REF.02 own project · skillsintelligence.tools · 2026
An AI glossary reconciled across twenty-two models
414 entries covering 2023–2026, each with a maturity rating, a Polish equivalent and a place in the relation graph — shipped as an embeddable component.
- I–IINeologism → Working termEarly signal & research jargon
- III–IVIndustry term → Canon414 entries · 10 categories
- VRegulatory term + PL status104 relations · 5 settled PL
- Glossary entries
- 414
- Models across three rounds
- 22
- Maturity rating coverage
- 414/414
- Relations in concept graph
- 75 → 104
-
Problem
AI vocabulary drifts faster than glossaries get written, and a single model asked for "important concepts" returns its own bias: it overweights whatever is loud right now and misses the rest.
-
Approach
The glossary was built in three rounds involving 22 models. Each round received the list of entries already collected and was asked to look for gaps, and results were merged with an explicit canonicalisation trail — 18 duplicate groups merged, 10 category headers stripped out of the entry corpus. Third-round candidates were checked against web search: 131 of 228 were admitted. Every entry received a maturity rating on a five-point scale, from neologism to regulatory term.
-
Outcome
A 414-entry glossary across 10 categories with full coverage on both maturity rating and Polish-equivalent status, shipped in three distribution packages and wired in as the glossary layer of a public service.
The maturity rating turned out to matter more than the definitions. A term three months old and a term written into regulation look identical in a glossary — and mean very different things to someone planning to build on them.
The five-point scale — neologism, emerging, industry, established, regulatory — is filled in for all 414 entries in this set.
What this engagement does not cover
The glossary and the competency atlas were built as independent corpora; 13 direct cross-links out of 414 entries connect them so far, with full cross-mapping kept as a separate stage. Five entries carry a settled Polish equivalent; for the rest, the status field explicitly marks the term as new, borrowed in its English form, or not yet stabilised in Polish rather than coining artificial calques.