The reasoner's verdict for ONE candidate, judged independently against the source. A matched
set can mix true duplicates and false positives (vector search returns the top-K neighbors),
so each candidate gets its own recommendation/confidence/reasoning — the detector stamps the
per-candidate verdict onto that candidate's match row, never a blanket set-level verdict.
The reasoner's verdict for ONE candidate, judged independently against the source. A matched set can mix true duplicates and false positives (vector search returns the top-K neighbors), so each candidate gets its own recommendation/confidence/reasoning — the detector stamps the per-candidate verdict onto that candidate's match row, never a blanket set-level verdict.