Begin with the method
Record the algorithm, precision requirements and validation reference. A hardware generation optimised for one AI workload is not automatically the most suitable configuration for a scientific solver.
Test the scaling boundary
Measure a representative small run, then a larger one with a defined strong- or weak-scaling objective. Look for time spent in communication, CPU work, data movement and serial parts of the application.
Preserve reproducibility
Keep software versions, compiler choices, dataset identifiers and job configuration. Plan scratch capacity and checkpoints. The evidence brief can capture these requirements before an infrastructure conversation.
Prepare a comparable test
Keep the workload, software version, data path, service objective and cost boundary attached to the result. Changing these assumptions changes what a comparison means.