6.7. Compare like with like🔗

Before attributing a change to TorchLean:

• retain the resolved TorchLean revision;

• use the same toolchain, build options, model arguments, seed, backend, dtype, and device;

• warm up compiled graphs and device libraries before active samples;

• collect several process runs;

• keep device completion policy identical.

The trace is the place to diagnose order and nesting. The summary is the input to an automated comparison.