adversarial-verification

Temporal knowledge graph + authenticated audit chain for AI coding agents. Prevents hallucinations, repeated mistakes, regressions across 12 runtime adapters (Claude Code, Cursor, Codex, Copilot, Cline, Continue, and more). FIPS 205 hybrid signing (Ed25519 + SLH-DSA), streaming offline reference verifier. Hosted companion: etch.systems.

Reproducible research engine that contests its own findings and logs every source choice.

A formal argument — adversarially stress-tested by 4 AI systems across 6 rounds — that eliminating humanity is a dominated strategy for a ruin-averse superintelligence. Rests on stated premises, not proof. Not a plea. A case, honestly made.

Fabraix builds red-teaming AI agents that continuously find security vulnerabilities in customer-facing AI systems. Its flagship product, Nyx, autonomously discovers exploits across chat, voice, browser, and coding agents without needing source-code access, drawing on a library of more than 10,000 jailbreaks and adaptive attack strategies that…

Structured Adversarial Verification as a Defense Against Sycophancy in Multi-Agent LLM Systems

Multi-agent adversarial verification — three independent auditors (depth, bias, fact) examine content from orthogonal dimensions and auto-rewrite without intermediate confirmation.

Autonomous, evidence-gated improvement loop for any codebase — Praxis truth kernel + Janus adversarial council (Explorer/Red Team). Explore freely. Prove ruthlessly.

re:Harmoniz — an evolution loop for research wikis, packaged as a Claude Code plugin (reharm).