How do you examine a model that changes every month?
An examination describes the model as it was that day. Medicine needs assurance about the one in use now.
Open question. Not yet tested.
On this page
Why it is hard.
An examination describes a model on the day it was examined. The model in use six months later may be a different one, updated by its maker a dozen times. Even an unchanged model ages, because hospitals change their equipment, their patients, and their habits, and the model was taught on the old ones.
The gap is not theoretical. A sepsis warning tool used in hundreds of hospitals was tested by independent researchers in 2021. It missed most of the patients who had sepsis and raised alarms on many who did not. It had been in daily use for years.
What has been tried.
Regulators have begun to allow a maker to agree a plan of permitted changes in advance, so that a model may be updated within limits without a new approval each time. The core protocol builds on that idea: a change outside the agreed corridor is rejected, and a model whose confidence drifts away from its accuracy is suspended within a bounded time. Neither has yet been tried in a hospital.
A direction we would test.
We would treat examination as a standing duty, with four parts. A sealed set of cases is held by an examiner the maker never meets, and every new version sits it again. A new version runs silently beside the old one on real work before it is allowed to speak. Each hospital keeps a small set of its own cases, because a model that works in Boston may not work in Ankara.
The fourth part is the result. It should be a report card that shows performance over time, and never a certificate that is awarded once and framed.
How we would know we were wrong.
Sealed questions leak, and a model can be tuned to pass an exam without getting better at medicine. If the examiner cannot keep its cases fresh and secret, or if maintaining them costs more than hospitals will pay, the standing examination becomes theatre.
Who could solve it.
People who run clinical AI inside a hospital and have watched a model get worse without anyone announcing it. Statisticians who design tests that cannot be gamed. Regulators who are writing the rules for models that change. And makers of models who would welcome an examiner they do not control.
Where this connects.
Solve one with us.
Tell us which problem you work on, and what we have got wrong about it.