A Metric Is Useful Only When It Can Change a Decision

Start with the choice, not the available data

A metric earns attention when a named person can explain which choice it informs and when that choice can still change. Page views, conversion, throughput, complaints, or refresh age may all be useful, but none has an inherent decision meaning. Begin by writing the decision, its owner, the latest useful review date, and the action options. Then ask which observation would distinguish those options. This reverses the common habit of displaying everything a platform makes easy to count.

DORA’s performance guidance uses balanced measures and warns against cross-context ranking and metric gaming. Editorial and commercial work differ from software delivery, so its measures should not be copied as universal targets. The transferable principle is to connect a small set of observations to system improvement while preserving context and trade-offs. A dashboard with no decision contract is an information display, not a management system.

Evidence: DORA / Google Cloud

Give leading, lagging, and safeguard measures different jobs

Leading measures indicate whether a mechanism is present before the final outcome arrives. Lagging measures describe the later result. Safeguards expose costs or harms that a favorable headline could hide. For an editorial workflow, work-item age may warn about congestion, reader task completion may describe a nearer outcome, and correction volume may protect against speed at the expense of accuracy. These roles should remain separate rather than being averaged into one score.

The Kanban Guide defines a compact set of flow measures and links them to active workflow management. That supports using work in progress, throughput, age, and cycle time for specific flow choices. It does not establish reader value or business impact. Likewise, a revenue measure cannot explain which workflow change caused it. A balanced metric set should cover the hypothesized mechanism, intended outcome, and unacceptable downside without pretending the relationship is proven.

• Leading: is the intended mechanism operating?
• Lagging: did the consequential outcome later move?
• Safeguard: what cost, exclusion, or instability could the headline conceal?

Evidence: DORA / Google Cloud; Kanban Guides

Set cadence from decision horizon and evidence latency

Review frequency should match how quickly evidence becomes interpretable and how long the decision remains reversible. A queue that can block within days may deserve a weekly age review. Search discovery or retained learning may require longer observation. Looking daily at a slow, noisy outcome creates pressure to react to ordinary variation; reviewing a fast operational risk monthly may arrive too late. More frequent is not automatically more responsible.

Map three times for each metric: collection interval, maturation delay, and action horizon. The meeting cadence must respect all three. If the evidence is incomplete, the correct outcome may be “wait until the next maturity date,” provided safeguard conditions remain acceptable. How People Learn II adds an independent reason for reflection and evidence-informed adjustment, but it also reinforces that context matters; a cadence cannot be validated simply because another field uses it.

Evidence: Kanban Guides; National Academies of Sciences, Engineering, and Medicine

Precommit thresholds without treating them as truth

A threshold defines when a metric prompts investigation or a specific action. Write it before the review and include a range or confidence label when precision is weak. Also define what would invalidate the number: tracking changes, small denominators, missing segments, seasonality, or a shifted product. Thresholds are decision policies, not natural facts. They should be revised with a dated reason rather than moved silently after an inconvenient result.

GOV.UK’s alpha guidance illustrates setting success measures while testing risky assumptions. Apply that discipline by separating a go/no-go measure from supporting diagnostics and by naming contradictory evidence. A threshold can prevent retrospective storytelling, yet it cannot rescue a poor proxy. If the measure does not bear on the intended mechanism, improve the observation before debating the cutoff.

Evidence: DORA / Google Cloud; UK Government Digital Service

Artifact: the metric-to-decision register

Create one row per reviewed metric with these fields: decision, owner, action options, metric definition, source, segment, update frequency, maturation delay, expected range, alert or decision threshold, safeguard, data-quality check, next review, and reconsideration trigger. Add the last action taken and whether the expected follow-up appeared. A metric with a blank decision or owner should not enter the recurring review.

The register is intentionally different from a dashboard. The dashboard shows current observations; the register preserves why they matter and how interpretation is governed. During review, open the register first, inspect data quality, then view only the relevant chart. Archive retired rows rather than deleting them so later reviewers can understand why a number stopped driving action.

Evidence: DORA / Google Cloud; Kanban Guides; UK Government Digital Service

Measurement cannot remove uncertainty or responsibility

Metrics omit experience, can be gamed, and may shift behavior toward what is easiest to count. Small samples and multiple comparisons can produce unstable patterns. Data quality can change without a visible warning. A sound cadence includes qualitative evidence, source checks, and a route to pause when rights, privacy, accessibility, or safety are implicated. It also permits retiring a metric whose decision value is lower than its collection or review burden.

For a first application, choose one upcoming decision and complete a single register row before building a new chart. Review it on the date when evidence can mature, record the action or justified wait, and revisit the threshold after the outcome window. This is an illustrative governance method, not a TenMultigure performance claim or statistical standard. Schedule full source and policy review by 2027-02-10.

Evidence: DORA / Google Cloud; National Academies of Sciences, Engineering, and Medicine