GMEOW AI Claim Module

• IRI: https://blackcatinformatics.ca/gmeow/slices/ai
• Tier: core
• Group: core

What This Slice Covers

This slice owns 51 terms and contributes 23 mapping or projection rows. Use it when its terms match the native fact you want to preserve; use the linkage tables to see how those facts leave GMEOW for consumer vocabularies.

Dependencies

• gmeow:slices/entities
• gmeow:slices/evidence
• gmeow:slices/kernel
• gmeow:slices/observations
• gmeow:slices/provenance

Consumers

• The flagship (P14): the gmeow PyPI client's Memory store, the MCP store/recall/revise triad, the hallucination-resistant KG pattern + gmeow audit (EvidenceSpan audit machinery), the claim-extraction eval suite, and Project Lillith via extensions/graphrag.

Local Map

Examples

Grounded Claim

• Source: slices/core/ai/examples/grounded-claim.ttl
• GMEOW terms: gmeow:Chunk, gmeow:Contradiction, gmeow:Document, gmeow:EvidenceSpan, gmeow:Goal, gmeow:MemoryItem, gmeow:ModelCard, gmeow:ModelInvocation, gmeow:Prompt, gmeow:PromptTemplate
• External prefixes: owl, xsd

# SPDX-FileCopyrightText: 2026 Blackcat Informatics® Inc. <paudley@blackcatinformatics.ca>
# SPDX-License-Identifier: CC-BY-4.0
#
# Worked example: LLM output as claim-not-truth, on the EXISTING machinery.
# A grounded claim and an ungrounded one (the flagged hallucination), both
# StandpointClaims whose vantage is the model agent (the unified observation
# stance, P9); their conflict SURFACED by a Contradiction; the grounded claim
# remembered (MemoryItem role); the verdicts about the hallucination expressed
# with the standpoint slice's bullshit modality and the deception slice's
# veridicality — from an auditor vantage, like every other claim.
@prefix gmeow: <https://blackcatinformatics.ca/gmeow/> .
@prefix ex:    <https://blackcatinformatics.ca/gmeow/examples/ai/> .
@prefix owl:   <http://www.w3.org/2002/07/owl#> .
@prefix rdfs:  <http://www.w3.org/2000/01/rdf-schema#> .
@prefix xsd:   <http://www.w3.org/2001/XMLSchema#> .

# --- The source and its retrieval segment (the spine's left half).
ex:handbook a gmeow:Document ;
    rdfs:label "facilities handbook, 2026 edition"@en ;
    gmeow:contentDigest "blake3:9f8e7d6c5b4a39281706f5e4d3c2b1a09f8e7d6c5b4a39281706f5e4d3c2b1a0" .

ex:chunk-042 a gmeow:Chunk ;
    gmeow:chunkOf ex:handbook ;
    gmeow:spanStart "18200"^^xsd:nonNegativeInteger ;
    gmeow:spanEnd "18950"^^xsd:nonNegativeInteger ;
    gmeow:contentDigest "blake3:1a2b3c4d5e6f708192a3b4c5d6e7f8091a2b3c4d5e6f708192a3b4c5d6e7f809" .

# --- The model, its card, and the invocation (the provenance anchor).
ex:assistant a gmeow:SoftwareAgent ;
    rdfs:label "facilities assistant model"@en ;
    # Teleology: the agent's goal, asserted like any attributed claim.
    gmeow:hasGoal ex:goal-answer-faithfully .

ex:goal-answer-faithfully a gmeow:Goal ;
    rdfs:label "answer facilities questions faithfully to the handbook"@en .

ex:card-assistant a gmeow:ModelCard ;
    gmeow:describesModel ex:assistant ;
    gmeow:modelProvider "Example Labs" ;
    gmeow:modelVersionTag "assistant-2026-05-01" ;
    gmeow:modelContextWindow "200000"^^xsd:positiveInteger ;
    gmeow:modelTrainingCutoff "2026-01-31"^^xsd:date .

