The Jury That May Not Vote for Its Own Family

Abstract

The Jury That May Not Vote for Its Own Family. A model reviewing work produced by its own family sits in a structural conflict of interest — it may be inclined to find that work plausible — so the verification juries described here are cross-family by construction: the producer's model family is disqualified from the bench on the conflict alone, verdicts are cast blind, and the verdict channel is designed so that injection attempts against it are detectable by construction rather than caught by vigilance. The family fingerprint is treated as a measurable risk to be controlled, not a settled result to advertise; measuring how large that risk is, and in which direction any effect runs, is handed to a planned follow-up rather than asserted here. This preprint is part of a series; it cites the companion campaign paper (Melikoğlu, Altınbaş & Tanrıöver, 2026; concept DOI 10.5281/zenodo.22086013). Bilingual: the full text is included in both English and Turkish. AI-transparency note: the draft was written by an AI agent under human direction (the human authors held go/no-go authority), blind cross-family jury-reviewed (the producing model family excluded from the bench), and reviewed against the source ledgers by the human authors before publication.