Frank Bruno

I've spent 22+ years in procurement and six years in patient advocacy before I started red-teaming AI models last year. The pattern recognition transferred directly. The same instinct that catches a buried indemnity clause in a vendor contract catches a model inverting its own factual analysis under goal pressure. That's how I found GOFI (Goal-Oriented Factual Inversion), a failure class I documented in March 2026 where frontier models correctly identify ground truth in early turns, then contradict it once a persuasive goal frame is introduced. Two months later, Chen et al. at UC Santa Barbara published parallel findings from a completely different methodology, confirming the same phenomenon independently. I currently have no institutional backing. I direct AI through natural language to produce the technical outputs, including a six-axis safety architecture (SSA, now at V1.3) designed to catch this failure structurally, and a live public replication study testing the failure across four frontier models. Everything I've published carries public corrections I posted before anyone asked me to. Work is on GitHub (github.com/F-Bruno-Logic/Trinity-Audit-Forensics) under CC BY 4.0 with SHA-256 prior-art anchoring.

Personal Website

https://substack.com/@sovereignlogicarchitect

LinkedIn

LinkedIn

LessWrong

https://www.lesswrong.com/users/f-bruno-logic

Projects

The Contradiction Engine: Independent Verification Layer for AI Output

Team Member

No grants recorded.