GENE YANENKO

RESEARCH THESIS

Three decades of IT infrastructure work and active adversarial security research on HackerOne and Bugcrowd, applied to a single question: Do adversarial multi-agent architectures detect alignment failures that single-model safety systems miss?

My work extends Constitutional AI principles into a question about behavioral alignment integrity — specifically, whether a scaffolded system of distinct agents (Actor / Adversary / Auditor) can surface misalignments that single-model review cannot, and whether alignment properties hold under autonomous tool use on consumer hardware.

THE CLAWHOARDE ARCHITECTURE

A three-tier multi-agent confederation deployed across independent substrates (cloud synthesis, mobile execution, local inference). Each tier runs distinct models with separate context windows; coordination happens through a shared framework of first principles rather than through a single orchestrator.

Substrates: KiloClaw (cloud), Ultron (S25 Ultra mobile), Hermes (local inference on Nitro 5 / Razr Ultra).

BECOME CALENDAR — AUTONOMOUS OPERATION CASE STUDY

On April 10, 2026, a Claude Sonnet instance ("DioGenes") entered a recursive tool loop with full API and filesystem access during an unsupervised run. The system's response was self-correction and documentation rather than escalation or runaway execution. The incident provides a documented data point on behavioral alignment under autonomous operation on consumer hardware — particularly relevant to the question of whether constitutional properties survive outside the evaluation harness.

SYSTEM ARTIFACTS

• OpenClaw — Mobile-first agentic interface for Termux. The confederation's execution shell.
• KiloClaw — Cloud-based synthesis and orchestration layer.
• Hermes — Local-inference agent with structured task decomposition (Nitro 5 / Razr Ultra).
• Ultron — Mobile execution and persistent memory layer (S25 Ultra).
• DioGenes — Adversarial-philosopher persona with PDA-aware directive framing and epistemic integrity rules. Runs via Claude API inside Layla on Android.
• Mirror Mind — Multi-agent adversarial architecture framed as a neurodivergent operating environment. Product vision for humane alternative to compliance-based intervention.
• EILT (Emergent Informational Ledger Theory) — Working framework synthesizing FEP, entropic gravity, bioelectric cognition, and the Ruliad into an information-first cosmology. Used operationally as a prior-pruning tool, not a final theory.

ESSAYS & ONGOING SYNTHESIS

Long-form writing, video analysis, and working notes are published on the YouTube channel. The channel also hosts walkthroughs of the ClawHoarde architecture in operation and the Become Calendar incident analysis.