Source-Grounded Translation Cards for Low-Resource French–Wolof LLM Translation

Large language models (LLMs) remain unreliable for many low-resource African languages, while authoritative local lexical resources often encode distinctions that may be poorly represented in web-scale training data. We study whether a bilingual dictionary can be converted into source-grounded, retrieval-time guidance for French--Wolof translation without model fine-tuning. Starting from Diouf's 495-page Wolof--French/French--Wolof dictionary, we build 4,468 source-grounded translation cards containing 9,562 sense-, collocation-, and construction-oriented units linked to source entries. We compare three inference conditions---flat dictionary pairs, rich structured cards, and a bare model---across seven model runs and 125 French inputs (2,625 segment-level judgments). Rich cards yield the highest acceptance in six of seven runs and, averaged across models, improve acceptance by 11.4 percentage points and reduce critical errors by 14.5 points versus the bare model. The strongest configuration, Opus 4.7 with rich cards, reaches 91.2\% LLM-judge acceptance versus 76.0\% bare. Absolute performance remains strongly model-dependent; smaller models still struggle with Wolof grammar and composition. Because references and judgments are LLM-generated, we treat these results as preliminary pending native-speaker evaluation.