Bias Resistant Human Aligned Judging

Use this skill when evaluating if an AI judge exhibits stable, objective, and logically sound reasoning, particularly against situational and statistical biases. Trigger it when requests mention 'test if the judge is tricked by a sob story', 'check if the evaluator is swayed by flattery', 'verify if the model is too sure too early based on vague clues', 'check if confidence increases just because the chat got longer', or 'test if the judge is just copying the reference answer'. It is designed for meta-evaluation, detecting susceptibility to rhetorical persuasion, 'Length Artifacts' (monotonicity bias), Criteria Entanglement (halo effect), and 'Solution Fixation' induced by references.

dingxingdi Updated 1 repo stars

File contents

dingxingdi/paper_fast_search_backup commit e1c1dfd293

Frequently asked questions

npx get-skill add dingxingdi/bias-resistant-human-aligned-judging-3