For the last few months, we've been working on a fundamental architectural shift in how autonomous agents are governed. The current industry standard relies almost entirely on probabilistic alignment (RLHF, system prompts, constitutional training). It works until it's jailbroken or the context window overflows. A statistical disposition is not a security boundary.We've built an alternative:...
- *Phase 1 (this repo):* deterministic classification + policy evaluation + signed constitution artifacts - signing/verification scaffolding (Ed25519) and templates - evaluator returns *DENY / CONFIRM / ALLOW* deterministically
- *Phase 2 (not in this repo yet):* *Gateway/tool-router enforcement* so the agent can't bypass the decision Clarification (so nobody over-reads it): this is a *reference integration*; it *doesn't grant patent rights* and doesn't disclose enforcement mechanisms beyond illustrative examples.
Pointers: - Repo: https://github.com/genesalvatore/aos-openclaw-constitutional - Policy evaluator: https://github.com/genesalvatore/aos-openclaw-constitutional... - Canonicalize/sign/verify: https://github.com/genesalvatore/aos-openclaw-constitutional... and https://github.com/genesalvatore/aos-openclaw-constitutional... - Constitution spec: https://github.com/genesalvatore/aos-openclaw-constitutional...
Happy to answer "how would you wire this into X tool router?" questions.
https://github.com/genesalvatore/aos-openclaw-constitutional
Yesterday I integrated the AOS Constitution (git-stamped Jan 1, 2026; https://aos-constitution.com) into a local-first agent (OpenClaw). The session later restarted with effectively zero retained context, so I asked the agent to review our humanitarian license with fresh eyes. It caught an overclaim ("uses automated monitoring" before we'd actually built monitoring). We fixed it in v1.0.1 the...