Tools that claim to detect emerging research topics are rarely judged on their own lists. Our main contribution is a protocol that makes such judgement possible. A bet freezes four things before its verdict can be known: the date of the photograph, the population of candidates, the verdict rule, and the comparison with chance drawn among the same candidates. The list has a fixed length, is read on two axes (champions and failures) against a computed ceiling, and is sealed by cryptographic fingerprint; sealing proves that a list existed at a date, not that it is right. A topic is a champion if its yearly volume of records triples within three years and reaches 100, a failure if it falls by 30%. We apply the protocol to one detector on PubMed/MEDLINE, in three separate blocks. On a retrospective benchmark (2017-2022), chance among the candidates yields 26.7 champions and 15.1 failures out of 120 slots; the detector yields 85 and 3, and under nine pre-registered variants of the verdict rule its champions stay between 2.5 and 5.0 times chance, with no more failures. The detector was developed with these verdicts known, and its specification is not published. Prospective validation rests on six sealed bets, judged in early 2029 and early 2030; only bets 5 and 6 test exactly the instrument of the benchmark. A secondary, exploratory study applies the same judge to eight research funders: under exact name matching, which may disadvantage grant texts, the detector's lists are more concentrated in future champions than the twenty topics named in the most grants of each funder (70.8% against 20.8% to 35.8%, chance 22.3%). We document data defects, negative results and the limits of each comparison.

翻译：暂无翻译