If I understand the article correctly the results are on models that can't be accessed and as such this is pretty expensive to actually verify. I'd be really, really sceptical.

Hi, I'm the post author. At the end of the article I share how you can access all three MoA models we've prepared via our public chat completions API!