Evidence & feedback
What inspection records miss when a model picks the cases
A model that sends inspectors back to places with more recorded defects can collect more evidence from the same places. Reserve some inspections for a sample chosen independently of the model.
Suppose a library checks returned books for missing labels. Staff find eight defects in 80 checks from one branch and two defects in 20 checks from another. Both sets yield one defect per ten inspections. A tool that sees only the counts might send more checks to the first branch, where it has more chances to record defects again.
When a finding changes where you look next
In a 2018 study, Danielle Ensign and colleagues modelled what happens when predictive policing sends officers to places based on previously observed incidents. In their simplest mathematical setup with unequal underlying rates, the system updated its priorities using only incidents officers discovered where they were sent. The researchers gave an example with underlying rates of 10% and 11%. Under their assumptions, the probability of visiting the second area approaches 100% over time.
The result describes a hypothetical allocation rule. The researchers also modelled reports from residents, which can arrive from places officers did not visit. Those reports could lessen the feedback loop in the model. In the library example, a reader’s report about a missing label would likewise be useful to act on, but it would tell you little about the condition of all returned books.
Look beyond the model’s priorities
In a low-risk workflow, keep a list of eligible cases and distinguish those checked, those awaiting a check and those with no check planned. Record an unchecked return as unknown. Set aside some inspections for a random sample chosen without relying on the model’s ranking. Draw from each group you want to compare and apply the same checklist. Keep these results separate from inspections the model requested.
When you review the schedule, put findings beside the number of checks and the way each set was selected. If the independent sample finds a recurring issue among returns the model rarely selects, check a larger sample before you decide whether to reallocate the inspections.
Sources and further reading
Sources checked 6 November 2026. Prepared with AI assistance and editorial review.