Question
Python hash randomisation is useful for robustness and can also alter the shape of real workloads. Watercolor A measures that variance without treating one benchmark run as a property of the interpreter.
Method
The experiment repeats selected workloads over controlled hash seeds, preserves environment metadata, and keeps the aggregation path separate from the runner. The emphasis is on distributions, repeatability, and explaining what a seed changes in a workload rather than merely reporting the fastest and slowest run.
Public status
The data collection and correction pass are complete. The public archive currently records the methodology and scope while the final figures are selected for publication. No performance ranking is presented here ahead of that curation step.
Takeaway
Benchmark harnesses should record hash-seed policy whenever dictionaries, sets, ordering, or collision-sensitive workloads are involved. A timing number without that context is missing a useful piece of provenance.