# Six Sigma

**URL:** https://kaizumi.com/dictionary/six-sigma

**Description:** Six Sigma is a statistical quality framework whose reported capability levels depend directly on how defect opportunities are counted.

**Category:** six-sigma

**Tags:** quality, metrics, foundational

## Counting | What a sigma level actually counts

Six Sigma measures how often a process fails to deliver what a customer specified, scaled against the number of chances it had to fail. It does not measure percentage yield, it does not count defective items, and it does not describe quality in the abstract. It calculates defects per million opportunities, looks up the corresponding distance from the process mean to the nearest specification limit in standard deviations, and adds an artificial constant of `1.5`.

Four terms must be separated before calculating anything:

1. **A unit** is the item being inspected: one circuit board, one insurance claim, one machined valve body.
2. **A defect** is any specific instance of failing a customer requirement: a missing solder fillet, an incorrect postcode, a scratch deeper than specification.
3. **An opportunity** is a distinct, measurable chance for a defect to occur on that unit.
4. **A defective unit** is a unit carrying at least one defect. A circuit board with `3` cold joints and `1` incorrect resistor is `1` defective unit carrying `4` defects.

Six Sigma counts defects, not defectives. A process that builds `100` units where `1` unit contains `20` wiring faults has `20` defects. A process that builds `100` units where `20` units contain `1` fault each also has `20` defects. Traditional yield treats the first batch as `99` percent good and the second as `80` percent good. Six Sigma treats their defect rate as identical because the count of operational failures is identical.

Take an electronics line shipping `20,000` circuit boards in a month. Inspection detects `130` defects across the run. Evaluating the board as a single pass-fail test (`1` opportunity per unit) gives:

```kz-formula
expr: DPMO = (D ÷ (U × O)) × 1,000,000
var: D | **Defects counted.** Total failures found across the inspection run.
var: U | **Units inspected.** Total batch volume evaluated over the period.
var: O | **Opportunities.** Designated failure chances per unit.
var: DPMO | **Defects per million opportunities.** Standardised failure rate.
```

`DPMO = (130 ÷ (20,000 × 1)) × 1,000,000 = 6,500`

A failure rate of `6,500` parts per million equals a yield of `99.35` percent. Under the standard normal distribution, a single-tailed tail probability of `0.0065` sits `2.48` standard deviations from the mean (`Z = 2.48`). Six Sigma reporting convention adds a mandatory `1.5` sigma offset to convert this short-term statistic into a nominal long-term capability score.

```kz-tiles
lead: Sigma level
tile: DPMO | 6,500 | failures / M
tile: Yield | 99.35 | percent
tile: Sigma level | 3.98 | long-term
```

```kz-formula
expr: Sigma level = Z + 1.5
var: Z | **Z-score.** Standard deviations from the mean to the specification limit on a standard normal curve.
var: 1.5 | **The shift constant.** The assumed long-term process drift adopted by Motorola.
```

`Sigma level = 2.48 + 1.5 = 3.98`

When teams describe an operation as a "four-sigma process", this is the calculation they use.

| Sigma level | Long-term DPMO | Equivalent yield | Defect drop to next level |
|---|---|---|---|
| `3` sigma | `66,807` | `93.32` percent | — |
| `4` sigma | `6,210` | `99.379` percent | `-91` percent |
| `5` sigma | `233` | `99.9767` percent | `-96` percent |
| `6` sigma | `3.4` | `99.99966` percent | `-98.5` percent |

```kz-figure
id: sigma-ladder
caption: The non-linear relationship between sigma levels and defects. Moving from 3 to 4 sigma removes 60,597 defects per million. Moving from 5 to 6 sigma removes just 229.6 defects per million, but requires cutting the remaining failure rate by 98.5 percent.
wide: true
```

Between `3` sigma and `4` sigma, yield moves by `6.06` percentage points (`93.32` to `99.379` percent). Between `5` sigma and `6` sigma, yield moves by `0.023` percentage points (`99.9767` to `99.99966` percent). 

