Process Performance vs Specifications

Process Performance vs Specifications (CSSGB BoK III.F.1 — Evaluate)

Quick Answer: Natural process limits (voice of the process) describe the range of variation the process produces—commonly estimated as μ ± 3σ. Specification limits (voice of the customer/design) are the allowed LSL and USL. Capability and performance analysis ask whether the process’s location and spread fit inside those specs. A process can be stable and still produce nonconforming product.

Two Different Voices

Green Belts constantly separate two questions that teams mix up:

1. What is the process doing? (process performance / natural variation)
2. What is allowed? (specifications / tolerance)

Critical exam trap: Control limits ≠ specification limits. Plotting specs on a control chart is allowed for context, but control decisions use control limits; conformance decisions use specs.

Natural Process Limits and Process Spread

Under a stable, approximately normal model, about 99.73% of individual observations fall within μ ± 3σ. That interval is often called the natural process limits or natural tolerance:

Natural process limits ≈ μ − 3σ to μ + 3σ
Natural process spread (process width) ≈ 6σ

In practice you estimate μ and σ from data (for example X̄ and s, or X̄ and R̄/d₂ for within-subgroup σ). The idea is the same: the process “claims” a band of values it will generate.

Worked sketch: Fill weight mean μ = 500.0 g, σ = 1.0 g.

• Natural limits ≈ 497.0 g to 503.0 g
• Process spread ≈ 6.0 g

If the customer allows 495–505 g, the natural band sits comfortably inside the specs if the mean stays at 500. If the mean drifts to 503 g, the upper natural limit moves to 506 g and nonconformances appear even though σ did not change.

Specification Limits and Tolerance

Specification limits are external requirements:

• LSL — lower specification limit
• USL — upper specification limit
• Target (T) — preferred value (often the midpoint, but not always)
• Tolerance / specification width — USL − LSL (two-sided)

One-sided specs (for example USL only on contamination, LSL only on strength) are common; then “width” is not a two-sided interval and capability formulas use the applicable one-sided form.

Specs answer conformance, not statistical control. Units outside LSL–USL are nonconforming regardless of how “in control” the chart looks.

Process Performance Metrics (Relative to Specs)

Process performance in this BoK sense means how the process behaves relative to requirements—location, spread, and the resulting nonconformance risk.

Metrics and views Green Belts use at Evaluate level:

1. Histogram or dot plot with LSL/USL overlaid — visual “does the mound fit the fence?”
2. Percent nonconforming — empirical fraction outside specs, or model-based estimate from normal (or other) assumptions.
3. Distance to nearest specification in sigma units — for a two-sided case, min[(USL − μ)/σ, (μ − LSL)/σ]. Larger is better.
4. Capability / performance indices (Cp, Cpk, Pp, Ppk, Cpm—developed in III.F.3) — formal ratios of tolerance to process variation, with or without centering and target.
5. DPMO / sigma level (linked to II.E.1) — long-run defect opportunity rates when the model fits.

Worked example — performance vs. specs: LSL = 10.0, USL = 16.0, process μ = 12.0, σ = 1.0 (stable, roughly normal).

• Spec width = 6.0
• Process spread = 6.0
• Natural limits ≈ 9.0 to 15.0 → lower tail crosses LSL
• Z_lower = (12.0 − 10.0)/1.0 = 2.0 → P(X < LSL) ≈ 0.0228
• Z_upper = (16.0 − 12.0)/1.0 = 4.0 → P(X > USL) ≈ 0.00003
• Roughly 2.3% nonconforming from the lower side alone—even though average “looks fine” inside the band.

The process is centered low relative to the two-sided window. Fixing performance may mean shifting the mean, reducing σ, or both—not “trying harder” on inspection alone.

Stable but Incapable vs. Capable but Unstable

Evaluate-level judgment: never treat a capability index from an out-of-control process as a durable customer promise. Performance vs. specs is only meaningful when you understand whether the process is in a state of statistical control (or you explicitly report short-term snapshot performance—see III.F.4).

Centering, Spread, and One-Sided Reality

Meeting specs is not only about 6σ versus USL − LSL:

• Spread problem: 6σ > USL − LSL → even a perfectly centered process cannot keep all (or nearly all) units inside specs under the normal model.
• Location problem: 6σ ≤ USL − LSL but μ is too close to one fence → Cpk/Ppk suffer; nonconformances cluster on one side.
• Target problem: specs may allow a wide band while the customer wants values near T (Taguchi loss)—Cpm and loss functions capture that (III.F.3).

Practical Evaluation Checklist

1. Are LSL/USL/T the correct, current requirements for this CTQ?
2. Are natural limits estimated from stable, representative data?
3. Is the distribution roughly normal (or handled with a proper method)?
4. Does the problem look like spread, centering, or both?
5. Are we confusing control limits with specs in team discussions?

Bottom Line for III.F.1

Process performance is what the process produces (location + variation). Specifications are the allowed window. Natural process limits summarize process voice; LSL/USL summarize customer/design voice. Your Evaluate job is to compare them honestly—visually and numerically—and decide whether the process can meet requirements as it currently runs, before you lean on formal indices or improvement projects.