[P] Cold Validation: Open-source system where one AI agent audits another with zero shared context

Raxe AI has open-sourced Cold Validation, a novel architecture designed to bring independent verification to AI-generated code. The core principle is simple but powerful: the AI agent that builds something should never be the one to review it. The system enforces this through strict separation of duties, employing two distinct AI agents—a Builder using Anthropic's Claude Code to produce plans and code, and a Reviewer using OpenAI's Codex CLI that audits only the final artifacts, completely blind to the original reasoning or development process.

An orchestrator component manages the entire workflow, enforcing phase gates and convergence between the builder and reviewer. Crucially, the reviewer operates in complete filesystem isolation, running in temporary directories with no access to the original repository or development context. All findings are tracked using durable fingerprints across review rounds, and a controller independently reconciles verdicts against any blocking issues. Released under the permissive Apache 2.0 license, the system comes with 35 mechanical tests to validate its operation, making it immediately usable for teams wanting to implement rigorous AI code review processes.

The architecture addresses a critical gap in current AI development workflows where the same model or agent typically both generates and validates its own output, creating potential blind spots and confirmation bias. By implementing what amounts to a 'four-eyes principle' for AI agents, Cold Validation introduces an audit trail and independent verification layer that could significantly improve code quality and security in AI-assisted development. The system's design prevents the reviewer from having any contextual knowledge about how or why code was written, forcing it to evaluate artifacts purely on their technical merits.