2026-11-12 meeting Reiji discussed life-story personas and cleaner contrasts. Child/engineer in E02.1 compared age/profession.


Pilot

  • Same setup as E02.1, but life-story personas
  • 20 generations
  • Persona conditions
    • child label
    • child life-story
    • adult label
    • adult life-story
  • Term: Beautifully
  • Data from pilot can be used in full E02.2 results

Full experiment

  • 20 generations
  • Personas
    • 2 child (child 3 and 4)
    • 2 adult
    • 2 artist
    • 2 engineer (engineer 3 and 4)
    • neutral
    • random
  • terms
    • Adorably
      • (chosen 2026-11-17 from E01.5 results)
    • Beautifully ?? might drop
  • runs per term: 12 conditions (personas) * 5 seeds (100–500) = 60 runs
  • reuse from E02.1 (20 generations, same seeds)
    • random: all 5, term-independent
    • neutral, beautifully
    • new runs: 50 beautifully + 55 adorably
    • beautifully ≈ 150 GPU-h
    • adorably ≈ 250 GPU-h (about two of three pairs rejected, so ~1.5× the calls)
      • ~400 GPU-h, ~8–9 days on blitzen + donner
  • type effect = difference between types > difference within a type (child-3 vs child-4)
  • noise reference: same persona, same generation 0, different GA random stream, 2–3 runs
    • needs a new evolve.py option (open in E02.1 results)
    • the judge is deterministic, so same persona + same seed = same run