SIE.4:2 - Problem

Correspondence candidates are cheap. Labels can be normalized, lexical similarity scored, hierarchies compared, and model structures aligned. Yet “similar”, “related”, “narrower”, “same field name”, and “safe to substitute here” answer different questions. A row can be a true relation and still lose distinctions that the receiving use requires.

When candidate generation and qualification collapse, downstream mapping code inherits an unstated ontology and loss policy. Counterexamples are discarded as data-quality defects, source versions disappear, and a reviewer cannot tell whether the relation itself or only its use is unsupported.