Knowledge Graph Edge Validator

Decides whether an edge an LLM extracted from a document (subject, relation, object) may go into the knowledge graph. Code first checks that both entities appear in the document (an invented entity is rejected for free). Then one Jev Choice reads the document and the edge's meaning in words, and answers supports, contradicts or not_stated. The traps are the ones that poison graphs: a reversed direction, a past fact stored as current ("previously", "before that"), a declined offer, a rumour, a wrong relation. The edge is stored only when P(supports) ≥ 0.5 (a threshold sweep shows the tradeoff). Against it: storing every extracted edge, and the LLM checking its own extraction. Measured: accuracy, bad edges stored, good edges rejected.