kaal:claim:7665358-028

Claim. One confirming audit finds that three choices can reverse the comparison between full replacement and the ten percent arm: retained real data was added to a full synthetic set while the public code keeps the original size and replaces the other rows with continuations, later generations were scored at their final epoch while generation 0 was scored at its best validation epoch, and the design fixed the learning rate schedule where the published methods left it open.

Type. empirical · Support. reported

Holds when

• Reports what one confirming audit identified, not what the paper concludes.

Source quote

One confirming audit finds that three choices can reverse the comparison between full replacement and the ten percent arm: retained real data was added to a full synthetic set, while the public code keeps the original size and replaces the other rows with continuations, later generations were scored at their final epoch while generation 0 was scored at its best validation epoch, and the design fixed the learning rate schedule where the published methods left it open.

From. Wulf A. Kaal, Selected Experience: A Preregistered Test of Reputation as a Data Filter Under Recursive Training (2026), page 10. Full text.

Cite as. Wulf A. Kaal, Selected Experience: A Preregistered Test of Reputation as a Data Filter Under Recursive Training (2026). SSRN: https://ssrn.com/abstract=7665358

Verify. sha256 of source PDF e23d44be0728f66ef8a1fdeebd56c7507b850097eb6f2559e31c2da593b76d26 at https://wulfkaal.com/papers/7665358.pdf

Content hash. 29729a607907fb2addb19310d265f53eced217791878724c2b0287a6809b2d8e