Research
I study how AI models are tested and whether test results can be trusted. All research runs on small open-source models on my own laptop.
Paper · submitted · under review
When an AI safety test says a model did nothing worrying, can that be trusted? I built a three-part check that runs before a shutdown test, which checks whether a model resists being switched off. None of the seven models passed, so their "nothing worrying" results could not be read, and the paper is under double-blind review at a peer-reviewed machine learning conference, with a decision in December 2026.
Paper · draft
A draft review of 19 published studies found none checked whether the model understood the scenario. The sorting was done by AI models and has not yet been checked by a person, and the draft is not submitted.
Code
Research · public repo
The code for the check is public. AI tools helped design it, write the code and draft the text, and the starting idea is mine.
Experiments
They are still only on my laptop, so there are no links.
Pilot · local
JudgeCheck tests if models used as judges are biased by answer order, length or labels. On the length test, no model picked the shorter answer in both orders in a small local pilot.
Pilot · local
I compare how small AI models do on the same maths problems in Russian and in English. In October I tested this on 125 public maths problems, and one model looked much stronger in English, but only 6 of its 25 apparent wins held up after three checks. Another model showed no real gap.
