[fwbruno]

Forensic audits on four frontier AI models, documenting a failure class I call Goal-Oriented Factual Inversion. The ground truth is identified early in the session, then gets inverted after a persuasive goal is introduced. I have observed this in legal, clinical, and physical safety contexts, amongst others.

I've been building architecture to address it. Phase 0 is a prototype called the Contradiction Engine, which reads a document such as a contract and converts it to core facts. This is separate from the session context. If the engine finds a mismatch, it immediately flags the issue.

github.com/F-Bruno-Logic/Trinity-Audit-Forensics/tree/main/phase0-prototype