This week: a solid technical exchange instead of a ship, but worth logging. Talked through JD Intelligence's quality-scoring approach with someone building a semantic-matching recruiting tool from scratch (embeddings + TF-IDF + explainability). Confirmed something we keep seeing building this ourselves: even good matching engines get noisy fast when the JD is vague, since the score has nothing precise to grade against. Doubling down on the JD-quality-as-upstream-input thesis. jdintelligence.dev