The Production Cut Should Be a Classifier

The middle of the software factory is already automatable. Spin up the environment, write the change, run the tests, run an adversarial review, move the card. What is not automated, and should not be waved through on hope, is the production cut. That decision is still a human queue in most companies, which means it becomes the bottleneck the moment the rest of the line speeds up.

The right shape for that gate is a classifier, not another reviewer. Most changes are boring in a way that is easy to name. No schema change. No public API change. No security-sensitive change. If a card clears those checks, and the automated review already beat it up, it should ship. If it trips any of them, a person looks. The point is not to remove judgment. The point is to spend judgment where the blast radius actually lives.

This only works if the earlier stations are real software, not a prompt hoping for the best. No Human Should See the PR Until It Has Been Beat Up is the precondition. A classifier sitting on top of an unreviewed diff is just a faster way to ship mistakes. The factory has stations: draft, test, review, security, ship. Agents staff the stations. The classifier is the policy at the last one.

Two gaps remain once you see the line this way. The funnel in, meaning what is worth building, held at the level of a business outcome rather than a pile of tickets. And the production cut, meaning which finished cards are allowed to leave without a person. Everything between those two should be background execution with a reviewable trail. That is the pattern that holds in an enterprise, because a leader can inspect the policy and the output without sitting in the path of every card.

Build the classifier with criteria your security and platform owners will sign. Then measure how much of the line still waits on a human who is not adding judgment. That wait is the next thing to remove.