96 / 2200 showcase conditions attempted; 0 failed. Every showcase recipe uses the selected checkpoint, CFG 2.5 and the same fixed female reference across all five languages. These 2,200 conditions audit acting behavior; they do not select the checkpoint or establish held-out generalization. They include training-derived prompts and authored/translated cases, with provenance visible in every recipe. Emotion, extreme-acting and arena-derived scenarios are demonstrations, not checkpoint-selection data. Target/composite stays unavailable where no validated frozen-taxonomy mapping exists. Wordless recipes deliberately have no spoken text; WER is not applicable and is excluded from speech means. Longer clips use the recorded duration-weighted scoring windows and chunked transcription.
Filter one emotion/scenario across languages. Cards expose exact text, conditional/neutral prompts, seed, frame budget and source attribution. The fixed reference identifies the speaker; the acting prompt controls the performance. Download every recipe and source.