ex:prompt-extract a gmeow:Prompt ;
    gmeow:promptRole gmeow:promptRoleSystem ;
    gmeow:filledFrom ex:template-extract-v1 ;
    gmeow:contentDigest "blake3:00112233445566778899aabbccddeeff00112233445566778899aabbccddeeff" .

ex:template-extract-v1 a gmeow:PromptTemplate ;
    rdfs:label "claim-extraction template v1"@en ;
    gmeow:promptRole gmeow:promptRoleSystem .

ex:invocation-19 a gmeow:ModelInvocation ;
    gmeow:usedModel ex:assistant ;
    gmeow:hasPrompt ex:prompt-extract ;
    gmeow:samplingTemperature 0.0 ;
    gmeow:atTime "2026-05-15T10:00:00Z"^^xsd:dateTime ;
    gmeow:eventTemporalFrame gmeow:temporalFrameUTCGregorian .

# --- The claims: reified statements observed from the MODEL's vantage.
# (The RDF 1.2 source authors these as statement annotations; the owl:Axiom
# form here is the generated downcast idiom, P3.)
ex:east-wing gmeow:validUntil "2026-12-31T22:00:00Z"^^xsd:dateTime .

ex:stmt-2200 a owl:Axiom ;
    owl:annotatedSource   ex:east-wing ;
    owl:annotatedProperty gmeow:validUntil ;
    owl:annotatedTarget   "2026-12-31T22:00:00Z"^^xsd:dateTime .

ex:stmt-2300 a owl:Axiom ;
    owl:annotatedSource   ex:east-wing ;
    owl:annotatedProperty gmeow:validUntil ;
    owl:annotatedTarget   "2026-12-31T23:00:00Z"^^xsd:dateTime .

# The GROUNDED claim: generated by the invocation, attributed to the model,
# pinned to the exact evidentiary bytes. ALSO remembered (MemoryItem role).
ex:claim-close-2200 a gmeow:StandpointClaim, gmeow:MemoryItem ;
    gmeow:vantage ex:assistant ;
    gmeow:observedFeature ex:stmt-2200 ;
    gmeow:observationMethod gmeow:methodLlmExtraction ;
    gmeow:claimModality gmeow:probable ;
    gmeow:wasGeneratedBy ex:invocation-19 ;
    gmeow:wasAttributedTo ex:assistant ;
    gmeow:groundedIn ex:span-close-2200 ;
    gmeow:memoryOf ex:assistant ;
    gmeow:memoryKind gmeow:memoryKindSemantic .

ex:span-close-2200 a gmeow:EvidenceSpan ;
    gmeow:spanOfChunk ex:chunk-042 ;
    gmeow:spanStart "310"^^xsd:nonNegativeInteger ;
    gmeow:spanEnd "366"^^xsd:nonNegativeInteger ;
    gmeow:supportPolarity gmeow:polaritySupports .

# The UNGROUNDED claim: no groundedIn span — the flagged hallucination,
# retained, never deleted (P10).
ex:claim-close-2300 a gmeow:StandpointClaim ;
    gmeow:vantage ex:assistant ;
    gmeow:observedFeature ex:stmt-2300 ;
    gmeow:observationMethod gmeow:methodLlmExtraction ;
    gmeow:claimModality gmeow:probable ;
    gmeow:wasGeneratedBy ex:invocation-19 ;
    gmeow:wasAttributedTo ex:assistant .

# --- The conflict, SURFACED and attributed — never resolved by rank (P9).
ex:contradiction-close a gmeow:Contradiction ;
    gmeow:contradictsClaim ex:claim-close-2200, ex:claim-close-2300 ;
    gmeow:contradictionKind gmeow:contradictionKindNumeric ;
    gmeow:detectedBy ex:auditor .

