A tuned model is an artifact with a version, a base, and a set of examples behind it. Serving it means those three are known, because a model served with the wrong template or the wrong base produces output that is confidently wrong rather than obviously broken.
About 1 min of reading
Run the three way eval.
Three scores
Before you run it, say out loud what it will print. Then open the answer.
The other case
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 same test set against the base, the prompted base, and the tuned model. Publishing the three numbers is what makes the choice defensible to anyone reading the deploy.
The previous model stays loaded or one step away. A tuned model that turns out bad has to be swapped in minutes, and a rebuild is not minutes.
Name the model with the run that produced it and the date. Two tuned models that differ by one epoch are indistinguishable in production unless one of them has a name.
The serving code carries the exact template used in training, in one place, tested. A template edited later is the bug where quality drops with no code change visible in the diff.
The held out test set runs against the served model on deploy. That is the check that catches a wrong base, a wrong template, and a bad run, and it is the only place all three are caught.
Each one is defined here on this site. Open the entry for the full definition and the other words it points at.