What production means for a language feature

Most language features fail in the gap between a good answer and a reliable one. The demo is allowed to be fluent. The feature is not. Production, for us, means a refusal path, a source the user can open, a latency budget, a log of what left the building, and a person who owns the metric when it slips.

We write those five down before we argue about models. A smaller model on your metal with a citation often beats a larger API that improvises. Sometimes it is the other way around. The point is to choose with the constraint in the room: privacy, cost, language, and the cost of a confident wrong answer.

Evaluation is a set of questions your operators already ask, scored by someone who knows the right answer, run again every time the prompt or the index changes. If you cannot rerun it, you do not have a feature. You have a performance.

Handover is part of the definition. If the only person who can explain a wrong answer is the consultant, it is not in production yet. The runbook, the rollback, and the owner’s name belong in the same pull request as the model.