What the headline multiples compare

That is the gap in the description. The 30% appears only as a chapter title, the real-world number, pointing at the dspy-typesafeify repo, and the entry notes the underlying table is not published. Do not cite it without opening the video and the repo.

The video opens on the question of what the headline speed numbers compare, and that framing is the argument, but the description does not spell out the baseline. Without the chapter content you cannot tell whether the comparison is fair.

He places it as a layer where a decision fits rather than a wholesale replacement, walking through Choice, Score and Noul. The description does not quantify that placement or say which calls he thinks should move, so it is a position, not a measurement.

A repo is named, dspy-typesafeify, which is the most concrete thing in the description. It is not linked directly and the description does not say what dataset or metric the 30% comes from, so the repo is where you would have to start the work yourself.

Nimrod published it, walking through Choice, Score, and Noul alongside the question of what the headline speed numbers compare.

dspy-typesafeify. The entry notes the repo is named in a chapter title but not linked directly and that the underlying table is not published.

It separates a schema guarantee from a correct decision, which is the same point other commenters make about typed answers still being wrong.