Working papers and technical reports from the Advanced Behavioral Cybersecurity & AI Institutehttps://abcai.org/en-ushttps://abcai.org/research/002-adversarial-multi-perspective-review/https://abcai.org/research/002-adversarial-multi-perspective-review/When autonomous artificial intelligence systems are tasked with reviewing their own implementations, conventional single-agent evaluation succumbs to confirmation bias, authority expansion, and semantic laundering. In this paper, we demonstrate how an independent adversarial lane ('Specter') operates with inverted incentives—specifically designed to disqualify implementation claims through bypass discovery, exploit path mapping, and boundary pressure. Evaluated across production CI pipelines, this adversarial topology reduced false closure events to zero while maintaining high throughput.Sun, 20 Sep 2026 00:00:00 GMThttps://abcai.org/research/001-deterministic-governance-ratchets/https://abcai.org/research/001-deterministic-governance-ratchets/Autonomous software organisms operating over long horizons inevitably face context compaction, architectural entropy, and behavioral drift. Traditional runtime guardrails fail closed intermittently or permit gradual degradation under operational pressure. In this paper, we introduce a deterministic governance architecture built on one-way upward ratchets: every consequential execution emits attributable feedback signal that ratchets constraints, test surfaces, and formal contracts upward. By enforcing strictly decoupled, parallel review lanes—pairing an adversarial exploitability lane with a formal invariant and custody lane—we prove that self-modifying autonomous systems can safely evolve in perpetuity without compromising security, provenance, or truth.Tue, 15 Sep 2026 00:00:00 GMT