Abstract
Small decision models choose among prompt options; this becomes slow and less accurate with hundreds of options, as in the Common Weakness Enumeration (CWE). We study Kev-Index, a configuration of the small decision model Kev in which a second pointer head on the same adapter scores cached option keys against the encoded record, keeps the top k, and lets the decide head choose among them. We compare original Kev, fine-tuned Kev with all options and Kev-Index on mapping Common Vulnerabilities and Exposures (CVE) records to CWE classes (205 options), MITRE ATT&CK techniques (51) and Common Vulnerability Scoring System (CVSS) severity (4), with one shared adapter, misses counted as errors and latency timed on one GPU. Fine-tuning raises CWE accuracy from 0.696 to 0.832, micro-F1 on ATT&CK from 0.328 to 0.610 and severity accuracy from 0.243 to 0.570. On CWE mapping, Kev-Index at k = 10 reaches 0.822 against 0.832 for a reference trained on fewer rows (paired difference −1.0 points, 95% interval [−1.8, −0.1], per-seed −0.2 to −1.8; inside a 2-point margin set on development data, but excluding seed variance) at 35.9 ms per record against 116 ms (3.2 times faster; eager mode, batch size 1). With 51 and 4 options Kev-Index is slower; on ATT&CK it is also less accurate; no k passed the pre-set rule. Limitations: three seeds, one many-option task, a larger index-head training pool (favouring Kev-Index), no training on shortlist prompts, no simple baseline on Tasks B and C, and no retrieve-then-rerank, encoder or caching baseline.