Fast, confident, and sometimes wrong: a prespecified holdout evaluation of a small decision model as an LLM pre-filter

A confidence-gated cascade (Laya 0.3.6 multilingual checkpoint answers when its top probability is at least 0.90, otherwise phi4:14b) was evaluated on five public English and Polish text-classification tasks. The threshold was chosen on a 230-item pilot and the analysis plan was frozen before a 204-item holdout was built. On the holdout the gate accepted 124/204 items (60.8 % LLM-call avoidance); the cascade scored 181/204 vs 180/204 for a reconstructed phi4 implementation (paired difference +0.5 pp, bootstrap 95 % CI -2.5 to +3.9). All three prespecified criteria were met, but the comparison rests on 11 discordant items, and the 8 errors among accepted small-model answers bypassed escalation. The results support selective LLM-call avoidance on this benchmark mixture, not general reliability or protection against high-confidence errors. Package: code, analysis plan, per-item predictions and an item manifest (no item texts).