Cover photo: Dylan Gillis / Unsplash. Illustrative photograph, not a record of a client engagement.

A founder needs to know whether the team can deliver useful changes safely, where work gets stuck and what needs a decision. Counting commits answers none of those questions reliably. One change may delete an unnecessary feature; another may add thousands of lines without improving anything for a customer.

Start with one service, a consistent reporting window and a few agreed definitions. Read delivery evidence alongside a customer outcome and operational context. The scorecard earns its place when it changes what you do next.

DORA now describes five delivery metrics

The current DORA guide, updated 5 January 2026, describes five software delivery performance metrics:

  • Change lead time: elapsed time from a committed change to production.
  • Deployment frequency: how often production deployments occur.
  • Failed deployment recovery time: how long recovery takes after a deployment failure requiring immediate intervention.
  • Change fail rate: the share of deployments requiring immediate intervention.
  • Deployment rework rate: the share of deployments that are unplanned responses to production incidents.

DORA advises interpreting these at application or service level and avoiding universal targets or comparisons between unlike applications. Use them to improve the delivery process together, not to rank individual engineers. The two rates answer different questions; a repair deployment can be rework without itself failing.

Agree the counting rules before discussing the trend

The scorecard needs a small measurement dictionary. Name the service, time zone, reporting-window boundaries and data source. Define what counts as a production deployment, how repeated deployment attempts are handled and which event marks recovery. Keep those rules beside the figures.

For lead time, specify how changes are linked to their production deployment. Choose and label your summary, such as the median, while retaining the underlying observations. An average and a median are not interchangeable. Do not quietly exclude awkward releases to tidy the chart.

Record coverage as well. “Nine of twelve deployments linked to commits” is more useful than a precise-looking lead-time figure that hides missing data. If the logging changed this month, annotate the break rather than presenting a clean trend.

The worksheet provides fields for numerator, denominator, unit, source, coverage and context. It contains manually calculated examples and blank rows, not an automated monitoring system. Replace the examples with your own agreed measures. Do not upload employee-level activity or customer-identifiable records to fill it in.

Read a small example without overclaiming

Consider a fictional service over four complete weeks. It has 20 production deployments. Two require immediate intervention; three deployments are unplanned responses to incidents. One of those repair deployments also fails, so the groups overlap.

The change fail rate is 2 divided by 20, or 10%. The deployment rework rate is 3 divided by 20, or 15%. Do not add them to claim a 25% failure rate. Average deployment frequency for this window is 20 divided by 4, or 5 deployments per week; that average hides whether releases were evenly spaced.

Suppose the two failed deployments recover in 30 and 90 minutes. Their median is 60 minutes. With only two observations, keep both durations visible. This is a description of the sample, not a service guarantee or a prediction about the next incident.

For a separate five-change lead-time sample of 2, 4, 6, 10 and 18 hours, the median is 6 hours. The example worksheet labels this as a five-change sample, not the complete deployment population.

If a reporting window contains no failed deployments, record recovery time as not applicable with zero observations. Zero minutes would imply instantaneous recovery, which was not tested.

Put a customer measure beside the release measures

A healthy release process can still ship the wrong work. Add one outcome connected to the current product priority: completion of onboarding, successful bookings or support contacts caused by a particular fault. State whose behaviour is included and what counts as success.

Illustrative comparison: onboarding completion moves from 72 of 100 eligible starts to 84 of 120. That is 72% then 70%, a fall of two percentage points, even though the number of completions rose. More customers finishing does not necessarily mean the journey became easier.

Check whether the populations are comparable. A new campaign, different eligibility rule or changed instrumentation could explain the movement. Release timing alone does not prove the release caused it. Add research, support evidence or a properly designed comparison before making a causal claim.

Keep a guardrail beside the outcome where it matters. Faster onboarding is less useful if support must repair every account afterwards. The guardrail should reflect the actual risk, not become another dozen charts nobody reads.

Use the review to choose one intervention

Bring product, engineering and operations into the same conversation. Ask the team to walk through one recent change, including time waiting for decisions or approvals. A founder may own the bottleneck even when the delay appears on an engineering dashboard.

Suppose lead time increases while the number of releases stays similar. Inspect the waiting points before buying more capacity. If reviews queue behind one specialist, a sensible experiment might be a shared review window or a second qualified reviewer. Record the hypothesis, owner, review date and a safety check so speed does not come at the expense of a necessary control.

When failure or rework rises, inspect the incidents and the size and nature of the changes. Do not simply tell the team to deploy less. A release freeze may suppress the count while leaving the underlying problem untouched.

Carry one agreed improvement into the next review. Record what changed, what remained uncertain and whether the intervention should continue. Stop collecting a measure if nobody can explain a decision it informs.

Keep the numbers out of individual league tables

Do not use commits, lines of code, ticket counts or these service-level metrics to score individual productivity. People doing support, reviewing difficult changes or removing obsolete code would be penalised by a crude activity total. It also gives everyone a reason to optimise the count rather than the product.

Discuss targets, where needed, in the context of a customer obligation or an explicit improvement hypothesis. “Every team must deploy daily” is not a substitute for understanding the service. A small team can start with a manual review of recent releases and incidents instead of spending its first month building a metrics platform.

Ask which constraint the evidence reveals, then give the team room to address it.

Download the editable worksheet (CSV) ↓

Practical questions

How often should a founder review the scorecard?

Choose a rhythm that gives the team enough observations and time to act. Review an urgent incident when it happens. For routine trends, retain the same window definition and label sparse samples rather than forcing a weekly conclusion.

What if there are too few deployments for a meaningful trend?

Show the individual observations and discuss the changes themselves. A rate based on one or two releases is easy to calculate but fragile to interpret. Use a longer clearly labelled window if that helps, without hiding recent incidents.

Does faster delivery prove engineering is creating more value?

No. Pair delivery evidence with the relevant customer outcome, and check for changes in the customer population or measurement. A faster release process is useful capability, not proof that the product decisions were right.

Do we need to buy a metrics platform first?

No. A manual review of recent deployments and incidents can establish the counting rules and expose a constraint. Consider tooling when the benefit of more reliable collection justifies its cost and operating effort.

Put this to work

Sources checked 8 November 2026. The examples and worksheet are suggested working aids, not client results, legal advice or guaranteed outcomes.