Tolerance stack-up example: worst case vs RSS on a prototype enclosure

A tolerance stack-up is a calculation that combines the dimensions and tolerances of every feature in a chain to predict the smallest and largest value of one functional result: a gap, a clearance, an interference, or an alignment. ASME Y14.5 is the standard most US drawings use to state those tolerances, and it frames dimensioning and tolerancing as the way a drawing communicates fit, form, and function (ASME Y14.5).

In a prototype program, a stack-up answers a practical question before you spend money on parts: if every part comes in anywhere inside its tolerance, will the assembly still go together and work?

A common electromechanical prototype: a circuit board sits on bosses inside a 3D printed base, and a lid closes over it. The tallest part on the board, for example a connector or a capacitor, must not touch the lid. We want at least 0.30 mm of clearance so the lid does not press on the part when the product flexes or is dropped.

The numbers below are illustrative, chosen to show the method. Process tolerances are taken from published sources:

Start at one surface of the gap and walk through every part until you arrive at the other surface. Each dimension you pass is a link. Dimensions that open the gap are positive; dimensions that close it are negative.

Our first version of the drawing dimensions the base from its floor:

Nominal gap = 14.00 + 6.00 - 5.00 - 1.60 - 12.50 = 0.90 mm.

Worst case assumes every part is at its limit at the same time, so the tolerances add directly. PTC's Creo documentation describes it as predicting the absolute upper and lower limits that all acceptable parts can produce (PTC Creo, types of stack-up analyses).

Worst case tolerance = 0.30 + 0.30 + 0.30 + 0.16 + 0.20 = ±1.26 mm

Gap range = 0.90 ± 1.26 = -0.36 to 2.16 mm. A negative gap is an interference: in the worst combination, the lid hits the part.

RSS treats each dimension as a random, independent variation and combines them statistically:

RSS tolerance = √(0.30² + 0.30² + 0.30² + 0.16² + 0.20²) = ±0.58 mm

Gap range = 0.90 ± 0.58 = 0.32 to 1.48 mm. That passes the 0.30 mm requirement, barely.

So which answer is right? It depends on whether RSS's assumptions are true.

The classic RSS method assumes each dimension is normally distributed, centered in its tolerance band, with the tolerance equal to three standard deviations, and that dimensions vary independently (Scholz, University of Washington, Tolerance Stack Analysis Methods).

Scholz also notes that practice has shown the results are usually not quite as good as advertised.

Tolerance analysis vendors make the related point that RSS also assumes linear geometry, and that non-linear cases such as cams suit Monte Carlo simulation better (Enventive on worst case, RSS, and Monte Carlo).

Statistics need volume. When you build three or five prototypes, any combination of parts can show up, and the base and lid may come off the same print job with correlated errors. That breaks the independence and centering assumptions RSS relies on.

Our default: use worst case for prototype builds, and use RSS (or Monte Carlo) when planning production, where real process data exists and you can state the assumptions next to the result. If a prototype stack only passes under RSS, treat it as a risk to test, not as a pass.

The worst case result above fails. Here are the levers, using the same example.

1. Fewer links: dimension to the function. Links A and C both start at the floor of the base, but the gap does not care where the floor is. Dimension the base from the rim straight to the top of the boss instead (9.00 ±0.30) and the chain drops from five links to four.

Worst case tolerance = 0.30 + 0.30 + 0.16 + 0.20 = ±0.96 mm, so the gap range becomes -0.06 to 1.86 mm. Much better, still a slight interference.

2. Tighter process on the links that matter. Moving the base and lid from prototyping-grade SLA (±0.30 mm) to industrial SLA (±0.13 mm at this size, per the Protolabs Network standard above) gives:

Worst case tolerance = 0.13 + 0.13 + 0.16 + 0.20 = ±0.62 mm, gap range 0.28 to 1.52 mm. Close, but still under the 0.30 mm target. Notice that the PCB and the component now drive most of the stack, and you cannot print your way out of those.

3. Shift the nominal. With the four-link chain and the original print process, raise the lid's inside ceiling by 0.40 mm (lid link 6.40 ±0.30). Nominal gap becomes 1.30 mm and the worst case range is 0.34 to 2.26 mm. It passes, and it costs nothing in process.

In a real program you check the side effects of each fix before choosing: a taller lid changes the product's overall height and may move a button or a display window, so the stack-up goes into the design review with those trade-offs visible.

A stack that works on 3D printed parts has to be redone for the production process. Published 3D printing tolerances vary widely by technology and grade: the same Protolabs Network standard lists prototyping FDM at ±0.5% with a ±0.5 mm lower limit, while Protolabs lists SLA at ±0.002 in.

(±0.0508 mm) in X and Y for the first inch plus 0.1% of nominal, and ±0.005 in. (±0.127 mm) in Z for the first inch plus 0.1% of nominal height (Protolabs, 3D printing tolerances). Build orientation matters too, because Z tolerances are often looser than X and Y.

When the part moves to injection molding, tolerances, shrink, and draft all change, and tighter tolerances drive tooling cost. Our guide to injection molding cost and what drives tooling covers that side, and soft tooling vs hard tooling covers the bridge between printed and molded parts.

AI tools can speed up the slow parts of a stack-up, mainly finding the dimensions and building the chain, but an engineer still has to confirm that the chain is the right one and that the inputs are real.

What vendors say their tools do (treat these as vendor claims): bananaz describes building stack-ups from the dimensions and GD&T already on your drawings and computing worst case and RSS automatically (bananaz tolerance analysis).

RD8 describes automatically detecting tolerance chains between two selected parts from a STEP file and running RSS, worst case, or Monte Carlo (RD8 automatic stack-up detection).

TolFlow, built for Onshape, says it pulls dimensions and GD&T from structured drawing data rather than relying on LLM guesswork (TolFlow).

That last point is the one we agree with most: let software extract and calculate deterministically, and do not let a chat model do the arithmetic from memory.

Where we find AI useful on a stack-up:

Where a human still decides:

For our broader view of AI in CAD work, see AI CAD agents for prototyping in 2026.

At LA NPDT, mechanical design, electronics, and prototyping sit in the same team, so the board, the parts on it, and the enclosure are checked together rather than in separate files. In practice:

If you have a design that needs to fit together, or an idea that needs a design, start with a free 20-minute call: book here. When a project needs a deeper look before any build, our Stage 1 Review/Analysis + Plan is a fixed $800 single payment. For enclosure-specific checks, see our enclosure design for electronics checklist.