Same seed, different results: how CPU architecture silently broke a data experiment

A developer running a reproducible data pipeline discovered that identical code with the same fixed seed produced different results across machines. The root cause was traced to NumPy's sort implementation, which selects its sorting routine based on the host CPU's capabilities. When sorting a column with thousands of duplicate values, the internal ordering of tied rows varied depending on whether the processor supported AVX-512 instructions. A positional slice applied after the sort then captured different subsets of data, shifting the experiment's apparent success rate from 8/15 to 14/15 on identical input. Testing across six cloud runtimes confirmed the split fell exactly along AVX-512 support lines, with no way to predict in advance which hardware a runtime would provide.