Direct answer
The short version.
Kill criteria are pre-agreed conditions that pause or end an experiment when spend, quality, safety or evidence makes continuation unjustified. They sit beside scale and learning criteria so the team knows all possible decisions in advance.
Key takeaways
Keep these three decisions.
- Define stop, change and scale outcomes together.
- Separate commercial loss from safety triggers.
- Record exceptions and why they were made.
The operating problem
Without criteria, weak campaigns continue because the team has invested in production or stop too early because an early average feels uncomfortable. The same data can support opposite decisions depending on who is in the meeting.
A practical system
Define minimum runtime or sample, maximum acceptable loss, non-negotiable quality and safety signals, and the evidence needed for iteration rather than termination. Separate reversible underperformance from policy, trust or technical failures that require immediate action. Give one owner authority to execute the rule.
What to measure next
Review avoided loss, false stops, time to decision and how often criteria were changed after seeing results. Good criteria do not eliminate judgement; they make the assumptions and risk tolerance available before pressure distorts them.
Primary and authoritative references
Sources used for context.
Frequently asked questions
Two useful follow-ups.
Should every campaign have kill criteria?
Every material experiment should have stopping and review conditions proportionate to spend, risk and conversion lag.
Can kill criteria be changed during a test?
Only for a documented reason such as tracking failure or a changed business constraint, not because the observed result is inconvenient.