In Constitutional AI, a plain‑text “constitution” encodes values and is incorporated into a model during post‑training. Creating such a constitution raises major challenges: who contributes, who decides which norms to include, and how participatory or democratic the process must be. Without addressing these issues, progress can stall and sovereign models risk misalignment or default to the norms set by frontier model providers.
We argue that decisive action is required. Treating constitution drafting as an evolving, citizen‑science‑like process, we propose five guiding principles: iterative, modular, justifiable, positional, and legitimate. The document should stay responsive over time and across the people it governs, with provisions marked as settled, contested, or underdetermined so that durable disagreements are recorded rather than erased.
To assess robustness, we built an agent‑based simulation. The results show that when 75% of participants act in good faith, the process tolerates up to 50% bad actors among citizens while still maintaining a stable living constitution. This demonstrates substantial resilience to adversarial participation.
Review: The paper presents a practical governance framework for sovereign models that balances technical implementation with democratic participation. Its empirical validation of robustness under high adversarial presence offers a promising direction for future model alignment efforts.