Protocol
Use small safe-to-learn probes in complex situations
When prediction is weak, buy information before buying the whole solution.
When it fits
- You cannot reliably predict which intervention will work because outcomes emerge from many interacting factors.
When to avoid it
- Do not call a high-stakes uncontrolled rollout a 'safe-to-fail experiment.' Safety and ethics constrain the probe first.
Why it matters
Design several small probes that differ meaningfully, cap their downside, and define what you will observe. Run them in parallel or sequence where appropriate, amplify useful patterns and stop or adapt the probes that fail. The goal is not to prove one grand theory in advance; it is to learn from bounded contact with the system.
Steps
- A failed probe is survivable and still produces information that changes the next move.
An example
Test two small onboarding changes with a subset of users before rebuilding the whole product around an untested explanation.
Check your result
A failed probe is survivable and still produces information that changes the next move.
Keep this limit in mind
- Do not call a high-stakes uncontrolled rollout a 'safe-to-fail experiment.' Safety and ethics constrain the probe first.
Evidence and sources
In the Cynefin framework's complex domain, Snowden and Boone recommend probes that can safely fail so patterns can emerge before a larger response.
A probe still needs bounded downside and a useful observation plan; 'experiment' is not permission for uncontrolled exposure.
A Leader's Framework for Decision Making · Abstract; complex context