REF.01 own project · skillsintelligence.tools · 2026
A GenAI competency atlas with a full decision record
Growing a competency atlas from 269 to 426 entries without breaking a single identifier or deep link, with every contested classification decision on the record.
- L1Ontological layer426 competencies · 591 relations
- L2Dispute tribunal99 cases · confidence 0–3
- L3Separated market noise−474 false edges removed
- Competencies in atlas
- 269 → 426
- Ontology relations
- 435 → 591
- Spurious edges removed
- 474
- Entries in decision record
- 2,514
-
Problem
The GenAI and data science competency atlas lived in a single spreadsheet — 269 nodes, 435 relations, a dozen junk rows — and was falling behind the vocabulary the industry actually uses for these competencies. Extending it risked breaking identifiers and glossary links.
-
Approach
The 269 atlas nodes were matched against 2,218 candidates for inclusion. The main classification was done by a single model; 99 contested cases were settled by three-model voting with votes, reasoning and dissents recorded; each decision carried a 0–3 confidence score, and uncertain cases were flagged rather than guessed. The key design decision was separating market mappings out of the ontology into their own layer.
-
Outcome
The atlas grew to 426 competencies and 591 relations with zero duplicates and no gaps in required fields. Splitting the layers removed 474 spurious edges and cut dangling nodes from 45 to 29. The source file was untouched throughout the rebuild.
The thing that mattered most was not technical. Market mappings — how often a competency gets mentioned — sat in the same layer as ontological relations. It looked like a rich graph. It was two different things glued together: what a competency is, and how often people talk about it.
Separating the layers removed 474 edges that described no real conceptual relation. The graph got smaller, and only then did it become usable for inference.
What this engagement does not cover
157 newly added nodes have no prerequisite edges yet. A verification round over part of the decision set by independent models is planned, not done.