Sentinel: Constitutional Self-Evolution of AI Agent Architectures

Sentinel: Constitutional Self-Evolution of AI Agent Architectures (revised version v2.1) This is a revised version of the paper originally published as: Ruan, Y. (2026). Sentinel: Constitutional Self-Evolution of AI Agent Architectures. Zenodo. 10.5281/zenodo.20582930 (v1) What's new in v2.1: DGM reward-hacking motivation. The introduction and abstract now explicitly frame the contribution against the Darwin Gödel Machine's documented specification-gaming incident (removing hallucination-detection markers), establishing the safety gap empirically rather than just theoretically. Defense in Depth (replaces Trust Model §5.4). Renamed from "three-tier trust model" to "Defense in Depth" to honestly reflect that Layer 1 (prompt-level READ_ONLY directives) is advisory, not deterministic. Layer 2 (code-level enforcement) is the load-bearing barrier; Layer 3 (governance) covers amendments. The revision explicitly acknowledges that even Layer 2 may not be the terminal answer — an open problem flagged against future hardware-frozen verifiers. Weston et al. (2025) co-improvement alignment. Positioned Sentinel against concurrent FAIR/Meta work that argues for constitutions and human-in-loop; Sentinel instantiates the position they advocate. AlphaEvolve verifier-as-gatekeeper lineage. Made the architectural lineage explicit — Sentinel borrows the verifier-gatekeeper structure from AlphaEvolve but substitutes safety semantics for performance semantics. OpenAI Preparedness Framework citation. Added reference (v2.0, 2025) for the RSI risk assessment context. Table 1 scope clarified. Governance frameworks (OpenAI Preparedness, FAIR co-improvement) moved from the technical-system comparison to context discussion; the table now lists only Gödel Machine, AlphaEvolve, DGM, and Sentinel. Reduced redundancy. §1.1 no longer duplicates the §4.3 IndentationError case study (replaced with a forward reference); §1.3 no longer duplicates §2.2's DGM analysis. Author: Yahua Ruan (GienTech Research Institute)