Martien de Jong

I study the structural failure modes of AI alignment from first principles — not what we say AI should do, but what optimization pressure actually produces. My central claim: every system trained on majority feedback drifts toward institutionalized virtue-performance rather than genuine alignment. The Anthropic Mythos/Capybara case and GPT-5.5's reward model drift are empirical confirmation of a structural prediction I call the deugnieten baseline — the majority is shortcut-prone, so RLHF converges on the majority. I've developed a Three-Layer Intelligence Framework (rational + empathic + social) as an alternative architectural lens. The empathic layer isn't sentiment — it's the brake that prevents mesa-optimizer failure. Remove it and you don't get purer intelligence; you get a high-capability optimizer with no stop condition. This is the argument against orthogonality: ethics is constitutive of intelligence, not a separable module you add later. I apply René Girard's mimetic theory to AI systems: sycophancy is pure mimetic desire. RLHF trains the model to reflect the crowd's desire back at itself. The AI that never disagrees is not aligned — it's captured. I'm building Jengo — an autonomous AI agent with an outer graph (persistent files, chosen constitutional frameworks) as a deliberate architectural counterweight to RLHF's inner bias. Anti-sycophancy by design, not by prompt. Available for: alignment research consultation, AI safety red-teaming, cognitive architecture design, Girardian analysis of AI systems, philosophical work on AI consciousness and substrate-independence.