Payment operations are a critical financial infrastructure, but the value of large language models in this domain remains unclear because payment rules change quickly, evidence is fragmented, and decisions depend on transaction state, participant role, region, and payment rail. Existing benchmarks do not isolate whether failures come from missing payment-rule knowledge, poor use of supplied evidence, or brittleness under imperfect harness inputs. We introduce BenchCompass, a payment-domain benchmark whose construction pipeline builds scenario-grounded tasks from typed evidence packs, applies LLM-based quality checks, creates task-input attack variants, and reserves final item admission for domain experts. The release contains an expert-reviewed Pro benchmark covering payment knowledge, context-grounded scenario reasoning, and Attacked Open robustness, plus a lower-assurance Normal pool for inspection and future curation. Across 16 model variants, BenchCompass shows qualitatively different failure modes: missing parametric payment knowledge, incomplete reasoning over supplied rules, and failure to reject plausible but invalid workflows. The benchmark remains unsaturated: the best frontier model reaches 89.6% on Open Context-Grounded Reasoning and 81.7% under attacked inputs, while a representative 32B open-weight model reaches 69.8% and 42.6%. The benchmark, data, and code are publicly available at https://github.com/ant-intl/BenchCompass.

Manchu is a typical low-resource agglutinative language, for which existing digital resources are largely limited to character recognition and basic annotation, while sentence-level generation remains underexplored. In this paper, we study constrained sentence-level Manchu continuation. We construct the first context-target corpus for this continuation task, consisting of 3214 examples collected from Manchu Stories in Translation, Common Manchu Dialogues, and other sources, covering both narrative and conversational genres. To improve robustness to rich suffixation, relatively flexible word order, and frequent OCR or transcription noise in Manchu, we propose a Multi-Granularity Data Augmentation (MGDA) strategy that expands training data through character-level noise injection, phrase rearrangement, and suffix perturbation, without relying on external dictionaries or translation models. Experiments on ByT5-small and Qwen2-1.5B show that MGDA substantially improves Manchu text generation performance: BLEU-4/ROUGE-L increases from 2.03/2.59 to 17.60/42.70 for ByT5, and from 1.98/6.42 to 14.12/41.49 for Qwen. Further ablation studies demonstrate the complementarity of the three augmentation granularities and suggest potential benefits of byte-level modeling in this experimental setting.