Short English sentences that two same-class LLMs from different providers reliably read differently. Each row carries two competing one-sentence readings of what is going on beneath the literal text; percept is the side google/gemma-3-12b-it measurably lands on, percept_A the side Qwen/Qwen3-8B lands on. Divergence is not designed-in by fiat — it is measured, per row, on both models, with… See the full description on the dataset page:
https://huggingface.co/datasets/cds-jb/synthpercept-v3.
