This is a grant covering work Jan did as part of MATS with me (during the period pre MATS where MATS isn't able to cover stipends). Jan did good work, e.g. working on this paper on meaningful improvements to activation oracle training to make them a better tool for researchers https://arxiv.org/abs/2606.02609 . I have an obvious conflict of interest here, but I think this is clearly above the bar.