About iso22989

A LinkML schema for ISO/IEC 22989:2022 (Information technology — Artificial intelligence — Artificial intelligence concepts and terminology), with a curated suite of SSSOM mappings to neighbouring AI, security, privacy and upper-ontology schemas.

The public schema ships with paraphrased descriptions only. The normative text of ISO/IEC 22989:2022 is © ISO/IEC and is not reproduced anywhere in this repository. License holders may maintain a private verbatim-text overlay — see OVERLAY.md.

Schema

Element	Count
Classes	85
Slots	105
Enums	42
Custom types	2
Subsets	9

Coverage spans the full ISO/IEC 22989:2022 vocabulary:

Core AI concepts — AISystem, AIModel, AIComponent, Dataset, InferenceEngine, AIApplication, AILifecycleProcess, Task, Prediction, Decision, Action.
Stakeholders — AIStakeholderRole plus specialised roles (AIProvider, AIProducer, AICustomer, AIPartner, AISubject, RelevantAuthority, AIPlatformProvider, AIServiceProductProvider, ModelDesigner, ModelImplementer, ComputationVerifier, ModelVerifier, AIUser, AISystemIntegrator, DataProvider, AIAuditor, AIEvaluator, DataSubject, PolicyMaker, Regulator).
Trustworthiness — TrustworthinessProperty with the full TrustworthinessPropertyType enumeration (robustness, reliability, resilience, controllability, explainability, predictability, transparency, fairness, bias mitigation, accountability, privacy, safety) and supporting BiasType, SocietalImpactCategory, JurisdictionalIssueType.
Lifecycle — AILifecycleStage, OECDLifecycleStage, OECDLifecycleMapping, VerificationValidationFramework, AutonomyAssessment, EvaluationMetric.
AI ecosystem — KnowledgeGraph, ExpertSystem, CognitiveComputingSystem, SemanticComputingSystem, NLPComponent, ComputerVisionFunction, Robot, IoTSystem, IoTDevice, CyberPhysicalSystem, HumanMachineTeam.
Data — InputData, DataProcess, DataSample, DataLabel, GroundTruthRecord.
Terminology — AbbreviatedTerm plus seven Term specialisations (AITerm, DataTerm, MachineLearningTerm, NeuralNetworkTerm, TrustworthinessTerm, NLPTerm, ComputerVisionTerm) anchoring every ISO/IEC 22989:2022 Clause 3 sub-clause.
Annex SL anchors — Organization, InterestedParty for cross-management-system reuse.
Risk — RiskItem for AI-risk records carried into ISO/IEC 23894 / ISO/IEC 42001 risk programmes.

AIConceptsCollection is the containment root for serialising full inventories.

Cross-framework mappings (SSSOM)

Nine SSSOM/TSV mapping sets are published under src/iso22989/mappings/. All follow the 10-column SSSOM convention with embedded YAML metadata, PascalCase class CURIEs and <EnumName>#pv_snake_case PV CURIEs. Mapping justification is semapv:LLMBasedMatching pending expert review.

NIST AI Risk Management Framework

File	Target	Rows
`iso22989-to-nist-ai-rmf-common.sssom.tsv`	NIST AI RMF common module (trustworthiness characteristics)	14
`iso22989-to-nist-ai-100-1.sssom.tsv`	NIST AI 100-1 (RMF 1.0 core)	24
`iso22989-to-nist-ai-600-1.sssom.tsv`	NIST AI 600-1 (Generative AI Profile)	21
`iso22989-to-merged-nist-ai-rmf.sssom.tsv`	Consolidated NIST AI RMF rollup	13

ISO sibling standards

File	Target	Rows
`iso22989-to-iso42001.sssom.tsv`	ISO/IEC 42001:2023 AI management system	4
`iso22989-to-iso27001.sssom.tsv`	ISO/IEC 27001:2022 ISMS (Annex SL + risk + privacy)	8
`iso22989-to-iso29100.sssom.tsv`	ISO/IEC 29100:2011 privacy framework	13

Upper ontologies

File	Target	Rows
`iso22989-to-gist.sssom.tsv`	gist minimal upper ontology (Semantic Arts)	10
`iso22989-to-uco-core.sssom.tsv`	Unified Cyber Ontology (UCO) Core	4

Total: 111 mapping rows across 9 mapping sets.

Verbatim-text overlay

License-holders of ISO/IEC 22989:2022 may locally swap the public paraphrased descriptions for the verbatim normative wording using a deep-merge overlay pipeline; the overlay file is git-ignored and never published. See OVERLAY.md.

Pipeline:

just create-empty-overlay — regenerate the empty scaffold iso22989-overlay.template.yaml (committed; contains no copyrighted text).
just overlay-licensed-text — merge the populated overlay into tmp/iso22989-merged.yaml for local use by downstream generators.

Testing

The full gate — just test (schema generation → pytest → example generation) — is green: 90 passing tests, 0 failures.

tests/data/valid/ contains 34 example fixtures. Every fixture is loaded through the generated Python data model and validated against the schema. Coverage spans all major classes plus the constraint-bearing ones — required-slot classes (ResourcePool, IoTDevice, DataProcess, GroundTruthRecord, DataLabel, SoftComputingSystem, EvaluationMetric, Threshold, OECDLifecycleMapping, VerificationValidationFramework, AbbreviationEntry), an ifabsent sub-role (AIProvider), inlined-list containment (IoTSystem, AIConceptsCollection) and object references (Robot, KnowledgeGraph).
tests/data/invalid/ contains 10 counter-examples, each failing for a single documented reason and together exercising every enforced constraint type: invalid enum value, missing required slot, minimum_value, maximum_value (custom type), pattern, and wrong scalar type. Each file carries a header comment naming the violation it triggers.
The harness (tests/test_data.py) adds structural guards so the data-driven tests can never pass vacuously: the schema must parse via SchemaView, both corpora must be non-empty, and every fixture's file-name stem must name a concrete (non-abstract) schema class.

Schema fixes landed alongside the tests

Removed a duplicate tree_root (only AIConceptsCollection is the serialisation root; AISystem is no longer a competing root).
Gave the ConfidenceScore custom type an explicit base: float / uri: xsd:float, clearing shaclgen/sqltablegen "unknown range base" errors while preserving its [0.0, 1.0] bounds.

Known upstream limitation

ifabsent defaults are applied by the Python loader but not by JSON-Schema validation, so a required slot with an ifabsent default must still be stated explicitly in validated fixtures (e.g. AIProvider). This issue was raised updstream.