<![CDATA[This paper presents SENTAKU (Sistem Elektronik Pertandingan Kenshi Unggul), a referee analytics platform for Shorinji Kempo built on a PHP-based competition management system. The core contribution is a structured referee analytics framework that (i) computes seven per-referee statistical indicators from the raw embu score matrix, (ii) submits them with a structured prompt to Llama 3.3 70B via the Groq API, and (iii) audits the generated report through an objective content audit of coverage, numerical grounding, and section completeness. Two research questions are addressed: whether the framework surfaces referee patterns invisible to manual workflows, and whether a zero-shot LLM with structured context produces text faithful to the underlying statistics. Deployment at POPDA DIY 2026 (5 contingents, 53 randori athletes, 10 judges, 111 matches) surfaced component-level calibration differences and same-regency scoring patterns unattainable through paper workflows. However, the LLM report named only six of ten judges, cited no numerical value from the prompt, and omitted two of five requested sections, while introducing no fabricated content. These findings indicate that absence of hallucination alone is insufficient to guarantee faithful LLM reporting: zero-shot prompting can fail through systematic under-utilisation of numerical input even when explicitly required. Schema-enforced decoding is motivated as the next step]]>
Copyrights © 2026
