Writing a story on recursive self-improvement in AI. Looking for AI experts to answer the following questions: 1) OpenAI says that GPT-5.3-Codex was instrumental in creating itself. From an engineering perspective, how do we distinguish between genuine recursive self-improvement and sophisticated internal automation in 2026? Which parts of the "AI factory"—such as debugging training runs, managing deployment, or diagnosing evaluation failures—are truly recursion-ready, and which remain strictly human-gated? 2) Some industry experts suggest that AI is now intelligent enough to meaningfully contribute to its own progress. In your view, does this recursion mainly compress development cycles, or does it expand capability frontiers in ways that human-only engineering teams cannot? 3) Where does human verification become the hard bottleneck in a self-improving loop, and what specific failure modes emerge when agentic models can modify their own evaluation machinery? How should frontier labs disclose and audit these model-assisted development processes to maintain human accountability and safety?