Stable identifiers
Concept IDs remain addressable as embedding models, vector dimensions, and retrieval techniques change.
About the project
Semantic identity should outlive the model that found it.
Embedded Semantics keeps stable ConceptCodes, definitions, reviewed language evidence, and registry versions authoritative while treating vectors and model scores as replaceable evidence.
Multilingual evidence
Language expressions attach to concepts with locale, review status, equivalence strength, and provenance.
Explicit abstention
The current exact resolver returns an unknown or ambiguity result when governed evidence does not justify one authoritative ConceptCode.
Inspectable evolution
Experimental retrieval, prototypes, model profiles, and evaluation evidence can change without redefining the stable concept identity they support.
Current capability
What works today?
Production resolution is registry-first and exact. Reviewed multilingual expressions can resolve to the same stable ConceptCode, exact ambiguity abstains, and unreviewed unseen expressions return unknown_expression.
Arbitrary unseen-query semantic generalization is still experimental. Embedding retrieval and language-neutrality evaluation exist as research infrastructure, but they are not presented as a production capability.
Design boundary
What the registry does not claim
It does not claim that every culture partitions meaning identically, that a single vector captures every nuance, or that translation is lossless. The system preserves unresolved distinctions as metadata and evidence instead of silently treating them as equivalent.