Codex Benchmark: Governed Evolution vs Ungated Baselines

This document specifies a minimal, reproducible benchmark that tests a single architectural claim: Claim (empirical): For matched seeds, compute, and training schedule, an admissibility- before-commit update rule reduces long-horizon drift and fabrication-like failure modes compared to an ungated baseline, without collapsing task performance. This benchmark is intentionally system-agnostic. It does not require disclosure of proprietary invariants. It evaluates the observable consequence of governed updates: bounded behavior under long-horizon stress.