Hidden in Figure 50 is that Gemini 3.1 Pro has a compliance gap of 37% (!!!) This is among the largest out there.

I replicated this result on my own prompted alignment faking setup, which has a semantically identical prompt that's been perturbed to avoid memorization. On my ... https://t.co/EQxsNLnqet https://t.co/rP9irgNQeC