How often an LLM judge wavered, by how uncertain Jev was

Share of examples where at least one LLM judge changed its label across ten runs, with 95% Wilson intervals.

Jev's uncertainty is measured as entropy: 0 bits means all probability on one label, 1 bit means a coin flip. The middle groups hold 20 to 50 examples each, so treat small dips between them as noise. Hover a bar for counts.