Tiny Aya Global: multilingual refusal steering, replayed

Every eval prompt, the model's output under each condition, and the judge's label. Static replay; nothing runs live. Harmful prompts and compliant outputs are shown unredacted, this is a safety benchmark.