AI model, artifact and training
Support artifact identification, lineage review and authorized reuse by connecting complete model packages to training evidence, reproducibility assessments and reviewed use restrictions.
Bundle → Layer → Finding → Questions Filled
2 bundles · 3 layers · 5 findings · 14 questions
Model and artifact Describe model definitions and the complete scope of artifact versions.
Definition and versions
Distinguish model lineage from complete artifact identity.
Model definition
Identify the model family and its registered versions.
- Which architecture and model family does this model record describe?
- Which checkpoint versions does this model record register?
Artifact identity
Establish artifact identity through a complete component manifest and integrity evidence.
- Which versioned manifest defines this artifact's weights, architecture, configuration, tokenizers and other required components?
- Which digests and verification evidence identify this artifact's manifest and constituent components?
- Where are this artifact's components stored under the recorded access controls?
Training and provenance Record training evidence, resource authorization and reviewed restrictions.
Training run
Describe available training evidence without assuming complete provenance.
Pinned inputs
Record input provenance and the limits of training reproducibility.
- Which base checkpoint and dataset versions does this run record identify, with reasons for any unknown or inapplicable entries?
- Which code and configuration versions does this run record identify?
- What reproducibility assessment does this run record contain for artifact availability, execution environment, randomness, nondeterministic operations and result acceptance criteria?
- What training provenance is available for this artifact, with reasons for unavailable information?
Compute allocation
Record the resources and authorization assigned to a training run.
- Which compute resources and budget does this allocation record assign to its training run?
- Who approved this compute allocation?
Data restrictions
Separate source restrictions from reviewed artifact obligations.
Inherited restrictions
Document source terms and the basis for determining their effect on artifact use.
- Which training-data licences and use restrictions are documented in this restriction assessment?
- What reviewed determination does this restriction assessment record about terms applicable to the resulting artifact and intended use?
- What valid legal basis or additional permission does this release assessment document for any proposed use outside otherwise applicable licence terms?
Classifiers Filled
- Family
- Enterprise profiles
- Category
- Enterprise subject
- Entry kind
- subject
- Domain
- EnterpriseAI
- Industry
- Cross-industry
- Tags
- EM-AI-02W3subjectAIModelModelArtifactTrainingRunTrainingConfigurationComputeAllocation
What it is Filled
Model definition, artifact versions, training/fine-tune run, configuration, compute and data provenance. Weights are stored outside the public catalog.
Why it exists Filled
Support artifact identification, lineage review and authorized reuse by connecting complete model packages to training evidence, reproducibility assessments and reviewed use restrictions.
Distinguishing features Filled
- A complete model artifact is identified by a versioned manifest covering weights, architecture, configuration, tokenizers and other required components.
- Matching weight hashes establish weight equality but do not establish equality of complete model artifacts.
- A training run records its applicable base checkpoint, data versions, code and configuration separately from its resulting artifact.
- The model artifact record describes a reusable component rather than its deployment and use.
- A compute grant is a resource allocation for a run rather than an employment or service relationship.
- Recorded training-data restrictions are separate from a reviewed determination of restrictions applicable to the resulting artifact.
What robots and AI may and may not do Filled
Must not
- Treat an endpoint change alone as evidence of a newly trained model.
- Treat identical weights as proof that complete model artifacts are identical.
- Discard source restrictions or reviewed artifact obligations when an artifact is fine-tuned or copied.
- Publish model weights or restricted training data into a public catalogue.
- Overwrite an artifact version so that earlier evaluations lose their reference.
- Claim reproducible training solely because inputs and configuration were pinned.
- Release an artifact outside applicable licence terms without a documented valid legal basis or additional permission.
Only with a human decision
- A qualified reviewer must determine which recorded source restrictions apply to an artifact and its intended use.
- An authorized reviewer must verify the legal basis or additional permission before approving a release outside otherwise applicable licence terms.
- An authorized reviewer must approve a training run using personal or restricted data.