Percentages lose visual resolution where reliable processes operate. A dashboard reporting `99.9` percent yield lumps processes running at `1,400` DPMO and `700` DPMO into the same single-decimal figure, hiding a two-fold difference in customer-bound failures. DPMO and sigma levels preserve numerical resolution in the tail of the distribution.

## The denominator | Why one line scores 3.98 and 5.09 sigma

A sigma level is not an objective physical property of a line; it depends on how the opportunity count is defined. Two engineers inspecting the same hardware with identical scrap bins can produce different capability scores by choosing different denominators.

Take the electronics line shipping `20,000` boards with `130` solder flaws.

```kz-compare
leftTitle: Rule A — Unit-level opportunity
left: `20,000` units inspected
left: `1` opportunity per unit (board passes or fails)
left: `20,000` total opportunities evaluated
left: `130` defects found
left: `DPMO = 6,500`
left: `Z = 2.48`
left: **Sigma level = `3.98`**
rightTitle: Rule B — Component-level opportunity
right: `20,000` units inspected
right: `40` opportunities per unit (40 solder joints per board)
right: `800,000` total opportunities evaluated
right: `130` defects found
right: `DPMO = 162.5`
right: `Z = 3.59`
right: **Sigma level = `5.09`**
```

Rule A treats the finished product as the deliverable: the customer receives a functional board, and any fault renders it defective. Rule B treats the assembly process as the deliverable: the soldering machine forms `40` independent joints per board, each representing a chance to succeed or fail.

Switching from Rule A to Rule B cuts recorded DPMO from `6,500` to `162.5` and raises the sigma level from `3.98` to `5.09` — an apparent gain of `1.11` sigma achieved without touching a machine parameter or reducing defects by a single unit.

```kz-figure
id: opportunity-scale
caption: The denominator effect. The physical defect count remains fixed at 130 across 20,000 units. Expanding the designated opportunity count from 1 to 40 inflates the denominator by 40x, compressing DPMO from 6,500 to 162.5 and moving the sigma score by 1.11 sigma on paperwork alone.
wide: true
```

To earn that same `1.11` sigma improvement under Rule A with the denominator fixed at `1`, the team would need to cut monthly defects from `130` to `3.25` — a `40`-fold reduction in defect generation.

| Opportunities counted per board | Basis of definition | Resulting DPMO | Yield equivalent | Sigma level |
|---|---|---|---|---|
| `1` | Entire board assembly | `6,500` | `99.35000` percent | `3.98` |
| `4` | Major circuit quadrants | `1,625` | `99.83750` percent | `4.44` |
| `12` | Integrated circuit packages | `542` | `99.94580` percent | `4.77` |
| `40` | Individual solder joints | `162.5` | `99.98375` percent | `5.09` |

This denominator spread makes cross-company sigma comparisons meaningless unless both organisations share identical operational definitions of an opportunity.

```kz-callout
tone: warn
title: A sigma level is comparable only to itself over time.
p: An organisation can arbitrarily inflate its sigma score by decomposing every assembly into microscopic opportunities — counting every thread, lead, label, and packaging tab.
p: Define your opportunity count once at project kickoff, document the boundaries in standard work, and never adjust the denominator to show progress. Progress must appear in the numerator.
```

Inflating the denominator also degrades metric sensitivity. At `40` opportunities per unit, eliminating `10` defects shifts DPMO by only `12.5`. At `1` opportunity per unit, those same `10` defects shift DPMO by `500`.

## The shift | Where 3.4 defects per million comes from

The widely quoted figure of `3.4` defects per million opportunities is not the tail area of a centred normal distribution at six standard deviations.