# --- The auditor's VERDICT about the hallucination: itself a claim from a
# vantage, using the standpoint slice's bullshit modality (Frankfurt:
# asserted without regard to truth) and the deception slice's veridicality.
ex:auditor a gmeow:SoftwareAgent ; rdfs:label "grounding auditor"@en .

ex:stmt-2300-is-claimed a owl:Axiom ;
    owl:annotatedSource   ex:claim-close-2300 ;
    owl:annotatedProperty gmeow:claimModality ;
    owl:annotatedTarget   gmeow:probable .

ex:verdict-2300 a gmeow:StandpointClaim ;
    gmeow:vantage ex:auditor ;
    gmeow:observedFeature ex:stmt-2300-is-claimed ;
    gmeow:observationMethod gmeow:methodNliDerivation ;
    gmeow:claimModality gmeow:bullshit ;
    gmeow:claimVeridicality gmeow:veridicalityUntrue .

Terms

Classes

Properties

Individuals

Linkages

• Rows: 23
• Projection profiles: schema-org, web-annotation
• External vocabularies: cito, mls, oa, prov, rdf, schema, wd

Guide

How GMEOW already models AI claims

GMEOW did not need an AI claim model bolted on — the constitution's unified observation stance (P9) is one: "a measurement, a standpoint-indexed claim, and a sensory perception are the same reified construct (a claim made from a vantage)." This slice mints only the genuinely missing pieces (the retrieval segment, the prompt artifact, the invocation activity, the model card, the surfaced contradiction, the memory role) and documents how the EXISTING machinery does the rest. Each pattern below is existing-terms-only unless marked.

1. A generated claim is a query, not a class

An LLM output is a gmeow:StandpointClaim (⊑ gmeow:Observation) whose gmeow:vantage is the model gmeow:SoftwareAgent and which gmeow:wasGeneratedBy its gmeow:ModelInvocation (new) — a direct triple on the claim individual. On the reified RDF 1.2 statement form, confidence / accordingTo / the clocks ride statement annotations, and the invocation link is gmeow:mappedFrom (the machine-derivation AnnotationProperty), because axiom annotations stay AnnotationProperty-clean for the DL downcast (P3) — see dsl/statements/ai.ttl. "GeneratedClaim" needs no class: it is the query claims whose vantage is a SoftwareAgent and which were generated by an invocation. The same claim re-asserted by a person sheds nothing but that provenance.

2. Hallucination: flagged, never deleted — and named honestly

A claim node with no gmeow:groundedIn span (new) is a flagged hallucination (P10: retained, surfaced by EvidenceSpan audit machinery's audit gates). The VERDICT about it is itself a claim from a vantage:

• the auditor's gmeow:claimModality gmeow:bullshit — Frankfurt's category is already in the standpoint slice: asserted without regard to truth;
• or gmeow:claimVeridicality gmeow:veridicalityUntrue (deception slice) when the standpoint holds it settled-false;
• declared fiction/roleplay output is gmeow:veridicalityLicensedFalsehood — the licensed-falsehood safety property separates harmless fictional assertion from deception, for model outputs exactly as for human ones.

3. Evaluation is the norms extension's Assessment

gmeow:Assessment ⊑ gmeow:Observation (norms): an LLM judge is just a vantage; two judges disagreeing are two coexisting cells with no winner. The RAGAS/AIS metric family ships as a published gmeow:Rubric with Criterion individuals in the eval suite — instance data, never TBox. This core slice deliberately mints NO evaluation machinery.

4. Hallucination as hazard (risk extension, instance data)

gmeow:Hazard borne by a deployment, gmeow:manifestedAsType a hallucination event type, grounding gates as gmeow:Mitigation — zero new TBox; see the ai-normative fixture.

5. System prompts and personas (norms extension, doctrine)

A system gmeow:Prompt (new) enacting a voice is the norms extension's gmeow:Persona (registers, activation Condition, StyleGuide) applied at prompt-assembly time. The seam stays instance-level until the projection that assembles prompts from personas lands extension-side.

6. Memory is claims, not vectors

gmeow:MemoryItem (new gufo:Role ⊑ Observation): any claim can be remembered; revision is supersession + gmeow:displayable false (P10); salience rides the EXISTING statement-level gmeow:importanceLevel; recency rides the statement clocks. Recall is a retrieval (extensions/graphrag's RetrievalEvent). This is the substrate under the gmeow client and the MCP store/recall/revise triad.

