Before each iteration of self-improvement (limiting it to successive generations if necessary). model the next generation of itself, and, simultaneously, model a maximally-realistic human capable of understanding/​controlling/​mastering both the next generation and the current generation, simultaneously. Halt activity and give an error report if such a human cannot be generated or modelled.
Maybe generate that human so it can model all previous generations as well, or even include inert training data of generations from different branches.
Before each iteration of self-improvement (limiting it to successive generations if necessary). model the next generation of itself, and, simultaneously, model a maximally-realistic human capable of understanding/​controlling/​mastering both the next generation and the current generation, simultaneously. Halt activity and give an error report if such a human cannot be generated or modelled.
Maybe generate that human so it can model all previous generations as well, or even include inert training data of generations from different branches.