Keeping comparison scope explicit

This overview explains how workload, host, output, lifecycle, and timed stages define what a measurement answers. Results that look contradictory can all be correct because they describe different tasks or boundaries. A useful comparison card identifies the input family and support, the required artifact, the lifetime of prepared objects, and the timed boundary on a specified host. These details determine which requests the evidence actually covers. Suppose the candidate kernel costs two units and the control kernel costs three. Within this explicitly limited kernel boundary, the candidate is cheaper. We have not yet counted how either kernel is prepared or how its answer reaches the caller. Add five units of candidate overhead and one unit of control overhead. Complete totals become seven and four. The kernel win and the request loss coexist because they refer to different scopes. Report both without silently moving the claim between them. Scope identifies inputs, host, versions, output and lifecycle. Boundary identifies the stages timed. Before pooling values, ask whether both studies estimate the same quantity and use a compatible sampling design. The corrected B2/B4 and accepted EPFL results occupy separate workload rows. Their compiled-kernel ratios answer different scoped questions. Preparation and public-wrapper results occupy separate columns from the kernel measurements. A cell's location now tells us what it measured. Adding a scalar-count result creates a different output category, not another sample of complete-vector execution. The retained program includes kernel gains, parity, slower wrappers and task-specific counting improvements. They can coexist because they answer different scoped questions. Cold, warm and reload observations must retain their lifecycle labels; a warm count is not a cold full-vector comparison. Imagine evidence for one family producing a fresh full vector. That cell in the comparison table is supported. It does not automatically fill a neighboring cell for point queries or resident queries. Changing any of those dimensions may change the work and the winner. An unmeasured cell should remain unmeasured rather than inheriting the nearest favorable result. Lack of evidence is not a measured loss either. A defensible conclusion repeats the scope that was actually tested. Broader application needs a matching experiment or proof, depending on the claim. Keeping that boundary visible makes the result more useful to someone choosing a method. Keep each result's scope and lifecycle attached. Combine estimates only under a justified common target and sampling design.