7. Teleology

Agents have gmeow:hasGoal; an gmeow:Intention gmeow:motivates a gmeow:ModelInvocation (an Activity ⊑ Event) — whose reading of the motive it is rides accordingTo on the statement, like every attributed-motive claim.

The spine (EvidenceSpan audit machinery preview)

source ─(chunkOf)─ Chunk ─(spanOfChunk)─ EvidenceSpan ─(groundedIn⁻¹, polarity)─ claim node
                                                       (StandpointClaim, vantage = model,
                                                        wasGeneratedBy = ModelInvocation)

Pipeline observability around this spine — corpus, embeddings, indexes, retrievals, the GraphRAG entity graph — is extensions/graphrag (consumer: Project Lillith).

Terms

The genuinely-new terms this slice mints, anchored to the doctrine above.

gmeow:Chunk · gmeow:chunkOf · gmeow:spanStart · gmeow:spanEnd

The retrieval segment: a gmeow:Chunk is a contiguous slice of a source information object, located in it by gmeow:chunkOf (functional, ⊑ partOf) plus the integer offsets gmeow:spanStart/gmeow:spanEnd (zero-based, half-open code points). The one genuinely missing information-object kind in the source → Chunk → EvidenceSpan → claim spine.

gmeow:spanOfChunk · gmeow:groundedIn · gmeow:supportPolarity · gmeow:SupportPolarity

The evidence seam: gmeow:spanOfChunk pins an EvidenceSpan into the retrieved chunk; gmeow:groundedIn (domain-free, per the unified observation stance) ties any claim node to a supporting span, and a claim with none is a flagged hallucination (P10, never deleted); gmeow:supportPolarity records the NLI/AIS stance from the closed three-value gmeow:SupportPolarity (supports / refutes / neutral).

gmeow:ModelInvocation · gmeow:usedModel · gmeow:hasPrompt · gmeow:samplingTemperature · gmeow:samplingTopP · gmeow:samplingMaxTokens

The model-invocation activity, provenance anchor of every generated claim: one call to a model agent (gmeow:usedModel), carrying its prompts (gmeow:hasPrompt) and sampling regime (gmeow:samplingTemperature, gmeow:samplingTopP, gmeow:samplingMaxTokens). Outputs hang off the existing gmeow:wasGeneratedBy; no forward output property is minted (P5).

gmeow:Prompt · gmeow:PromptTemplate · gmeow:filledFrom · gmeow:promptRole · gmeow:PromptRole

The prompt artifact: a gmeow:Prompt is the concrete instruction presented to a model — content-addressed and role-tagged (gmeow:promptRole, drawn from the open gmeow:PromptRole vocabulary of system/user/assistant/tool). A gmeow:PromptTemplate is the reusable, versioned, variable-bearing source a concrete prompt links back to via gmeow:filledFrom.

gmeow:ModelCard · gmeow:describesModel · gmeow:modelProvider · gmeow:modelVersionTag · gmeow:modelContextWindow · gmeow:modelTrainingCutoff

The model card: a structured document gmeow:describesModel the SoftwareAgent, carrying gmeow:modelProvider, gmeow:modelVersionTag, gmeow:modelContextWindow, and gmeow:modelTrainingCutoff. The card is about the agent; capability claims stay attributed, contestable claim-layer observations.

gmeow:Contradiction · gmeow:contradictsClaim · gmeow:detectedBy · gmeow:contradictionKind · gmeow:ContradictionKind

The surfaced contradiction: a relator binding two or more claim nodes in conflict (gmeow:contradictsClaim), attributed to its detector (gmeow:detectedBy) and kind-classified (gmeow:contradictionKind over the open gmeow:ContradictionKind — factual / temporal / numeric / framing). It surfaces conflict, never resolves it by rank (P9).

gmeow:MemoryItem · gmeow:memoryOf · gmeow:memoryKind · gmeow:MemoryKind

The memory role: a gufo:Role on the universal claim construct — any claim can be remembered. gmeow:memoryOf names the remembering agent and gmeow:memoryKind its cognitive register (open gmeow:MemoryKind — episodic / semantic / working / procedural). Revision is supersession + gmeow:displayable false (P10), never deletion.