The Marginal Gauge and the 900 ppm

Supported

• %GRR = 29% of tolerance is in the marginal band (10–30%). It is not automatically unacceptable, but it is too much measurement error to support a capability decision on a ±0.05 mm tolerance. Any Cpk quoted from this gauge carries material uncertainty.
• Reproducibility (22%) exceeds repeatability (18%). The dominant problem is between operators, not the gauge itself: differences in clamping force, where on the shaft the measurement is taken, or how the zero is set. This is cheap to fix - work instruction, fixture, retraining, and a verification step - before spending capital.
• ndc = 3 means the system resolves only three distinct categories across the parts studied. AIAG guidance wants ndc ≥ 5. On that evidence alone the gauge cannot distinguish good parts from marginal ones.

Not supported / wrong to draw

• "29% is under 30%, so we pass." The threshold is not a pass/fail line; it is a statement about risk. With %GRR near 30%, roughly a tenth of the observed variance may be measurement, and near a specification limit that shifts reject decisions by a meaningful amount.
• "ndc = 3 proves the gauge is bad." The 10 parts came from one hour, so part-to-part variation was artificially compressed. ndc is computed partly from part spread, so it is biased low. The %GRR-of-tolerance figure is unaffected by that sampling, so 29% is the number to trust - and the study must be repeated with parts spanning the real process range.
• "Repeatability 18% is fine." It is only fine relative to a 29% total; both need work.

Credibility problems, in order of severity

1. Capability requires a stable process. Cpk computed from three weeks of pooled data with no control chart is a Ppk in disguise if the process drifted. Run an I-MR or Xbar-R over the same period first; if it is unstable, no capability index is interpretable.
2. The measurement system attenuates the index. Observed variance ≈ true variance + measurement variance. With %GRR at 29% of tolerance, part of the spread used in Cpk is gauge noise, so observed Cpk is biased. The direction of the bias on the true Cpk depends on whether the measurement error is random (inflates σ → understates Cpk) or includes operator bias (can shift the mean).
3. Two gauges, one studied. The 300 points mix shifts and mix gauges. If the older night gauge reads low by even 0.005 mm, the pooled σ and mean are contaminated and the Cpk describes no real process.
4. Cpk vs. observed ppm. Cpk 1.12 on a two-sided tolerance implies a defect rate on the order of a thousand ppm - which is broadly consistent with 900 ppm, so the numbers do not contradict each other. That consistency is not reassurance; it means the process genuinely sits where the customer says it does.

What to report at the gate

• Process performance is roughly at the voice of the customer (≈900 ppm), not better than it.
• The capability figure is provisional because the measurement system is marginal and the studies are not comparable across shifts.
• Measure gate status: conditional pass - Analyze may start on non-diameter questions, but no conclusion about diameter capability or about the effect of any future improvement is defensible until the MSA is redone.

Measurement system

• Retrain all operators on a written method: measurement location, clamp torque, zeroing, part temperature soak. Add a fixture if the parts can rotate in the jaws.
• Repeat the crossed R&R with 10 parts spanning the full production range (select from the last 4 weeks of output, not one hour), 3 operators, 3 trials, randomized order, operators blind to part identity.
• Acceptance before proceeding: %GRR ≤ 10% of tolerance (≤ 30% with a documented interim justification and an explicit statement of how decisions will be made), ndc ≥ 5, and no operator with a mean bias against the reference.

Gauge equivalence (the neglected half)

• Run a 30-piece study comparing the two gauges against a reference/calibrated standard, same parts, same operator, blind. If the bias is not statistically and practically negligible, either qualify the night gauge or standardize on one gauge and re-baseline.

Ongoing data

• Sample by shift, machine, and part family - a stratified plan, not 30 pieces from whichever bin is nearest. Record subgroup, time, operator, gauge ID, and lot.
• Minimum 25 subgroups of n = 5 before any capability claim; keep the rational subgrouping consistent (consecutive parts within a short window).
• Add a short-term, high-frequency check (5 pieces/hour for one shift) to separate within-lot from between-lot variation.

Why this order: every downstream tool - hypothesis tests, control charts, capability - inherits the measurement error. Redoing the MSA costs about two days; discovering at Control that the savings were gauge noise costs the project.