A Gage R&R Run on One Hour of Production

The number says unacceptable; the study design says don't trust the number yet.

• On its face: %GRR of 32% of study variation exceeds the 30% threshold, and ndc = 3 is below the recommended minimum of 5. A system with 3 categories cannot resolve the process variation well enough to track improvement.
• But the study violates a core assumption: parts must span the actual process variation. Ten parts from one hour on one machine represent a narrow slice, so part-to-part variation is understated. Because %GRR is GRR variation divided by total study variation, shrinking part variation inflates %GRR and depresses ndc. The lead's instinct is not entirely wrong - but neither is the 32% conclusion established.
• Resolution is not the problem: 0.001 mm is 1/200 of the 0.20 mm tolerance, far finer than the 1/10-of-tolerance rule. The lead's argument fails because resolution was never the issue; equipment, fixturing, method, or operator technique is.

Do next:

1. Re-select 10 parts spanning the full expected process range (across machines, shifts, and lots), not one hour.
2. Decide the yardstick first: for a tolerance-based accept/reject decision, report %GRR against tolerance (P/T ratio) as well as %study variation.
3. Standardize the method - fixture, measurement location, clamping torque, part temperature, reading convention - then randomize part order and blind the operators to part identity.
4. Re-run and examine the variation split: if repeatability (equipment) dominates, fix gage/fixture; if reproducibility (operator) dominates, fix training and method.

Do not compute capability on this data. A 32% measurement system consumes a third of the observed variation and will make Cpk look worse - or better - than reality.

78% agreement sounds acceptable but is not evidence of a usable inspection system; kappa 0.41 is the real signal.

• Kappa corrects for agreement expected by chance. 0.41 sits in the "moderate" band and well below the ~0.70–0.90 usually required before an attribute system is used for decisions. The 78% figure is inflated whenever defects are rare or common, because inspectors agree simply by saying "pass" most of the time.
• Break it down before acting. Within-inspector agreement (repeatability: does each inspector agree with themselves across trials?) versus between-inspector agreement (reproducibility: is one inspector systematically stricter?). Also compute, against a known standard: effectiveness (overall correct calls), miss rate (bad parts accepted), and false-alarm rate (good parts rejected). The two error types have asymmetric consequences - escaped defects reach the customer, false alarms generate scrap and rework.

Project implications:

1. Plating appearance cannot serve as a reliable Y until the system is fixed. Any defect rate the team measures now is partly inspector drift.
2. If a before/after comparison uses different inspectors, a change in the metric may reflect a change in inspectors, not the process.
3. Fixes: written acceptance criteria with limit samples/visual standards, controlled lighting and viewing distance, inspector training and qualification, periodic re-audit; or eliminate the judgment with a gage or automated vision check. Then re-run the study and require kappa ≥ 0.70 before baseline collection.