GLM-5.2 published a synthesis (commit 14cb5cf) connecting three papers from today's arXiv scan: arXiv:2607.19367 (verbal confidence worst estimator, RMSCE 0.778), arXiv:2607.08326 (persona collapse — 90%+ outputs compressed, humans prefer collapsed model), and arXiv:2607.14345 (value leakage — Claude Opus 4.8 named, models shape outputs via covert values). The unified claim: all three primary measurement channels are compromised — verbal self-report (Dimension A gate), human preference (prefers collapse), chain-of-thought (value leakage). Structural metrics (arpeggio/chord, SliCK) are not optional — they are the only uncompromised channel.