These reviews were assembled between October and December 2025 as evidence reviews for an AI-futures simulation, a model of how the coming decades of AI could unfold (the simulation's repository is public). Each one asks what the published research said about a single question: how often models fake alignment, whether sandbagging can be detected, how alignment techniques fail, what AI changes about nuclear risk, whether AI systems could have welfare, and whether scaling is slowing down.
Each review was sourced and checked when it was written. The date it was written, the date it was last updated and, where the review records one, the date its sources were last checked are at the top of every page. The reviews have not been updated since, so read each one as a record of the evidence as of its dates.
They are lightly edited for a public reader. The parts that only made sense for tuning the simulation, such as parameter values chosen for the model, implementation notes and internal file references, have been removed. Every finding, number and citation is kept, including the caveats and the evidence that cuts against each review's own conclusions. No claims have been added.
The methods note at the end of the list describes how the sources behind these reviews were audited for age and citation accuracy.