DMAIC¶

If you already run DMAIC (Define, Measure, Analyze, Improve, Control) efforts, this page shows where each part of a DMAIC metrics document comes from in an impact model, how to mark the programs that suit DMAIC, and what the impact model checks before it lets you start.

impact-model export dmaic impact-model.yaml -O observations.yaml -o dmaic.json

The output is a DMAIC metrics document in the JSON format of prism-roadmap's goals/dmaic package, which validates it and can render it as slides or as Markdown for PDF.

Which programs suit DMAIC¶

DMAIC improves a process with statistics: control charts need many measurements taken on a steady schedule. A program whose numbers arrive weekly, such as an online course catalog with completion rates and a post-course survey, suits it. A program measured once a year does not: twenty points would take twenty years.

So DMAIC is opt-in, per pathway. Mark a pathway, the impact model's program or product, with projections:

pathways:
  - id: pw-catalog
    title: Course catalog
    projections:
      - dmaic

Only marked pathways are checked by the DMAIC lint rules and exported.

What lint checks first¶

impact-model lint checks every indicator of a marked pathway:

Rule	Checks	Why
`IMP-DMA-001`	It has a `frequency` and a data source (`dataSource` or `collection.system`)	Without them nothing is measured on a schedule
`IMP-DMA-002`	It is measured at least 20 times a year	Control limits need 20 to 25 points; fewer a year means waiting years
`IMP-DMA-003`	A survey indicator gets at least 30 responses per measurement	Below 30 a score swings with chance; an NPS needs about 100 for a margin near ±10 points

Responses per measurement are collection.respondentsPerYear, times collection.sampleRate when set, divided by measurements per year. A weekly survey of 6,500 learners a year gives about 125 per week. The same survey of 780 learners gives 15, and lint suggests measuring every month, pooling across courses, or asking more people.

The rules are warnings, not errors: an indicator that fails them can still be exported, but its chart will not be ready for a long time.

The mapping at a glance¶

DMAIC	Comes from	In impact model terms
Category	each pathway marked `dmaic`	A program or product: a group of outcomes and the interventions that bring them about
Metric	each indicator of the pathway with a numeric baseline and target	An indicator measures one outcome or intervention
Metric baseline, target	`baseline.value`, and the value of the latest-dated target
Metric current	the latest observation, or the baseline before any
Metric data points	the indicator's observations, given with `-O`	Observations are dated values recorded against an indicator
Metric control limits	computed from the data points, only from 20 or more
Metric trend direction	`direction`: increase, decrease, or maintain
Metric status	the indicator's health on the status date: Green, Yellow, Red
Metric phase	derived; see Phases
Metric root causes	the causes the measured outcome addresses	The problem tree: `causes`, each immediate, underlying, or structural
Initiatives	the pathway's interventions aimed at the measured outcome	An intervention is something the organization does to change an outcome

One metric, end to end¶

The course catalog example is a fictional nonprofit's free on-demand certificate courses. Its North Star is the weekly completion rate:

indicators:
  - id: ind-completion-rate
    measures: oc-mt-certified
    name: Course completion rate
    unit: percent
    direction: increase
    baseline: { value: 38, date: "2025-12-29" }
    targets:
      - { value: 45, date: "2026-12-31" }
      - { value: 50, date: "2027-12-31" }
    dataSource: Learning platform
    frequency: P1W

With 24 weeks of observations it becomes this metric, shown trimmed:

{
  "id": "ind-completion-rate",
  "name": "Course completion rate",
  "baseline": 38,
  "current": 40.9,
  "target": 50,
  "unit": "percent",
  "phase": "Improve",
  "trendDirection": "higher_better",
  "controlLimits": { "ucl": 42.855, "lcl": 35.987, "centerLine": 39.421, "sigma": 1.145 },
  "frequency": "Weekly",
  "dataSource": "Learning platform",
  "status": "Green",
  "dataPoints": [
    { "timestamp": "2026-01-05T00:00:00Z", "value": 38 },
    { "timestamp": "2026-01-12T00:00:00Z", "value": 39 }
  ],
  "rootCauses": [
    { "id": "cause-no-time", "description": "Staff cannot set aside an hour at a time for a course.", "category": "Immediate", "validated": true }
  ],
  "initiativeIds": ["int-course-redesign"]
}

The baseline is the indicator's baseline and the target its 2027 target, the latest one.
The data points are the 24 weekly observations, and the current value is the last of them.
With 24 points, the control limits are computed.
The outcome it measures, learners completing courses, addresses the cause "Staff cannot set aside an hour at a time". The cause cites evidence, so it is marked validated.
The intervention aimed at that outcome, shortening modules, is in progress, so the metric is in the Improve phase and lists it as its initiative.

Data points and the current value¶

Data points are the indicator's observations from the files given with -O, in date order. An observation broken down by a group, with a disaggregation such as member size: small, describes only that group, so it is not a point on the total's chart and is left out.

The current value is the latest data point. Before there are any, it is the baseline.

Control limits¶

Control limits are those of an individuals chart (I-MR), the usual chart for one measurement per period:

the center line is the mean of the data points
sigma is estimated as the average moving range, the mean difference between consecutive points, divided by 1.128
the upper and lower limits are the center line plus and minus three sigma

They are computed only from 20 or more data points. With fewer the metric has no limits, because limits from a handful of points mislead more than they help.

A point outside the limits is a signal that the process has changed. A metric that is improving on purpose will cross a limit; that is the improvement showing, and the limits should be recomputed from the new level.

Process capability (Cp, Cpk, sigma level) is never filled. It compares the process with specification limits, the bounds a result must stay within, and an impact model records targets, not specification limits.

Phases¶

DMAIC tracks which phase each metric is in. The impact model does not record a phase, so the export derives one, as a convention:

Phase	When
Control	The indicator has achieved its target on the status date. Hold the gain
Measure	There are fewer than 20 data points, too few for control limits
Improve	An intervention aimed at the measured outcome is in progress
Analyze	Otherwise: enough data, the target not yet met, and nothing in progress to change it

Define is never assigned: a metric with a baseline and a target has already been defined.

Root causes and initiatives¶

A metric's root causes are the causes its outcome addresses, from the model's problem tree. Each cause's level becomes the category, such as Immediate or Structural, and a cause that cites evidence is marked validated.

The document's initiatives are the interventions of the marked pathways that aim at a metric's outcome, or that the metric measures directly. Each lists the metrics it should move, and its expected impact names the outcomes it targets.

Initiative status	From the intervention
Planned	`planned`
In Progress	`in_progress`
Completed	`completed`
Cancelled	not exported: a cancelled intervention is not an initiative

A paused intervention is exported with no status. External links come from initiativeRefs, such as a VisionStudio or Jira item.

What is left out¶

Left out	Why
An indicator with no numeric baseline or target, such as a qualitative rating	A DMAIC metric is a number moving from a baseline to a target. Each is named on standard error
Pathways not marked `dmaic`	They are not DMAIC efforts
The causal links, assumptions, and evidence	`report pathway` and `report assumptions` show them
Thresholds and process capability	The model records targets, not warning thresholds or specification limits

When the export is refused¶

export dmaic writes nothing and exits with status 2 when no pathway is marked dmaic, or when no indicator of a marked pathway has a numeric baseline and target:

impact-model: impact-model.yaml: the impact model cannot be drawn as DMAIC: mark a pathway with projections: [dmaic]