For a centred normal distribution midway between the Upper Specification Limit (USL) and Lower Specification Limit (LSL), the clearance to each limit is `6.0` standard deviations. The area beyond `6.0` sigma in a single tail is `0.00099` parts per million. Summing both tails yields `0.00197` DPMO — roughly `2` defects per billion opportunities.

```kz-pass
title: Centred process — true statistical 6 sigma
step: Mean centered at nominal specification: **distance to limit = 6.0 σ**
step: Defect probability per tail: **0.00099 per million** (0.99 ppb)
out: Both tails combined: **0.00197 DPMO** (~2 defects per billion)
```

```kz-pass
title: Shifted process — Six Sigma reporting standard
step: Process mean drifts over time: **assumed shift = 1.5 σ**
step: Distance to nearest specification limit: **6.0 σ - 1.5 σ = 4.5 σ**
out: Tail beyond 4.5 σ: **3.3977 DPMO** (conventionally rounded to 3.4)
```

The value `3.4` comes from shifting the process mean towards one limit by `1.5` standard deviations, reducing clearance to the near limit from `6.0` sigma to `4.5` sigma. The probability in a single normal tail beyond `4.5` sigma is `0.0000033977`, or `3.3977` per million. The far limit is `7.5` sigma away, where the tail area is negligible. Quality literature rounds `3.3977` to `3.4`.

```kz-figure
id: the-shift
caption: The 1.5 sigma shift mechanics. The top curve shows a perfectly centred distribution with 6.0 sigma clearance to both limits, yielding 2 defects per billion. The bottom curve shows the mean drifted by 1.5 sigma, reducing the near-tail clearance to 4.5 sigma and creating 3.4 defects per million.
wide: true
```

Motorola adopted the `1.5` sigma offset empirically in the `1980`s. Short-term capability studies capture only immediate machine variation, whereas long-term production encounters tool wear, raw material batch changes, ambient temperature swings, and calibration drift. Motorola introduced the constant 1.5 sigma shift in the 1980s as an empirical allowance for long-run process drift, though it has no general mathematical derivation.

There is no mathematical law dictating that processes drift by `1.5` standard deviations. In practice, physical drift varies substantially across processes: tightly controlled mechanical operations may remain exceptionally stable over time, while batch chemical processes can experience pronounced shifts between raw material feeds.

In classical statistical process control, process capability is expressed as `Cpk`:

`Cpk = (USL - Mean) ÷ (3 × σ)  [for upper limit clearance]`

Under this index:
- `Cpk = 1.00` corresponds to `3.0` sigma clearance.
- `Cpk = 1.33` corresponds to `4.0` sigma clearance.
- `Cpk = 1.50` corresponds to `4.5` sigma clearance.
- `Cpk = 2.00` corresponds to `6.0` sigma clearance.

Six Sigma programs define the reported sigma level as `3 × Cpk`.

When a line runs at `Cpk = 1.33`, true clearance to the limit is `4.0` sigma. If it remains centred, it produces roughly `32` DPMO (single tail). Under Six Sigma convention, that `Cpk = 1.33` process is assigned a nominal "`4.0` sigma level", which assumes the clearance to the near limit falls to `2.5` sigma after a `1.5` sigma drift and generates `6,210` DPMO in standard lookup tables.

<!--TOOL:process-capability-->

```kz-callout
tone: note
title: Measure actual drift instead of assuming 1.5 sigma.
p: Calculate short-term capability (`Cp` / `Cpk`) and long-term performance (`Pp` / `Ppk`) directly from multi-month subgroup data.
p: The difference between short-term and long-term Z-scores is the true operational drift. Using an arbitrary `1.5` constant indicates an absence of long-term control chart data.
```

## The hidden factory | What a 99.35 percent final yield hides

Final yield measures only what passes inspection on the final attempt. It conceals internal loops of disassembly, touch-up, re-soldering, and re-testing that occur between the first station and packaging.

This rework loop is [the hidden factory](/dictionary/hidden-factory). It consumes capacity, adds indirect labour, masks process instability, and introduces latent reliability flaws into reworked units.

