I’d argue repeatability is impossible to guarantee. Any cloud based computing with floating point math is subject to variance. Bare metal local LLM computing can be relatively stable, but still suffers from precision error from your GPU.