We present a humor generation system forSemEval-2026 Task 1, Subtask A (Castro et al.,2026) that produces short jokes under lexicalor headline-based constraints. For each input,our system generates multiple candidate jokesusing a large language model across diverse hu-mor styles and prompting strategies, includingzero-shot, few-shot, and structured prompting.Constraint satisfaction is explicitly enforced,either by requiring exact lexical inclusion orby approximating semantic relevance to a head-line using sentence-embedding similarity. Allvalid candidates are ranked using a weightedhumor score that combines semantic incon-gruity, emotion-based humor potential, ironylikelihood, linguistic fluency, and novelty withrespect to a large external jokes corpus, andthe single highest-scoring joke is selected foreach constraint. This approach follows a best-candidate selection paradigm, leveraging auto-mated humor proxies to improve joke qualitywithout task-specific fine-tuning.