kaal:claim:7665358-020

Claim. The difference between full replacement's perplexity rise and the ten percent arm's came out at -231.24, the retention arm having climbed 287.28 to reach 330.44.

Type. empirical · Support. reported

Holds when

• Reports measurements on the replication gate, substrate P, as five-run means.

Source quote

The record reports, for each arm, the five-run mean perplexity on the test split at generation 0 and at generation 9, and the rise between them. Full replacement: generation 0 42.22, generation 9 98.26, rise 56.04. Ten percent retention: generation 0 43.15, generation 9 330.44, rise 287.28. Full replacement's rise minus the ten percent arm's rise: -231.24.

From. Wulf A. Kaal, Selected Experience: A Preregistered Test of Reputation as a Data Filter Under Recursive Training (2026), page 9. 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. 1d988cb5667b3b1bcd1b48f7c7c47f117ce4c4b4382950eeeaee4f5a8c281d6d