Quality-aware synthetic-data selection assumes that examples rated highly by an LLM judge will help downstream learning. We test this assumption in a controlled replay across two low-resource African-language classification tasks, four languages, and seven selectors. Six selectors retain 64 examples per task-language cell. CoSDA-V2 yields the highest judged label correctness (0.904 versus 0.767 for naive) and the lowest measured shortcut score, but AlpaGasus leads CoSDA-V2 on downstream Macro-F1 in both the main replay (0.202 versus 0.163) and the three-seed refit (0.193 versus 0.157).

Adding the audit’s strict gate to AlpaGasus reduces Macro-F1 by 0.029, driven by one cell where only 31 of 64 candidates survive. Because the gate changes pool composition and training-set size, this comparison measures their joint effect. LESS also falls below naive when unconstrained top-k collapses label balance, but recovers with quotas. At this budget, retention constraints account for more of the observed differences than a consistent ranking advantage. In this replay, audit scores do not reliably predict downstream utility. We therefore recommend reporting both metrics for the same retained sets.

For very low-resource language pairs such as Bahnar-Vietnamese, the relative effectiveness of generic multilingual, language-pair-specific lexical, and task-adapted retrieval remains unclear. We study Bahnar-Vietnamese parallel-sentence retrieval (bitext retrieval) as a controlled candidate-ranking task and compare lexical, alignment-based, neural, and hybrid methods under a development-based selection protocol. The strongest off-the-shelf encoder, LaBSE, reaches 24.24% Recall@1, while symmetric IBM1 reaches 49.33%. Among four in-domain modern dense retrievers, development selection chooses DPR-XM, which reaches 47.43% Recall@1 and 72.21% Recall@10. XLM-R LoRA reaches 64.02% Recall@10, and the development-selected XLM-R-first IBM1 reranker gives the highest Recall@1 point estimate, at 51.62%, with an MRR of 0.5579. Paired tests do not establish a significant difference between the hybrid and IBM1 at the 0.05 level, although reranking substantially improves over its neural first stage. Under pair-specific training, LoRA projection with CSLS also performs strongly on Khmer-Vietnamese and Lao-Vietnamese, reaching 80.14% and 84.58% Recall@1. Overall, the results show complementary strengths between lexical alignment and lightweight neural adaptation in low-resource bitext retrieval. Data and code are available at https://github.com/NgThTy/translation-vn-bahna.