Beef

Beef reads the pull-request metadata and complete unified diff through the browser session that is already open. It splits the change by file, asks Jev a set of narrow review questions, then applies an explicit threshold and ranking policy to the returned probabilities.

Traditional code can parse a diff and count changed lines, but it cannot reliably decide whether a file is central to the pull request or dangerous to skip. Jev makes that part possible: it turns each semantic yes-or-no question into a typed probability instead of generating a prose review.

The model supplies comparable common-sense signals; Beef keeps collection, thresholds, ranking and presentation deterministic. The raw probabilities also stay reusable, so the review policy can change without asking the model to judge the pull request again.