Its search unit is an item-hypothesis pair: a candidate and a natural-language explanation of why it fits a decision maker. The system maintains an archive, repeatedly selects an item, generates a hypothesis from earlier entries, and decides whether to retain it.
QD-Rec uses a contextual bandit to choose items, with an additional diversity bonus. A proposer model can mutate, combine, or explore ideas from three archive examples. An LLM evaluator retains a proposal if it is novel relative to nearby archive entries; otherwise it uses a blind pairwise comparison against its closest neighbor.
The reported tests include acquisition-target search across about 6,300 US companies for Walt Disney and JPMorgan Chase, marketing over 10,000 fictional customer profiles, and paper recommendation over 17,500 recent conference papers. The company reports higher archive scores in some M&A and marketing tests, but little difference in the paper task. The source does not establish performance in real acquisition, marketing, or research workflows: evaluations rely on language-model judges, and part of the evaluation uses fictional profiles.