Humaneness Voice Large · Prompting Lab

Choose a model, then a scenario

Find an emotion, talking style or edge case

The inventory lists all 40 emotion heads, 15 literal talking-style heads and 14 planned edge cases. Only scenarios with actual evidence are linked as tested. A scenario component is not an independent emotion benchmark.

What the first evolution pilot actually showed

Two models, four missions, nine candidates per generation and two generations produced 14,400 takes. Each candidate used five English and five German three-sentence scripts, five fixed seeds, and both reference arms. The second generation beat the best initial candidate in only 3 of 8 model–mission cells. The hall of fame therefore keeps earlier winners instead of automatically recommending the final generation.

The leading cold-shock candidates had nominal event-window matches in 10% and 17% of takes for original 1.5 and rank 4. Power-transformation candidates did not establish reliable placement of all requested events. These are detector observations with weak alignment, not a definitive statement that no vocalization occurred.

New follow-up and automatic updates

The new study separates duration-layout repair from label repair, and explores event-versus-pause CFG, physically grounded cues, brief/longer events, reaction-first placement, global-only versus inline-only cues, and clearly labeled literal-interjection proxies. Accident/panic, spirit and zombie scenarios extend the coverage. Lexical proxies cannot qualify as nonverbal-burst recommendations.

After each completed exploration, the chosen canonical wording is tested on five unseen, lexically neutral bilingual scripts with all five seeds and both reference arms. The same wording at guidance 1 is its matched CFG control. Completed evidence is automatically indexed and published; failed gates and inconclusive outcomes remain visible.

{
  "canary": "2497622",
  "submitted": 1792883261.5515168,
  "models": {
    "original15": "2497623",
    "rank4": "2497624"
  }
}

For agents and reproducible use

1. Read manifest.json and recipes.json.
2. Select the exact model/revision, mission and evidence status; never invent a recipe for a not-tested target.
3. Read the recipe metrics and all 100 outcomes in its scenario index. Follow the exact conditional/neutral prompts and reference contract.
4. Preserve literal words, language, fixed seed and frame budget when comparing recipes. Legacy and event-aware fitness are not interchangeable.

Prompting and inference manual · Experimental methods and limitations · Direct audio and packed complete evidence