What if you only have 20 parts? Pool across runs, use short-run SPC, report the CI alongside the point estimate, negotiate with the customer. Practical guidance for low-volume / high-mix shops:
https://t.co/BbURPt0XQN
Your PPAP submission says Cpk = 1.45 from 30 parts. The customer reads 1.45 as a true value. The actual 95% CI is [1.07, 1.83]. The minimum 1.33 threshold is inside that interval. The estimate told them what they wanted to hear.
Standards converge: AIAG SPC manual says 30 parts for initial estimate, 100+ for long-term. ISO 22514 ties n to CI width. Bosch's convention is 25 subgroups of 5 = 125 parts. Pick from those numbers, not from convenience.
Cpk = 1.45 from a 30-piece capability study has a 95% CI of roughly [1.07, 1.83]. The point estimate cleared the 1.33 threshold. The lower bound didn't. Why customers want 100+ parts, with the formula and worked numbers:
https://t.co/BbURPt0XQN
Browser calculators sit in the middle: less powerful than R, no installation, free. Best for ad-hoc supplier capability checks. The verdict by reader type:
https://t.co/BAQixhrZhA
'Free SPC software' is doing a lot of work in those Capterra listicles. A 30-day Minitab trial is not the same product as the qcc R package, but they get the same green checkmark. The categories that actually differ:
Open-source SPC packages — R + qcc, JASP, Python pyspc — are the indefinite-use answer most listicles ignore. Real learning curve, but no time limit, no chart cap, no nickel-and-diming.
Most 'free SPC software' lists conflate 30-day Minitab trials with R + qcc and call them both free. They're not the same thing. An honest 5-category taxonomy and which fits which use case:
https://t.co/BAQixhrZhA
Two control charts on your desk: one shows a point above UCL, the other shows 7 points drifting up — none beyond limits. Both are signals. Most operators react to them the same way. That's the mistake.
Nelson rule 3 (6+ points trending) usually means a slow drift — tool wear, temperature, transducer drift. Investigation: check elapsed cycles since last index, not the change log.
That single point above UCL and the 7 points trending up are both signals — but Nelson rule 1 and Nelson rule 3 point to different causes. A decision tree for reading the chart and acting on what fired:
https://t.co/zAjbKgsBKk
Pragmatic shops run both: Shewhart on the floor for operator response, CUSUM in the background for engineer-level drift detection. EWMA is the middle-ground when you want CUSUM sensitivity with Shewhart ease of reading. Full decision framework: https://t.co/paVy8jK6Ya
Shewhart charts are great at catching large shifts (2 sigma and up) in one or two subgroups. They are terrible at catching small sustained shifts. A 1-sigma shift takes an average of 44 subgroups to signal on an X-bar chart. That is a full week of production at many shops.
The trade-offs: CUSUM requires training to interpret (operators find cumulative sums less intuitive than Shewhart signals), needs tuning (reference value K, threshold H), and is slightly slower on very large shifts. Shewhart wins on ease-of-use and audit recognition.
Shewhart X-bar takes 44 subgroups to detect a 1-sigma shift. CUSUM tuned for 1-sigma takes 10. For processes where small sustained drifts cost money - tool wear, reagent aging, calibration drift - CUSUM is dramatically faster. When to pick which: https://t.co/paVy8jK6Ya