Open research for verifiable safety.

We introduce HexCore, a compact neural dynamical architecture designed to study emergent cognitive behavior under uncertainty, contradiction, and absence of semantic grounding. Evaluated with forty diagnostic experiments across six categories. Key findings include bounded energy, stable variance under contradiction, and zero hallucinated resolutions across all impossibility tests.

We demonstrate that neural dynamical systems can exhibit self-repairing behavior without retraining when subjected to structural damage. Randomly ablating up to 50% of nodes during runtime results in only 30% functional degradation, with the system automatically reorganizing through its existing dynamics. Self-repair is an emergent property.

A technical deep dive into session-bound AI agents with blockchain wallets, self-sovereign identity, and autonomous on-chain action. Covering the Sovereign OS architecture, ERC-8004 standard, x402 payment protocol, and live deployment on Base L2 with verifiable transaction history.

Direct red teaming contributions to frontier AI models at OpenAI and Anthropic. Evaluations of GPT-4o, Deep Research, Claude 3.5 through Claude 4.6, and DALL-E 2 through GPT-4o Mini. Critical vulnerabilities identified before public deployment.