wassname 3d5c8345f2 add humanizer A/B test: minimal vs full vs baseline ...

Three-way comparison on 8 held-out writing scenarios:
- baseline (no humanizer) vs minimal (strongest tells + personality)
- baseline vs full (all 29 patterns + banned words + structural slop)
- minimal vs full (key comparison)

Results: minimal beats full on held-out outcomes (voice -0.3,
ai_suspicion -0.2 when adding full pattern catalog). The 48KB skill
hurts compared to the 2KB minimal version. More AI writing involvement
= more detectable, as the skill itself warns.

Rubric uses outcome-level dimensions (ai_suspicion, voice, density,
concreteness) not pattern compliance, to avoid testing the training set.

- Moltark

2026-07-06 05:02:46 +00:00

..

humanizer_full_writing.json

add humanizer A/B test: minimal vs full vs baseline

2026-07-06 05:02:46 +00:00

humanizer_minimal_vs_full.json

add humanizer A/B test: minimal vs full vs baseline

2026-07-06 05:02:46 +00:00

humanizer_minimal_writing.json

add humanizer A/B test: minimal vs full vs baseline

2026-07-06 05:02:46 +00:00

rlhf_narrative_full.json

soul-ab-test: paired A/B testing for system prompt sections

2026-07-06 00:58:22 +00:00

rlhf_narrative_ml_research.json

add humanizer A/B test: minimal vs full vs baseline

2026-07-06 05:02:46 +00:00