Consider `20,000` circuit boards passing through five sequential operations:

| Process step | Units entering | Units reworked | First-pass yield |
|---|---|---|---|
| `1`. Form | `20,000` | `300` | `98.50` percent |
| `2`. Weld | `20,000` | `160` | `99.20` percent |
| `3`. Coat | `20,000` | `520` | `97.40` percent |
| `4`. Assemble | `20,000` | `40` | `99.80` percent |
| `5`. Test | `20,000` | `180` | `99.10` percent |

To assess true line capability, calculate [first-pass yield](/dictionary/first-pass-yield) at each operation and multiply them to get Rolled Throughput Yield (RTY):

```kz-formula
expr: RTY = Y1 × Y2 × Y3 × Y4 × Y5
var: Y1 | **Form yield.** (20,000 - 300) ÷ 20,000 = 0.9850.
var: Y2 | **Weld yield.** (20,000 - 160) ÷ 20,000 = 0.9920.
var: Y3 | **Coat yield.** (20,000 - 520) ÷ 20,000 = 0.9740.
var: Y4 | **Assemble yield.** (20,000 - 40) ÷ 20,000 = 0.9980.
var: Y5 | **Test yield.** (20,000 - 180) ÷ 20,000 = 0.9910.
var: RTY | **Rolled Throughput Yield.** Probability that a unit clears all five steps without rework.
```

`RTY = 0.9850 × 0.9920 × 0.9740 × 0.9980 × 0.9910 = 0.9413 = 94.13 percent`

Across the five steps, operators performed `1,200` separate rework actions (`300 + 160 + 520 + 40 + 180`). Assuming defects are distributed independently, only `18,825` boards cleared all five operations untouched (`20,000 × 0.9413`), while `1,175` boards required rework intervention.

```kz-tiles
lead: Final yield
tile: Final yield | 99.35 | percent
tile: Rolled yield | 94.13 | percent
tile: Capability gap | 0.91 | sigma
```

```kz-figure
id: hidden-factory
caption: The hidden factory visualized across five steps. Final yield measures only the output of Step 5 (99.35 percent). Rolled Throughput Yield compounds the losses of all five operations (94.13 percent), exposing 1,200 rework events that final inspection metrics conceal.
wide: true
```

The difference between final yield and rolled throughput yield transforms the capability evaluation:
- Final yield of `99.35` percent corresponds to `6,500` DPMO (`3.98` sigma).
- Rolled throughput yield of `94.13` percent corresponds to `58,737` DPMO (`3.07` sigma).

The hidden factory masks a capability gap of `0.91` sigma. 

Calculating the normalised step yield gives:

`Normalised yield = RTY^(1/5) = (0.9413)^(0.2) = 0.9880 = 98.80 percent`

Each step averages `98.80` percent first-pass yield (`12,000` DPMO, or `3.76` sigma). Compounding five steps reduces total line output to `94.13` percent. If the line contained `50` sequential steps at this capability, rolled throughput yield would fall to `0.9880^50 = 54.7` percent — over `45` percent of production would require rework despite every supervisor reporting individual step yields near `99` percent.

## The project | DMAIC, belts and what a project costs

Six Sigma executes improvement through DMAIC (Define, Measure, Analyse, Improve, Control). Unlike rapid continuous improvement events that adjust physical workstations within a shift, DMAIC investigates root causes masked by interacting process variables.

| Phase | Core activity | Concrete deliverable |
|---|---|---|
| **Define** | Identify the defect, define scope, state customer requirements. | Charter with problem statement and baseline metrics. |
| **Measure** | Map the process, validate measurement systems, collect continuous data. | Gage R&R report (measurement error `< 10` percent); baseline capability. |
| **Analyse** | Formulate hypotheses, isolate critical input variables (`X`s), test significance. | ANOVA tables, regression models identifying root causes (`Y = f(X)`). |
| **Improve** | Formulate countermeasures, run designed experiments (DOE), verify reduction. | Pilot run data demonstrating verified shift in mean and variance. |
| **Control** | Standardise process parameters, implement SPC, hand over to operations. | Standard Operating Procedures, control charts, process owner sign-off. |

Attribute data (pass/fail counts) becomes impractical for verification as quality improves because required sample sizes grow exponentially.

| Target sigma level | Associated DPMO | Average units produced per single defect |
|---|---|---|
| `3` sigma | `66,807` | `15` |
| `4` sigma | `6,210` | `161` |
| `5` sigma | `233` | `4,299` |
| `6` sigma | `3.4` | `294,319` |

At `6` sigma, finding a single defect requires inspecting an average of `294,319` consecutive units. Constructing a statistically valid hypothesis test using pass/fail counts at that level requires observing dozens of failures across millions of parts.

