Self-Preference Bias in Rubric-Based Evaluation of Large Language Models

This project identifies a specific nuance in the 'LLM-as-a-judge' paradigm: that bias persists even when using structured rubrics rather than just pairwise comparisons. While the insight is academically valuable, the project currently lacks a moat. With 0 stars and 3 forks at 9 days old, it is effectively a fresh research release. The defensibility is low because once the methodology for detecting self-preference in rubrics is published, it becomes a commodity metric that evaluation platforms (like LangSmith, Arize Phoenix, or WhyLabs) can implement in a weekend. Frontier labs (OpenAI, Anthropic) have a high interest in this space as they rely on self-critique and recursive improvement loops; they are likely already building internal mitigations for this exact bias. The 'displacement horizon' is short (6 months) because the field of LLM evaluation is moving at extreme velocity, and this specific finding will likely be absorbed into larger meta-evaluation frameworks quickly.