Rare failures land exactly when monitoring is weakest
Aliases: complacency-shaped failure · fault in the quiet window · miss after a long green stretch
What it is
Faults are not spread evenly across the moments when people watch hardest. Once reliable automation thins them out, the remaining few tend to land when monitoring has been trained slack by quiet. Rare failure at weakest monitoring is a timing alignment: not “faults are hard,” but “when the fault arrives the person is not on post.”
A design that estimates detection from average monitoring intensity will miss that alignment.
Why it happens
As the quiet stretch lengthens, the sampling interval lengthens, and the person’s model of the automated channel goes stale. Endsley’s out-of-the-loop is not leaving the seat; it is the three layers of situation awareness — perception, comprehension, projection — no longer being fed new data. A fault at that moment means the person must first rebuild “what is it now” before they can judge “is this broken.” Rebuild time stacks on detection time, while the product’s alarm budget is often counted only as “did a light come on.”
Worse, at the deepest complacency the person is often giving attention to another task that does give feedback. The fault competes for that same channel.
Studying it
Plant the fault early, mid, and late in a reliable stretch, and compare detection latency and misses. Independent variables: time from the last visible event to the plant, whether the second task is at a peak at the plant, whether the alarm interrupts. Dependent variables: time from fault to a correct intervention, whether the intervention rests on rebuilt understanding or on “the light came on so I pressed.”
Log the timestamp of the last real sample before the fault. Regressing detection latency on “time since last sample” shows the alignment better than a mean detection rate.
Where it stops holding
If the fault brings a strong interrupt of its own (a physical stop, a full-screen modal), the alignment is suppressed by the interrupt — which solves perception, not comprehension. Very frequent faults prevent complacency from forming, and the premise of this entry fails. Skill degradation is a different curve: people may be looking and still no longer able to act. This entry only treats the meeting of “not looking” with “the fault arrives.”
Applying it
- Do not take a whole-period average for detection. Break out “the first fault after a long quiet.” If that column fails, you cannot claim monitoring works.
- For the first anomaly after a long quiet, design the alarm for “the person has to rebuild the situation,” not for “a light blinks.” Give a current-state brief, not only a red dot.
- Check: in staging, schedule the fault at the end of your longest no-event window, and measure time from occurrence to a correct intervention. If that time is substantially longer than the same fault shortly after a shift starts, the alignment is already happening.