You cannot count your way to Six Sigma.

For this reason, DMAIC projects rely on continuous variables data. Instead of classifying a solder joint as good or bad, the project measures continuous physical parameters: wetting angle in degrees, paste volume in cubic micrometres, reflow peak temperature in degrees Celsius, and conveyor speed in millimetres per second. Continuous data lets sample sizes of `30` to `50` parts establish changes in mean and spread with high statistical power.

Continuous data collection requires evaluating measurement tools through a Gage Repeatability and Reproducibility (Gage R&R) study. If measurement error consumes more than `30` percent of the specification band, the data captures gauge variation rather than process variation.

```kz-callout
tone: warn
title: Never analyse process data before passing Gage R&R.
p: If operators cannot measure the same part twice and get the same value (repeatability), or if two inspectors record different numbers on the same feature (reproducibility), process data is invalid.
p: Running statistical analysis on unvalidated measurement tools models gauge noise rather than production tooling.
```

<!--TOOL:gage-rr-->

Organisations structure project roles into hierarchical belt tiers:

- **Yellow Belt:** Baseline operational contributor. Understands core terminology, participates in data collection, and supports project teams at the gemba.
- **Green Belt:** Part-time project leader. Typically an engineer, supervisor, or operations lead who runs `1` to `2` DMAIC projects per year within their functional area.
- **Black Belt:** Full-time technical specialist. Trained in Design of Experiments, regression, and multivariate hypothesis testing. Black Belts lead cross-functional projects and mentor Green Belts.
- **Master Black Belt:** Technical advisor and coach. Manages program governance, audits statistical standards, and consults on complex modeling.
- **Champion / Sponsor:** Executive leader with budgetary authority. Charters projects, removes roadblocks, approves equipment modifications, and validates financial closure.

No single standardisation body governs Six Sigma certification. The American Society for Quality (ASQ), the International Association for Six Sigma Certification (IASSC), commercial consultancies, and individual companies each maintain separate certification criteria.

## Fit | When Six Sigma is the wrong tool

DMAIC is designed for complex problems with unknown root causes. Applying full statistical architecture to obvious failure modes adds project overhead without adding insight.

```kz-compare
leftTitle: Reach for Six Sigma when
left: The root cause of variation is genuinely unknown
left: Multiple variables interact simultaneously (`Y = f(X_1, X_2, X_n)`)
left: The process is already stable and in basic control
left: Historical improvements have plateaued
left: High-volume continuous data is affordable to collect
left: Defect volume justifies a multi-month investigation
rightTitle: Reach for something else when
right: The root cause is visible on the shop floor
right: A single physical failure mode exists (use [5 Whys](/dictionary/5-whys))
right: The line lacks standard work or basic stability
right: The defect can be designed out physically (use [Poka-Yoke](/dictionary/poka-poka))
right: The primary issue is queue time, inventory, or flow
right: Defect volume is too small to justify project overhead
```

If an operator occasionally inserts a connector backwards, do not run a four-month project and build an ANOVA table. Machine an asymmetric guide pin onto the fixture so the connector physically cannot enter backwards. Mistake-proofing ([poka-yoke](/dictionary/poka-yoke)) eliminates the opportunity entirely rather than managing its probability.

Statistical characterisation requires a stable baseline. If a cell lacks [standard work](/dictionary/standard-work), if operators follow different routines, or if maintenance occurs only after breakdowns, common-cause variation is drowned out by special causes. The resulting data is non-stationary.

