← All questions
EvaluationHard· LLM Evaluation· Statistical Significance· Decision Making

An engineer wants to replace the majority-vote counter...

A three-point improvement on 500 questions is likely within the margin of error. Beyond the statistical weakness, replacing a deterministic counter with an LLM selector introduces new risks. LLM selectors are often sensitive to the order of candidates, which creates systematic errors that are difficult to debug. To validate this, you should perform a paired test with randomized candidate order and compare the performance against the original voting method at a matched call count.

Code playground (Python)