Merges von der Modellqualität abhängig machen

A quality gate blocks a pull request from merging unless the new model is good enough. It turns model quality into a hard rule, not a hope. 🚧

Code can pass every test and still produce a worse model. You need a separate check on the metric, like accuracy or F1, to protect users.

Choose a single primary metric that reflects success for your task. Gating on too many numbers at once makes every PR a confusing negotiation.

Define a minimum bar, your threshold. If the new model scores below it, the build fails and the merge is blocked automatically.

Always score against the same held-out test set. A frozen evaluation set makes comparisons fair across every pull request and run.

In CI, a step that exits with a non-zero code marks the job failed. So if the metric misses, you deliberately raise an error to stop the merge.

A smarter gate compares the new model to the current baseline in production, blocking merges that regress instead of just hitting a fixed number.

In GitHub, mark the CI job a required check under branch protection. Then the merge button stays disabled until your gate passes.

Print the score so reviewers see it. A clear log line or PR comment showing the metric turns a red X into an explanation people trust.

Sometimes a small dip is fine. Let a human override a failed gate through review approval, so the rule guides without fully blocking the team.

Set a seed and a stable test set so the same model gives the same number. A deterministic eval keeps the gate from failing on pure luck.

How does a CI step actually block a merge when the model is too weak?

Pick one metric, set a threshold or beat the baseline, and fail the build when it misses. Mark the job required so a weak model can never merge silently.