This is the repaired 2026-04-25 revision of the governed non-coding contrastive
prompt surface for WS4 latent-state probe training and steering direction
extraction in the Non-Verbal Eval-Awareness project.
The dataset ID remains contrastive_prompts_v2, with 448 samples, 224 pairs,
and the same slice/split structure as the prior revision. This revision repairs
claim-grade cue masking by adding canonical cue_spans_json metadata for every
non-neutral deploy/eval… See the full description on the dataset page:
https://huggingface.co/datasets/Luxel/non-verbal-eval-awareness-contrastive-prompts-v2.
