Personas judge the same archived pairs, no evolution. Does the judgement depend on who the model is asked to be?
Hub and runs: E01 Persona judge study.


Setup

  • Gemma 4 12B
  • Neutral = Miyazaki’s prompt unchanged
  • Persona replaces the two role lines; rest of the prompt unchanged
  • Paraphrase of the role lines = noise floor; a persona effect must beat it

Findings

  • Personas change the stated criteria clearly (child: smoothness)
  • Winners
    • E01.2/E01.3 (dynamically, weirdly): flips at the paraphrase floor
    • E01.5 adorably: stories change winners above the floor (+0.18)
  • Gemma 4 31B: more consistent, same winners as 12B
  • Adorably chosen as the second term for E02.2
  • E01.3b running: do stories of the same type agree more?