Write Your Kill Criteria Before You Fall in Love With the Idea

A kill criterion states before an experiment which result will make the team pause, change or abandon an assumption. It protects against sunk cost, metrics selected after the result and attachment to a solution already built.

Stopping an assumption does not mean abandoning the problem. It means refusing to continue the same answer without enough evidence.

Five elements

Hypothesis

What you believe. Example: “sales managers will approve a draft in under 90 seconds.”

Observation

What you measure: review time, correction and use of the next action.

Threshold

The boundary: “at least four out of five cases under 90 seconds with no critical error.”

Date or volume

When to decide: after five cases or on October 22.

Consequence

What happens below the threshold: remove automatic assignment, redesign the workflow or stop.

Four criterion families

• Value: the result is unused, the problem has low priority or no commitment appears.
• Feasibility: data is unavailable, exceptions dominate or review cost erases value.
• Trust: consent cannot be clarified, severe errors remain undetectable or recourse is absent.
• Distribution: the segment is inaccessible, the signal predicts nothing or the message damages reputation.

One experiment may have several stop conditions, including an immediate safety stop and an economic stop after the batch.

Avoid false criteria

“If it does not work” is not measurable. “If feedback is bad” depends on interpretation. “If nobody buys” defines neither offer, segment, volume nor time.

A good criterion is observable, bounded and connected to an action.

Continue, modify or stop

Write three branches before seeing data:

• Continue: the mechanism has enough evidence for the next step.
• Modify: the problem remains supported but a secondary assumption fails.
• Stop: value, feasibility, trust or distribution no longer justifies the cost.

Sales example

Hypothesis: an organizational-change signal improves conversation quality.

Test: manually contact 30 people in the segment.

Continue with at least five useful replies and three conversations where the problem is active. Modify if replies are useful but reveal a different problem. Stop with fewer than two useful replies or a deliverability warning after message and address quality are checked.

When to override

Major new evidence may justify revision, but record it before extending the test. Do not move the threshold merely because the result disappoints.

Legal, security or trust risk may stop the test immediately even when commercial metrics look good.

FAQ

Do kill criteria make a team pessimistic?

They make the investment explicit. A team can be ambitious while defining what it needs to learn.

Can a stopped idea return?

Yes, when new evidence, technology, distribution or constraints change the hypothesis. Preserve the original stop reason.

Who decides to stop?

The owner named before the test, together with people holding authority over the relevant risks.

Next step

Take your three active experiments and write a threshold, date and consequence for each. Use the Kill Criteria Generator. For a review, email contact@yumea.fr with “STOP” in the subject.