Preview with incomplete coverage. These points test the interface; they are not a full-corpus comparison.
Inspect a point
Select a point, search for a file, or choose a row below.
| Source | Our time | Radar instructions | Share difference |
|---|
Rows are ordered by the absolute difference in cost share. Search includes every matched source.
What these comparisons measure
CPU instruction counts, frontend elapsed time and native compilation wall time have different measurement boundaries. Our frontend timer includes imports and instrumentation, and ends before trace export. Source-attributed time allocates overlapping source events without counting their time twice. Radar counts instructions for native file compilation; its wall time also includes output work.
Rank correlation measures ordering. Share disagreement is half the sum of absolute differences between normalized cost shares. Neither statistic is a percent timing-accuracy estimate. Hosted Radar uses different hardware and concurrent compilation. Local measurements use pinned sources and unchanged public Radar code. The methods record hardware and runtime evidence, measurement boundaries, and profiling-only options. The original LeanPool capture has no independent hardware fingerprint.
A clock check after capture found about a 4.45% scale difference between Linux and Windows timers. Original timings are shown. The methods preserve the samples and explain why this observation cannot retrospectively calibrate the captures.