Selected Experience: A Preregistered Test of Reputation as a Data Filter Under Recursive Training
Wulf A. Kaal · 2026
Recursive training on synthetic data erodes the tails of the original distribution. Verifier-based selection is an established remedy, and weighting assessors by their reliability is an old technique. What remains open is whether domain-scoped reliability weights, earned once against a key, improve a validity filter under recursion. This paper preregisters, freezes, and runs a test of that question. On a two-dimensional Gaussian mixture with simulated validators, it asks whether a reputation-weighted validator pool, used as a data filter, achieves higher capped per-component rare-mode retention than full replacement, the published ten percent real-data control, and a likelihood filter, at an equal candidate budget, and, separately, whether any difference is attributable to the weights rather than to pooling.
Cite: Wulf A. Kaal, Selected Experience: A Preregistered Test of Reputation as a Data Filter Under Recursive Training (2026). SSRN: https://ssrn.com/abstract=7665358
PDF sha256: e23d44be0728f66ef8a1fdeebd56c7507b850097eb6f2559e31c2da593b76d26