Over the past year, frontier AI has produced novel breakthroughs in mathematics, biology, and cybersecurity.

While of economic interest to research labs, these domains share a more relevant characteristic: they are highly verifiable. This is not a coincidence. Contemporary learning mechanisms hillclimb fastest in verifiable domains.

We believe that AI R&D can be similarly mapped onto a closed set of verifiable problems, and that leveraging those problems will yield comprehensive gains in model AI research capabilities and greatly accelerate RSI.

Jigsaw distills frontier AI R&D into verifiable post-training artifacts, develops rigorous methods for evaluating their quality, and measures the capability gains produced by training on them.

Our goal is to accelerate the path to ASI by making recursive self-improvement tractable, and to translate AI research into faster scientific progress, stronger academic research, and broader benefits for humanity.