Iterations done between the last converged substep and the bisection are discarded so the error is no longer relevant. Iterations restart from the last converged substep using the smaller time step, so if there are no new warnings, then the result will be reliable.
It looks like the last step could finish in fewer iterations if you allowed a larger step size. The second last step may finish in fewer iterations but you won't be able to use the same large step size that you use in the last step, it will need to be smaller than that.
You are navigating away from the AIS Discovery experience