Guides · 29 Oct 2026
Tolerance stack-ups with a language model: a checking workflow
Language models are quick at setting up a tolerance stack-up and surprisingly confident about the result. This guide walks through a simple one-dimensional stack on a shaft assembly and shows a checking routine that catches the usual mistakes.
The assembly
A shaft carries a bearing, a spacer and a gear, held by a circlip against a shoulder. We want the axial clearance between the gear and the housing wall to stay between 0.1 and 0.6 mm.
| Contributor | Nominal, mm | Tolerance, mm |
|---|---|---|
| Housing depth | 42.00 | ±0.10 |
| Bearing width | 12.00 | 0 / −0.12 |
| Spacer | 8.00 | ±0.05 |
| Gear hub | 21.50 | ±0.08 |
What the assistant got right
Asked for a worst-case and a root-sum-square result, the assistant built the loop correctly, assigned directions and converted the asymmetric bearing tolerance to a mean and a symmetric band. That part saved real time.
Where it went wrong
- It dropped a contributor. The circlip groove position was never mentioned in the prompt, so it was never in the loop. An assistant can only stack what you describe.
- It mixed up the bearing direction after a follow-up question and silently changed a sign. Always ask for the full table again, not just the final number.
- It applied RSS to four contributors and called it statistically safe. With so few contributors that assumption is weak; we keep worst-case as the acceptance check.
The checking routine
Draw the loop yourself first, on paper. Then give the assistant your table and ask it to recompute, not to build the loop. Compare both results and resolve every difference by hand. It takes about ten minutes and catches all three errors above.
Use the assistant for arithmetic and formatting. Keep the loop, the directions and the acceptance criteria in your own head.