When two parts with independent manufacturing tolerances are joined into an assembly, the combined tolerance can be estimated a few different ways — and they don't agree, which is exactly the point of this tool.
Worst-case is guaranteed safe but almost always far too pessimistic — in practice, it's extremely unlikely that every part in an assembly lands at its extreme dimension at once. RSS is the standard correction for that: for a manufacturing process running at ±kσ (σ = tolerance/k), it predicts a tighter, statistically realistic assembly spread, and it's exact in the limit of infinitely many parts. A process running at ±3σ has a process capability Cp = T/(3σ) = 1.00; at ±4σ, Cp ≈ 1.33 — tighter control, narrower real spread, for the same stated tolerance.
What RSS can't show you is what actually happens with a finite batch of real parts randomly paired on an assembly line — which is exactly what the "Assemble" simulation does here: it draws real, discrete populations for Part 1 and Part 2 (1008 of each, matching the classic textbook batch size), randomly matches them one-to-one without replacement, and measures the assembly spread that actually results. The outcome is always close to the RSS prediction, but not identical to it — the gap between "Actual (this run)" and "RSS-Predicted" is the honest, expected effect of finite-sample randomness, not an error in either number.
The tool also reports escapees — parts that a ±3σ Normal process still produces just outside its own stated tolerance band (the well-known ~0.27% tail of a true ±3σ process), which a Uniform/Linear distribution can never produce since it's bounded by definition. Switching either part between Normal and Uniform, or between ±3σ and ±4σ, updates every number on the page live, so you can build intuition for how process capability and distribution shape actually move the needle on assembly-level risk.