A Certification Gate for Emerging-Topic Detection: Pre-Registration, Negative Controls, and a Second-Domain Transfer Test on arXiv

Methods for detecting emerging research topics keep multiplying, but they are rarely validated against ground truth chosen independently of the method: most are assessed on internal coherence or a few hand-picked success stories. This paper contributes not a new detector but a certification gate, a pre-registered, negative-controlled validation protocol any emergence detector can be run through. Before results are seen, the gate fixes (i) a binary "flagged rising" criterion, (ii) an externally-sourced positive set and a matched negative-control set, and (iii) Wilson-95% reporting of recall, specificity and precision. Applied to the temporal component of a five-dimension emergence system over arXiv stat.ML (2018–2023; 76 topics), the gate yields recall 93% (14/15, Wilson 95% [70,99]) and specificity 100% (13/13, [77,100]). The identical protocol is then run on a second, structurally different domain (arXiv cs.CR; 2018–2023; 129 topics) under a benchmark frozen and timestamped before scoring. This is a transfer test, not a replication: with only two domains it characterises how the protocol behaves out-of-corpus, not whether the first result reproduces. Recall falls to 62% (10/16, [39,82]) and specificity to 69% (9/13, [42,87]); pooled, recall is 77% (24/31, [60,89]) and specificity 85% (22/26, [66,94]). The intervals are wide by construction: at these n the benchmark is underpowered to rank detectors and demonstrates only the gate. The honest degradation is the point: a validation that reported 100% on one corpus and hid the second would be the overfitting it claims to detect. The protocol, both benchmarks with timestamp proofs, and a disclosed pre-scoring amendment are released, and the gate's numbers are reported with and without that amendment.