Methodology

Four-layer model

Every concept is assigned to exactly one of four enterprise-architecture layers:

Layer	Question	Examples
Strategy	Why?	Objectives, drivers, KPIs, principles, policies
Process	What activities? When?	CAD Design, BOM Management, Change Management (ECO/ECN)
Logical	What capability/data? (vendor-neutral)	CAD 3D Design, Part Master, Digital Thread
System	Concrete implementation	Siemens NX, ISO 10303-242, STEP AP242, Siemens Teamcenter

Typed relations

Concepts are connected by a closed set of typed edges (implements, realizes, supports, uses, produces, conforms-to, part-of, synonym-of, etc.). Rules enforced in CI:

• implements — only System → Logical.
• realizes — only Process → Strategy.
• part-of and synonym-of must stay within one layer.
• Every relation has a defined inverse; the build step materializes missing back-edges.

Full relation list: see ontology/plm.ttl and pipeline/plm_pipeline/model/relations.py.

Sources

Each concept carries an explicit sources list in its frontmatter. Preferred sources, in order:

1. Wikidata (for stable identifiers; owl:sameAs linkage).
2. Wikipedia (definition + first paragraph, CC-BY-SA 4.0 attributed).
3. ISO / OMG / NIST (for standards and normative references).
4. arXiv / OpenAlex (for academic citations on deeper topics like MBSE, digital twin).

No vendor-site scraping. Vendor product pages are referenced by URL but not copied.

LLM-assisted enrichment

A Python pipeline drafts new stubs via the Claude API (Haiku 4.5 for bulk classification + normalization, Sonnet 4.6 for relation typing). Prompt caching keeps cost low. Every LLM output lands in a pull request for human review before merging to main.

Status lifecycle

• stub — pipeline-generated skeleton, unreviewed, not published in production.
• draft — has a definition and some relations; published, flagged as draft.
• reviewed — complete, sources cited, relations validated; canonical.