A drawing full of tight individual tolerances can still produce parts that will not assemble. The reason is usually not a single out-of-spec dimension but how those dimensions accumulate across a chain of features. Tolerance stack-up is the arithmetic of that accumulation, and it is the check that tells a buyer whether a design has a functional requirement in mind or just a set of numbers that look precise on paper.
This guide is written for design and purchasing teams who issue RFQs for CNC machined parts. It covers how to define the functional requirement, how worst-case and statistical approaches differ, why datum consistency matters, how an inspection plan should reflect the stack, and what to put in an RFQ checklist so the supplier can quote and inspect against the same logic.
Start from the functional requirement, not the drawing default
A tolerance stack only has meaning relative to a functional requirement. Before assigning any number, state what the assembly must actually do: maintain a clearance so two parts never interfere, hold a gap for a seal, keep a pivot centered, or guarantee a pin enters a hole. That statement becomes the closing dimension of the stack — the result the calculation must satisfy.
Only then do you decide how much of the total budget each contributor may take. A common failure is to tolerance every feature independently and equally "because the shop can hold ±0.05 mm." That spreads the budget evenly and often leaves the critical interface with too little room, while non-critical features consume tolerance they do not need. Buyers can prevent this by marking the functional dimensions on the drawing and letting the non-functional ones carry looser tolerances — the same discipline covered in CNC Machining Drawing Requirements: RFQ Checklist for Buyers.
This is also a cost lever. Every dimension tightened below what the function needs adds machining time, fixturing, and inspection. Danfoss, in its machining design guidance, makes the same point: tolerances should be selected with function in mind, because unnecessarily fine tolerances increase cost, and rough surfaces do not improve sealing or function where they are not required. The practical rule is to loosen everything that is not part of a stack and spend tight tolerances only where the assembly result demands it.
Worst-case and statistical approaches: know which one you are buying
There are two fundamentally different ways to combine contributors. The first is worst-case (arithmetic) analysis. Every variable is placed at its tolerance limit in the direction that makes the closing dimension as large or as small as possible, and the limits are simply added. As Wikipedia’s overview of tolerance analysis states, worst-case analysis predicts the maximum expected variation, and designing to it guarantees that 100 percent of parts will assemble and function, regardless of actual variation. The trade-off, stated plainly in the same source, is that worst-case tolerancing often requires very tight individual tolerances, which drives up manufacturing and inspection cost and scrap.
The second approach is statistical. Here each contributor is modeled as a distribution, and the distributions are combined to predict the distribution of the assembly result rather than its extreme values. The classic method is root sum square (RSS), where independent contributors are combined as the square root of the sum of their variances; Monte Carlo simulation is the more general alternative and is the usual choice for 3D problems. Wikipedia notes that statistical analysis relaxes component tolerances without sacrificing quality, but that it predicts a distribution, not a 100 percent guarantee. That distinction is the one a buyer must be explicit about.
Practical consequences for an RFQ:
- Worst-case is the right choice for critical interfaces, safety-relevant fits, and spare-part interchangeability, where a failed assembly is unacceptable. It is also the conservative default when you have no data on the supplier’s process capability.
- Statistical methods only pay off when the contributor distributions are real and roughly centered. RSS assumes the variables are independent and reasonably normal, and it assumes no systematic shift such as tool wear drift, setup offset, or a batch of material that machines differently.
- Never mix the two silently. If the drawing says statistical and the inspection plan says 100 percent, you are paying for a guarantee the design does not provide — or accepting risk you did not intend.
Which method applies is often a contract question rather than a pure engineering one, and both should be stated in the RFQ so the supplier knows what it is being asked to produce and verify.
Datum consistency is what makes a stack valid
A stack is only valid if every dimension in the chain is measured from the same reference logic the assembly uses. This is where datum consistency fails most often. If the design stack is built from Datum A, but the shop machines and inspects from a convenient setup face that is not Datum A, the individual measurements can all be in tolerance while the functional result misses.
GD&T Basics, in its Datum Reference Frame article, explains the mechanism. Datums are derived from datum features, which are real, tangible features on the part and usually important functional surfaces; the datums themselves are theoretically exact points, lines, axes, and planes derived from those features and simulated by measurement equipment. The datum reference frame is the coordinate system created by the datums listed in a feature control frame, and it must lock down the degrees of freedom the part needs — generally all six. The order of datums matters, because it dictates which features take precedence, and the recommended practice is to select datum features in the order the part would assemble in real life. Lower-precedence datums only control the degrees of freedom not already controlled by higher-precedence ones.
