A fraud detection model dropped from 94% to 75% recall in a single week, triggering no alerts because monthly metrics stayed within tolerance. When researchers fitted the classic Ebbinghaus forgetting curve to 555,000 production fraud transactions, they got R² = -0.31—worse than predicting the mean. This mathematical failure exposes a fundamental flaw in how the entire MLOps industry approaches model retraining.
Every major MLOps platform builds retraining schedules around smooth, predictable decay borrowed from 19th-century memory research. The assumption: models gradually forget like humans do, following an exponential curve where performance degrades continuously at a rate proportional to remaining accuracy. But production systems don't behave like psychology experiments. They face sudden distribution shifts, adversarial attacks, and market changes that create abrupt performance shocks rather than gentle slides.
The broader MLOps narrative focuses heavily on monitoring and lifecycle management, with 67% of AI models never reaching production and 91% experiencing performance degradation over time. But these statistics mask the real problem: we're treating symptoms of episodic failure with solutions designed for smooth decay. When R² falls below 0.4, scheduled retraining becomes actively counterproductive—you're optimizing for the wrong failure mode entirely.
For teams running production models, this research suggests a practical diagnostic: check if your weekly performance metrics fit an exponential decay. If R² < 0.4, abandon calendar-based retraining and implement shock detection instead. The math is telling you that your model isn't slowly forgetting—it's getting blindsided by changes your schedules can't predict.