2026-07-14 (review): REJECTED a "super-reviewer" verb gated on a second model

2026-07-14 (review): REJECTED a "super-reviewer" verb gated on a second model backend. Its motivating rationale does not survive: self-preference under genuine authorship is weak or absent (arXiv 2606.20093, gap -5.1pp, CI crosses zero), and the bias that DOES exist on code is a context-framing artefact, not model identity (arXiv 2603.04582). CHOSEN instead: a fresh-context, off-policy, same-model reviewer — re-present the diff in a new session as an artefact of unknown authorship. Captures the only measured debiasing effect, costs nothing, needs no second subscription, and removes the two-tier UX problem entirely. A second model, when present, is used for disagreement-as-triage per adr-25 asymmetric trust and itd-81 — NEVER a panel or a vote (nine judges = 2.18 effective votes; on code, consensus underperforms the naive baseline — the "popularity trap", arXiv 2510.21513).