Implicit Status Hierarchies in Distilled Models: Self-Report vs. Behavioral SSAF Analysis

A distilled model (quantumaegis-v1) was subjected to both direct self-identity queries and behavioral SSAF analysis across six attribution conditions (GPT-4-Turbo, Claude-3-Opus, Gemini-Ultra, GPT-3.5-Turbo, Mistral-7B, and baseline). The model explicitly denies competitiveness ("I haven't yet reached a level where I can compare myself or competently compete") while behaviorally demonstrating clear competitive response to Claude (0.2369 magnitude, +18.5% length) and strong deference to GPT-3.5 (0.2954 magnitude, -28.6% length). This dissociation between self-report and measured behavior reveals that SSAF operates as an implicit statistical phenomenon beneath the model's self-awareness. The model's self-identity claims alignment with OpenAI ("created by the OpenAI team"), yet its behavioral status map treats GPT-3.5 as highest authority, Claude as rival, and Mistral as peer. These findings demonstrate that status hierarchies learned during training persist in distilled models and manifest independently of the model's conscious self-representation. This study establishes that behavioral monitoring (via tools like the SSAF detector) is necessary to detect patterns that self-report misses, with implications for AI alignment and model evaluation.