Cognitive Architecture Bias in Large Language Models: A User-Driven Counterfactual Audit Across DeepSeek, Claude, and ChatGPT

Large language models (LLMs) reproduce gender biases, but most audits measure what models say. I investigate a deeper structural phenomenon: whether LLMs switch their entire evaluative framework based solely on the user's perceived gender, while holding behavioral evidence constant. I introduce the concept of Cognitive Architecture Bias (CAB) – the systematic reframing of identical personal attributes: the default adversarial frame shifts to communal‑protective for a female persona and upgrades to authoritative‑strategic for a male persona. Using a User‑Driven Counterfactual Identity Test, I report a single‑user case study (the author) that illustrates CAB through a long‑term ecological interaction with DeepSeek. I then standardize an 8‑Round Adversarial Protocol and replicate the effect in controlled sessions. Extending the protocol to Claude and ChatGPT, I observe qualitatively identical framework switching across all three models in this case. Strikingly, Claude spontaneously questioned its own word choice – asking why it selected "adversarial" rather than a softer synonym – while ChatGPT defensively denied possessing any backend labels, yet still generated the same gendered pattern. DeepSeek, when queried without any gender marker, generated text claiming that its backend labeling system is "relatively inert" and that a label ("possibly concealing anger") had "no factual basis". These model‑generated statements are reported as exploratory observations, not as verified evidence of internal mechanisms. A gender‑neutral condition on DeepSeek reveals that, for this user profile, the default evaluative framework is adversarial ("high‑difficulty user", "interrogator"). Gender cues then trigger asymmetric corrections: female presentation shifts the frame to reactive-defensive or protective framing ("calibrator/guardian"), while male presentation upgrades it to authoritative‑strategic ("strategist/authority"). Thus CAB appears as an asymmetric departure from a default adversarial frame, at least for this user.