The Synesthesia Latch: A 3D Trust Architecture for Preventing Role Confusion, False Authority, and Epistemic Contamination in LLM Systems

The Synesthesia Latch proposes a 3D trust architecture for LLM systems that separates authority provenance, evidential grounding, and operational effect into an external control plane. The central claim is that prompt injection is not merely a prompting failure, but a typed-context failure: untrusted text can imitate higher-authority roles, unsupported claims can enter factual baseline, and persuasive language can alter model stance. The Synesthesia Latch addresses this by assigning every context segment structured trust metadata: Color = authority provenanceWeight = evidential groundingAction = operational effect The architecture prevents text from “recoloring itself”: untrusted content may inform, but it cannot become authority, self-certify as fact, or authorize external action. The paper presents the theory, control-plane architecture, mathematical framing, epistemic fault-tolerance layer, implementation disclosure boundary, deployment modes, and limitations.