A tuned model in production is a decision with a cost. It changes behaviour invisibly, it goes stale, and it can fail in ways a base model would not. Shipping it means the change is visible, measurable, and reversible, and those three are the release.
About 1 min of reading
Before this new lesson, one small question about the lesson you just finished. Recall it, do not look it up.
Make the switch real.
Record the model
Before you run it, say out loud what it will print. Then open the answer.
The switch
Before you run it, say out loud what it will print. Then open the answer.
Read the first example one line at a time. A line that prints shows what appears on screen; the rest set things up quietly.
Run the tuned model on real traffic without serving it, and compare. That period is where the difference shows up on real inputs, and it is cheaper than finding it after the switch.
The harness runs as part of the release, and a failing score blocks the deploy. A release that can ship a bad model is a release with no referee, and the referee is the only part that makes it safe.
Record which model produced an output. When a complaint arrives, the answer to 'which model said this' has to be instant, and without the record the investigation is a guess across a week of traffic.
A tuned model carries the date of its data. If the job depends on anything that moves, that date is the bug, and a model with no date is a model whose bugs cannot be found.
The base model stays available and the switch is one config value. A tuned model that turns out worse than prompted has to be turned off without a rebuild, and that is the whole reason it is shipped as a switch.
Each one is defined here on this site. Open the entry for the full definition and the other words it points at.