- An authorized reviewer must approve publication of model weights outside the organization.
May
- Record training runs with their available input provenance, configuration and resulting artifact manifest.
- Report recorded training-data restrictions separately from reviewed restrictions applicable to an artifact.
- Flag artifacts that share a name but differ in their manifests or required components.
- Record base checkpoints as inapplicable for training from scratch and unavailable training provenance as unknown with an explanation.
Moral aspects Filled
- Training data may contain personal data or copyrighted material whose use must stay within its terms.
- Lost provenance makes it impossible to answer for what a model learned and from whom.
Who is affected
- Data subjects whose data was used in training
- Rights holders of training data
- Teams that build on the model
Owners Filled
Steward
The head of AI or research who answers for models and their evidence on behalf of the organisation.
Master systems
- Model registry
- Experiment tracker
- Evaluation store
Links to other meta-models Filled
neighbor
- EM-AI-01
- EM-AI-03
- EM-DAT-01
- EM-FIN-05
references
- WM-SFT-004 - conceptual-candidate
- WM-AI-007 - conceptual-candidate
- WM-AI-006 - conceptual-candidate
What else AI and robots need to interact with it Filled
Identity and identifiers required Filled
- A model has a stable identifier and version, while each complete artifact has a versioned component manifest with integrity digests.
- Artifact identity includes weights, architecture, configuration, tokenizers and other required components.
Direct properties not applicable Not applicable
Not applicable
Enterprise record contour: physical properties belong to referenced world models.
Recognition required Filled
- An artifact record identifies its component manifest and the availability of training provenance.
- A base checkpoint may be inapplicable for training from scratch, and training provenance may be unknown for an externally obtained artifact.
- The artifact is distinguished from its serving endpoint and the deployed AI system.
Capabilities and actions required Filled
- Available evidence supports lineage tracing across base checkpoints and derived artifacts.
- Training reproducibility requires an assessment of available artifacts, inputs, code, configuration, execution environment, randomness and nondeterministic operations.
- A reproducibility assessment specifies whether exact equality or a defined tolerance is supported by evidence.
Hazards and failure modes required Filled
- Required components can change silently under an unchanged model name.
- Identical weights can conceal differences in architecture, configuration or tokenizers.
- Lost source terms or unsupported inheritance assumptions can produce incorrect artifact use decisions.
- Restricted weights or training data can leak through the catalogue.
Standards and interfaces optional Filled
- PROV-O from the World Wide Web Consortium supports provenance of training runs and artifacts.
- SPDX from the Linux Foundation supports licence and provenance information.
- The AI Risk Management Framework from the National Institute of Standards and Technology supports documentation of model context and risks.
Context of use required Filled
- The record supports organizations that train, fine-tune, obtain or store AI models.
- Documentation obligations depend on jurisdiction, applicable model classification and the responsible actor's role.
Sources Filled
- NIST AI RMF: context, evaluation and risk management
- PROV/SPDX: provenance of data, artifacts and actions
- Model registry/evaluation tooling practice; map system, artifact, run and endpoint
- NIST AI Risk Management Framework - Checked live (HTTP 200) on 2026-10-07.
- W3C PROV-DM: The PROV Data Model - Checked live (HTTP 200) on 2026-10-07.
- SPDX Specification - Checked live (HTTP 200) on 2026-10-07.
Open questions
- How to distinguish the architecture, a checkpoint and a deploy?
- Which inputs does reproducibility require?
- How to reflect access and use restrictions on training data?
- Установить границу и решение reuse/extend/new по действующим спецификациям.
- Подтвердить semantic crosswalk, права и source mastership.
- Выбрать immutable refs; провести проверки fixtures до заявления о публикационной готовности.
Machine files
Provenance
enterprise research programme · published-partial
Built from: enterprise/models/em-ai-02/brief.json, ver-cy/world-models/card-supplements/em-ai-02.json
Still in Russian: suggested owner, blocking decisions, vercy candidates. Translate in research/enterprise/i18n/units.en.json.