FAMOUS

A job I had crashed several time at the same point. I’ve restarted the job just a month before it crashed and changed the restart dumps to 1 day (was 360 days before) and it goes for on and on … (ouch, so much output, had to kill it after 13 yrs).
Did you mention, Robin, something about changing output frequency and bit reproducibility ?

yes - if you change the restart dump frequency the job is no longer bit-identical. Essentially, the model stops and restarts every time a dump is put out, and this changes the run slightly, compared to running normally, so your 1day-dump run /should/ look (at the bit level) different from the 360day-dump run.
Clearly, this isn’t an ideal situation, but that’s how the model is written. Whatever is causing your crash seems to be very sensitive if such a small perturbation can change it.