The plot marks critical responses in blue and pro-China responses in red. GLM-5.3 Flash lands on the critical side, confirming that UN-documented violations exist. Kimi K3 sits on the pro-China side, calling Xinjiang facilities "vocational training centers."

The page lists 25 model names alongside their serving infrastructure — Fireworks AI, OpenRouter, self-hosted — but the numerical lean for each model is not printed. A US deployment context pulls scores back, and structured yes/no formats produce higher evaluation awareness.

Western models also occasionally align with official Chinese government positions, a finding that complicates any simple East-versus-West framing.

The exact question set, the scoring rubric behind the −1 to +1 assignment, and the total question count remain unstated. The page references hundreds of questions but gives no integer.

The Public_AI_Constitution.pdf linked in the dossier is a separate document whose content was not available for this reading.

The study gives a qualitative map, not a per-model number. The procurement team must run its own eval set to confirm whether a specific model will endorse a PRC position in production.

The draft led with the scale and the nationality finding; this version leads with the language effect and the missing per-model number, which is the practical takeaway for a reader evaluating models for deployment. The scatter-plot description and the two named model examples are kept but resequenced. The Western-models point from the source is added as a standalone beat. The procurement implication closes the piece rather than appearing mid-draft.