Steady-State Hypothesis
The measurable statement of normal behaviour that a chaos experiment predicts will hold while a fault is injected, without which the exercise is not a test.
A chaos experiment is a test, and a test needs a prediction. The steady-state hypothesis is that prediction, expressed in the terms users experience: order submission success rate stays above 99.9%, checkout p99 stays under 400 ms, throughput stays within ten percent of baseline.
The sequence follows from it: state the hypothesis and measure the baseline; define the smallest fault that would test it; define the blast radius and the abort condition; inject; observe whether the hypothesis held. If it held, the resilience mechanism is confirmed. If not, you have a specific finding.
The framing matters because "let us break something and see" produces incidents rather than knowledge. Choosing a hypothesis forces the team to say in advance what they believe the system does under that fault, and the interesting outcome is frequently that nobody agrees on the prediction — which is itself the finding, and it arrives before anything is injected.
Two conditions before running one in production: it must be abortable within seconds, and the team must already be able to observe the hypothesis in real time. Without the second, you cannot tell whether it held.