 1 and dash:uniqueValueForClass, scoped within each identity-bearing sortal and never cross-sortal. A narrow hand-curated owl:FunctionalProperty on a genuine world-fact singleton is admissible only when the modal marker is explicitly wanted.

Binding ground (for permanence): a published IFP asserts the negation of ODR-0005’s bounded-context-identity ruling — it promotes a contingent identifier (a UFO Quality) to a constitutive global identity criterion. This is the binding reason; redundancy-with-the-safe-SHACL-substitute (the session-035 drop-redundant-axiom rule, ODR-0026 §R3) is corroborating only. The carve-out re-opens only if ODR-0005 is itself overturned; it is not re-openable by a modelling-economy argument.

R5 — Disjunction form: repeated rdfs:domain, never owl:unionOf

A multi-bearer (“any-of”) domain is authored as repeated rdfs:domain read as a disjunction (the domainIncludes-in-RDFS idiom), under a CI-gated module-header convention note and a per-property rdfs:comment, with the authoritative per-class disjunction carried in SHACL sh:or. owl:unionOf is not used — it is an excluded boolean class constructor, and a union of mutually disjoint sortals reifies a non-sortal that carries no identity criterion. This keeps the corpus boolean-constructor-free.

R6 — Permitted/excluded OWL-construct table (re-keyed to OPDA’s frozen closure)

This is the per-construct disposition for every OWL/RDFS modelling construct OPDA may encounter — the ODR-0030-adopted construct table from the sibling semantic-modelling corpus, re-keyed to OPDA’s frozen 7-rule closure (ODR-0025 §R1). OPDA’s documentary-only band is LARGER than hm’s: hm runs disjointness (and more) as inference, whereas OPDA validates disjointness via SHACL + the ADR-0035 consistency gate and infers it never. Three bands:

Permitted + inferred (the 7-rule closure, ODR-0025 §R1 — the only constructs OPDA reasons over):

Permitted + documentary-only (authored as AI-signal, NEVER inferred — R1/R2):

Excluded entirely (never authored — neither inferred nor documentary):

CI enforcement. The excluded band is enforced corpus-wide by the ci-excluded-construct meta-shape gate (ADR-0049 task 4): it fails if ANY excluded construct appears anywhere in the class graph, detected as a real triple component (a predicate, or the object of an rdf:type axiom) — never as a substring of an annotation literal, so this record’s own prose mentions of e.g. owl:unionOf do not trip it. This generalises the FP/IFP limb (b3) of the ci-object-property-coverage gate (which catches FP/IFP on object properties only) to the full excluded set, corpus-wide. The corpus has zero excluded constructs today, so the gate passes and arms drift protection. owl:FunctionalProperty is not failed corpus-wide (a hand-curated singleton is admissible, R4); its object-property carve-out is enforced narrowly by ci-object-property-coverage limb (b3).

More Information

• Consolidates (these records keep their decisions; this names their shared doctrine):
  • ODR-0025 §R1 (the frozen 7-rule closure — the only inference) and §R2 (the excluded set — non-evaluation, not a prohibition on authoring).
  • ODR-0026 §R2 (model-but-don’t-evaluate) and §R3 (drop the redundant axiom where a safe evaluable substitute carries the signal — the session-035 corroborating ground for R4).
  • ODR-0027 §R5 (rdfs:domain/range are documentation; SHACL enforces, value-keyed).
  • ODR-0029 §R1 (the inference/validation boundary), §R2 (frozen rule logic), §R3 (domain/range-as-SHACL-constraint).
• Identity ground for the carve-out: ODR-0005 §R5 (bounded-context identity; no owl:sameAs) — the binding reason IFP is excluded.
• Engineering realisation: ADR-0049 (the per-construct adoption, relationship-layer correction, and the excluded-construct + coverage gates); ODR-0032 §R1/§R2 (the relationship-layer correction this doctrine motivated); ADR-0035 (the zero-domain/range-triple proof obligation).
• Council deliberation: session-048 (the per-construct admission test; the withdrawal of the “everything-becomes-a-Person” premise as applied to OPDA) and session-050 (5–0; the FP/IFP carve-out and its ODR-0005-binding ground; the disjunction form).
• Prior art (project-neutral statement of the idioms above): the domainIncludes-in-RDFS idiom for “any-of” domains is mainstream prior art (schema.org domainIncludes, DCMI, FIBO declares domain/range and enforces via SHACL); the multiple-rdfs:domain-is-conjunction reading is RDFS 1.1 §3.2; the IFP→owl:sameAs semantics are OWL 2 §2.3.5; annotation properties carry no model-theoretic consequence per OWL 2 §10.1. The sibling semantic-modelling corpus is the prior-art source from which OPDA adopted this OWL coverage (see ADR-0049 §More Information for the per-construct adoption map).