I'm an independent researcher. I study how AI systems fail when they make decisions one step at a time.

The question I keep coming back to is pretty simple. A system builds something piece by piece. A sentence, a plan, a string of actions. But it can only see part of what it's building at any given moment. So what can it actually promise about the finished thing? Usually less than people assume. That gap is where a lot of familiar problems live. Hallucination in language models. Dead ends in planning. Runs that fall apart in ways no local check would have caught.

I don't work at a university or a lab, so the work has to carry itself. Every main result gets written up, formally verified, and put somewhere public with a permanent link. Anyone can check it. They don't have to know me or trust my background. For an independent researcher that's not a nice extra. It's the whole point.