Research & methodology
Measure meaning before claiming universality.
The project treats multilingual semantic alignment as an empirical retrieval problem with explicit failure cases. The current production resolver is exact and registry-governed; model-backed generalization is evaluated separately.
00 · Current boundary
Separate production evidence from research targets
Reviewed exact multilingual expressions can resolve in the production registry today. Arbitrary unseen-query retrieval, reranking, and confidence calibration remain experimental until their evidence gates are satisfied.
01 · Objective
Concept resolution, not generic similarity
The target research task is to identify which registry concept best represents an unseen expression. Topic similarity is useful evidence, but it is not equivalent to semantic identity.
02 · Training
Cross-language positives plus hard negatives
A future semantic model should bring equivalent expressions across languages into compatible neighborhoods while keeping closely related but incorrect concepts distinguishable. Evaluation therefore uses reviewed translations, paraphrases, definitions, and hard negatives.
03 · Evaluation
A benchmark that rewards discrimination
- Concept Recall@1 and Recall@5
- Mean reciprocal rank
- Cross-language concept agreement
- False-neighbor rate
- Centroid and language-family spread
- Top-1 / top-2 score margin
- Abstention precision and recall
04 · Abstention
Unknown is a valid result
The exact production resolver already abstains when governed exact evidence is absent or ambiguous. Any future model-backed resolver must preserve that behavior by declining to assign a concept when its evidence is weak or competing candidates are too close.
05 · Provenance
Every semantic claim should be traceable
Concept definitions, expressions, renderings, relationships, model profiles, prototype evidence, and evaluation outcomes are tracked independently so changes can be inspected without rewriting ConceptCode identity.