TermGuard combines neural and symbolic AI. An LLM drafts a definition guided by the ISO 1087 evaluation grid; a deterministic rule engine with packs for 16 EU languages scores it and explains every failed rule; the model then redrafts against that feedback. A separate semantic check tests whether the definition can replace the term in a real sentence without changing its meaning.
The loop sits inside a terminology workbench: upload a corpus, extract term candidates, draft and validate definitions, extract concept relations, ground terms in IATE and EuroVoc, and export to TBX (ISO 30042). Validated terms then constrain machine translation, with an automatic check that the approved terminology was used.
The system is provider-agnostic: any LLM can be the backbone, including European models. In our benchmark across six models and 90 runs, structural validity rose from 56% at first draft to 98% after the loop, in about two iterations. It is deployed as a containerised service on EU cloud.