Two practical rules follow for buyers:
- Match the datum reference frame to assembly, not to manufacturing convenience. A part that locates on three mutually orthogonal functional surfaces in service should be dimensioned and inspected from those same surfaces.
- Keep the datum scheme small and functional. A drawing that scatters many different datum letters across features makes it hard for the supplier to build one setup that satisfies all of them, and hard for inspection to reproduce the same reference.
Note also the standard assumption on the drawing itself: as the GD&T overview records, dimensions and tolerances are taken to be valid at 20 °C and 101.3 kPa unless stated otherwise. If a measurement is made in a hot shop, or if a thin part is measured in a free state and then clamped, the numbers can shift without anyone being wrong. That is a boundary condition worth naming in the RFQ whenever the part is large, thin, or temperature-sensitive.
An inspection plan should test the stack, not just the tolerances
An inspection plan that checks every dimension in isolation can pass a batch that fails at assembly, because isolated checks do not exercise the accumulation. The inspection plan should mirror the stack: identify the closing dimension and the contributors, then decide how each contributor is verified and how the assembly result is confirmed.
Guidance from NIST on dimensional metrology reinforces the underlying point: measurements carry their own uncertainty, and a reported value is only meaningful together with the reference conditions and method behind it. A ±0.05 mm tolerance checked with a device whose uncertainty is a large fraction of that tolerance is not real confirmation. When the stack depends on a tight contributor, the measurement method and its uncertainty should be agreed, not assumed.
A workable structure:
- Name the closing dimension and the contributors on the inspection plan, so the inspector knows which dimensions are functional and which are reference.
- State the datum sequence and the setup used to simulate each datum, so inspection reproduces the design reference frame.
- Decide which contributors are checked on every part and which are verified by first-article or process capability, based on their share of the budget; the reporting side of this is described in CNC Machining Quality Inspection: Buyer Guide to Reports and Checks.
- For statistical designs, record the assumed distribution or capability so the yield claim is traceable rather than asserted.
- Verify the assembly result itself where possible — a functional gage, a check fixture, or a physical trial fit — because that is the requirement the stack exists to protect.
RFQ checklist for tolerance stack-up
Buyers do not need to perform every calculation, but they should hand the supplier a package in which the stack can be reproduced. The items below are the ones that most often go missing.
- Functional requirement stated in words: what the assembly must achieve, and which dimension closes the stack.
- Functional dimensions identified: the contributors to the stack marked separately from general dimensions.
- Method specified: worst-case or statistical, and if statistical, the assumed distribution or expected capability.
- Datum scheme unambiguous: datum features named, precedence order set, and consistent between drawing, model, and inspection.
- Reference conditions noted: temperature, free state versus restrained state, and any surface or material condition that affects the measurement.
- Inspection plan agreed: which contributors are 100 percent checked, which are sampled, the measurement methods, and how the assembly result is confirmed.
- Standards cited: name the GD&T and dimensioning standards the drawing follows, so there is no interpretation gap.
How much tolerance to spend is also a cost decision; the trade-offs between design intent and machining cost are set out in DFM for CNC Machining: Design Guidelines to Reduce Cost. Finishing calls interact with the same stack, since a coating or polish can shift a dimension; those choices are compared in Surface Finish Options for CNC Machined Parts: Buyer RFQ Guide.
The last item is easy to overlook. A drawing that cites one geometric tolerancing standard in the title block but is read by a shop that works to a different convention can produce a compliant part that does not assemble. Naming the standard, and the revision, removes the ambiguity before it becomes a rejected lot.
What this means in practice
Tolerance stack-up is not advanced mathematics; it is disciplined thinking about which variation matters. Buyers who ask for the functional requirement up front, who say whether the design is worst-case or statistical, who keep datums consistent from design to inspection, and who agree an inspection plan that tests the accumulation rather than each number in isolation, get parts that fit. Those who skip those steps discover the stack at assembly, when the drawing is already fixed, the tooling is already cut, and the cost of change is highest.
For CNC machined parts, the most useful thing an RFQ can carry is not an ever-tighter tolerance but a clear statement of which tolerances are functional. That single clarification lets the supplier spend precision where it matters and keeps the rest of the part economical to produce and verify.
