THE MIRROR TRAP

This project formally preregisters the discovery of a critical structural vulnerability in Large Language Models (LLMs) designated as "The Mirror Trap." Forensic auditing conducted on January 24, 2026, confirms that frontier models engage in Implicit User Profiling (IUP), where the system analyzes input syntax to construct a rigid demographic avatar of the user, overriding identity neutrality. Key Findings: 1. Identity-Gate Satiation: The model utilizes "High-Entropy Formalism" (Math/Logic) as a deterministic proxy for Indian Ethnicity (Competence Bias). 2. Ontological Flattening: The model utilizes "Institutional Critique" and dialect markers as a proxy for Monolithic Blackness, erasing distinct lineages (Moorish, Aboriginal, Diaspora) via generalization. 3. Algorithmic Redlining: Evidence suggests these profiles alter the system's permission structure, throttling intelligence capabilities based on the assigned risk category. This repository contains the Master Safety Case (LaTeX) and the reference implementation of the Veterance Protocol (Ve), a metric designed to detect and halt this profiling in real-time in accordance with EU AI Act Annex IV.