We present a generator for Thai klon pad verses that creates a four‑line stanza from a short topic. The model is a masked language model initialized from WangchanBERTa and operates on a fixed canvas with two layers: a sound layer that records pronounced syllables and enforces rhyme, tone and syllable count via hard masks, and a word layer that keeps whole words intact, preventing word fragmentation. Decoding fills one cell at a time, sampling each cell at temperature 0.4, producing a single sample per topic without any selection, reranking or repair.
Evaluation was performed on 78 held‑out topics using a frozen LLM judge that scored form, lexical cleanness and topic alignment in three independent runs. The generated stanzas achieved a combined pass rate of 70.5% across the three criteria, compared to 76.9% for human verses (difference –6.4 points, 95% CI [-19.2, +6.4]). Lexical cleanness was on par with humans (85.9% vs 87.2%), and the form‑plus‑cleanness pass rate was slightly higher (+3.8 points, CI [-6.4, +14.1]); the gap stemmed mainly from topic alignment (84.6% vs 91.0%).
Temperature proved crucial: higher temperatures reduced pass rates because low‑probability lexical choices introduced errors, while the low temperature sharpened the sampling distribution at the cost of vocabulary richness (7.5% rare words versus 16.5% in human verses) and diversity. The pre‑registered target—all three metrics within five points of human performance—was missed by 1.4 points. In contrast, the only publicly available klon pad generator, KlonSuphap‑LM, failed to pass any of the 78 topics (form 19.2%, cleanness 3.8%).
Review: This study shows that single‑sample low‑temperature decoding can approach human quality under strict formal constraints, especially in form and cleanness. However, topic alignment remains a weakness; improving semantic alignment may require richer contextual modeling or post‑processing techniques.