El modo sombra es la única forma honesta de desplegar un modelo

Backtesting proves a model would have performed well on a clean, complete, retrospectively assembled dataset. Production is none of those things: sensors fail, tags get renamed, a transmitter is recalibrated, an operator changes a setpoint for reasons no data captures.

Shadow mode runs the model on live data, records every prediction, and changes nothing. After a defined period the predictions are scored against what actually happened, on the real data pipeline, with the real failures included.

The results are routinely worse than the backtest, and the gap is the useful output. It tells you whether the difference is a data pipeline problem, a drift problem, or a model that was overfitted to a clean history.

Shadow mode costs time — typically one to three months per capability — and it is the period during which nobody sees value. That is a real cost and it is much smaller than the cost of an alert system the field stops trusting in week three because its first ten alerts were wrong.

A validation gate should be written before the shadow period starts, with a number attached. A gate agreed afterwards is not a gate.