Standardise work sequences, stabilise maintenance, and balance flow before calculating statistical capability.

```kz-callout
tone: note
title: Establish Lean flow before calculating Six Sigma capability.
p: Eliminating lead time and excess work-in-progress inventory exposes defects minutes after generation rather than weeks later at final test.
p: Use Lean methods to establish standard work and continuous flow; deploy Six Sigma statistical tools when multivariate variation persists in stable processes.
```

The final criterion is project economics. A DMAIC project consumes engineering time, machine availability for designed experiments, and metrology resources. If a defect mode causes minor scrap that can be resolved with standard maintenance or simple tooling fixtures, chartering a full analytical project is an inefficient use of engineering capacity.

## Straight answers | Six Sigma questions people ask

```kz-qa
q: What is Six Sigma?
a: Six Sigma is a data-driven quality methodology designed to reduce process variation and eliminate defects. It uses statistical analysis, structured project execution (DMAIC), and a standardised metric — defects per million opportunities (DPMO) — to achieve predictable, highly capable processes.
q: What does 3.4 defects per million mean?
a: It is the defect rate that occurs when a process with normal variation experiences a `1.5` standard deviation shift in its mean, leaving `4.5` standard deviations of clearance between the mean and the nearest specification limit. A perfectly centred six-sigma process produces just `2` defects per billion opportunities.
q: How do you calculate a sigma level?
a: Count total defects (`D`), total units inspected (`U`), and defect opportunities per unit (`O`). Compute `DPMO = (D ÷ (U × O)) × 1,000,000`. Determine the single-tail normal distribution value `Z` corresponding to that defect rate, and add `1.5`. A rate of `6,500` DPMO gives `Z = 2.48`, which equals a `3.98` sigma level.
q: What is DPMO?
a: Defects Per Million Opportunities. It standardises failure rates by dividing the defect count by the total opportunities evaluated across the entire batch, multiplied by `1,000,000`. It allows processes of varying complexity to be compared using a common mathematical denominator.
q: How do you determine the number of opportunities per unit?
a: You define opportunities by identifying distinct, measurable features where execution can fail a customer specification. A solder joint, a critical diameter, and a label position can each be counted as `1` opportunity. The choice of denominator is an operational definition: expanding the count inflates the resulting sigma level without reducing actual defects.
q: Why is there a 1.5 sigma shift?
a: Motorola engineers introduced the `1.5` sigma shift in the `1980`s to account for long-term operational drift caused by tool wear, temperature changes, raw material variation, and machine aging. It is an empirical convention rather than a mathematically derived natural law.
q: What are the Six Sigma belt levels?
a: Belts designate levels of methodological expertise. Yellow Belts understand basic concepts and support data collection; Green Belts lead part-time projects in their functional areas; Black Belts lead full-time multivariate projects and mentor teams; Master Black Belts manage enterprise deployments and audit statistical standards.
q: How does Six Sigma differ from Lean?
a: Lean focuses on eliminating non-value-added waste, shortening lead times, and optimizing flow using visual tools and continuous kaizen on shop-floor clocks. Six Sigma focuses on reducing variation, improving capability, and eliminating defects using statistical hypothesis testing on multi-month project clocks.
q: Is Six Sigma only applicable to manufacturing?
a: No. Transactional processes — such as healthcare records management, financial invoicing, software deployment pipelines, and customer service ticket resolution — can be evaluated with Six Sigma if they have clear customer specifications, measurable defect criteria, and continuous or high-volume discrete data.
q: Did Motorola and General Electric actually save billions of dollars with Six Sigma?
a: Both corporations published multi-billion dollar savings figures during their commercial rollouts in the `1980`s and `1990`s. Those savings figures were self-reported internal accounting metrics that included standard operational cost reductions, and neither figure was ever independently audited by third-party accounting firms.
```

Source: https://kaizumi.com/dictionary/six-sigma
Licence: free to quote and cite with attribution to Kaizumi.