Drift
Also called: configuration drift, model drift
The gap that opens between how a system is defined and how it actually is — in infrastructure, because someone changed a console; in AI, because the inputs changed.
Two different problems share the word, and both matter.
Configuration drift
Your infrastructure is defined in Terraform. At 2am during an incident, someone changes a security group in the console to get service back. That was the right call.
The problem is what happens next: if nobody notices, the running system and its definition have permanently diverged. The code is now a lie, recovery from it would produce something that does not work, and the divergence is undocumented.
The fix is not preventing console access — it is detecting drift and then either codifying the change or reverting it. Any estate without drift detection has drift; it just does not know where.
Quality drift in AI systems
Accuracy slides from 94% to 81% over a quarter and nobody notices, because nothing errored. Causes are mundane: the input mix shifted, a supplier changed their invoice template, a provider updated a model, or three prompt tweaks each helped one case and broke four others.
Detecting this needs an evaluation set running against real traffic on a schedule, with alerting on the number. Application monitoring cannot see it — the service is up, the responses are well-formed, and they are wrong.
The common thread
Both are silent. Neither throws an exception, and both are discovered late and expensively — configuration drift during a recovery you needed to work, quality drift when a customer complains.
The instrument is the same in both cases: declare the expected state, measure the actual state, and alert on the gap.
Related terms