What is Tessary's cost_drift classifier?

It detects a call site spending more per turn than its own history says it should, with token counts attached as evidence. The comparison is always against that call site’s own past, never a fixed dollar figure.

The increases it exists for don’t throw errors: an agent paying for more context than it used to, extra retries nobody counted, a cache that stopped being hit. Nothing in the stack notices until the bill does.

Sudden jumps and slow creep both register, and a drop in cost shows up as clearly as a rise, which is worth as much: spend falling off a cliff usually means a call stopped happening. A detection writes a finding, and triage decides whether it’s a real deviation before it becomes anyone’s problem.