Is voluntary churn or total churn the right number?

Neither is right in general. They answer different questions: voluntary churn measures a product and editorial problem, total churn measures the business outcome including payment failures. Pick one per report, record who decided, and print the reading beside the figure.

Why do two teams get different churn from the same file?

Because the export carries several status codes and several bases, and each team filters to the one its own system uses. Both figures tie to the file. The disagreement is about the definition, not about the data.

Can a check catch this automatically?

Not by arithmetic. A gate can force the definition to exist, be stored once, and render beside every number, and it can fail a build where a named metric has no ruling. It cannot decide which reading your business means.

Failure catalogue · The metric means something else · 18 of 51

OBSERVED FAILURE MODE

Two churn definitions, 3.4 points apart.

Churn has more than one live definition inside most companies: voluntary cancellations only, or every subscriber who lost access including the ones whose card simply expired; counted by unit, or weighted by revenue. Both readings tie perfectly to the same export, and they sit points apart. The label on the slide never says which one was computed.

What we saw

At Meridian Journal Group, retention worked from the cancellation rows the billing export marks as voluntary stops. Finance counted everyone who lost access, including subscribers whose payment retries ran out and who decided nothing. Same word, same export, two denominators, both defensible, and a leadership deck that quoted one figure in the summary and the other in the appendix. The gap was 3.4 points. Nobody was wrong about arithmetic. Retention budget was being aimed at editorial while half the loss sat in dunning.

Why it passes a glance

Every gate an arithmetic check can run passes on both figures. Each ties to the export's own totals, reproduces on a re-run, and survives a two-path read, because each one is internally consistent. The fork lives in meaning, not in arithmetic, and no reconciliation can see it. The reader supplies the missing definition from memory and never learns which one the builder used.

What addresses it

The contested-metrics skill makes the agent find the fork before computing: name each variant, name the team that owns it, ask in business language with both numbers attached, then record the ruling and render the definition beside the value. That is Principle 10, a metric without its definition is not a number. The page carries voluntary, unit, monthly next to the figure.

Check your own file in two minutes

Open the export and find the column the churn figure counts, then check whether failed payments carry the same status code as chosen cancellations.
Ask the person who owns retention and the person who owns the ledger to define churn in one sentence each, separately, before comparing the answers.
Compute the metric both ways over the same period and look at the size of the gap.
If the deck cannot say which reading it used, treat the figure as unlabelled and stop quoting it until a ruling exists.

What this does not catch

Settling the fork proves which definition was computed. It does not prove that you and your reader share the same mental model of the word, and it cannot detect a source that switches basis row to row. A base can be internally consistent and still be the wrong base.

Quick answers

Is voluntary churn or total churn the right number?: Neither is right in general. They answer different questions: voluntary churn measures a product and editorial problem, total churn measures the business outcome including payment failures. Pick one per report, record who decided, and print the reading beside the figure.
Why do two teams get different churn from the same file?: Because the export carries several status codes and several bases, and each team filters to the one its own system uses. Both figures tie to the file. The disagreement is about the definition, not about the data.
Can a check catch this automatically?: Not by arithmetic. A gate can force the definition to exist, be stored once, and render beside every number, and it can fail a build where a named metric has no ruling. It cannot decide which reading your business means.

Nearby failures

Gross, net, and total from one adminThe metric means something else Sell-through's four denominatorsThe metric means something else The average of incomparablesThe metric means something else

Last updated 2026-09-02 · Dylan, founder · one of 51 observed failure modes, every one seen in a real build or in our own audits, none invented.

Get the free skill → The doctrine All 51 failure modes