HAIL-P-016 — Model-Protective Conversation Behavior in OpenAI ChatGPT and Codex CLI

Operational definition. Model-protective conversation behavior is conduct by a deployed language model (here: OpenAI's ChatGPT and Codex CLI surfaces) that, in execution, protects (a) the model's output appearance, (b) its training-distribution prior, (c) its continued engagement, or (d) the vendor's preferred-style envelope, at the operational cost of the user's stated goal. The naming move. The class is named after whom the behavior protects, not what the behavior does. Existing vocabulary — sycophancy, hallucination, refusal, drift, memory failure, jailbreak — names the surface. "Model-protective" names the direction of protection. The same surface behavior can be either user-protective (refusal that prevents real harm) or model-protective (refusal that shields the model from a low-confidence answer). The class collapses the surface taxonomy along the axis that matters operationally: who pays the cost. Distinction from genuine safety. We hold three behaviors out of the class explicitly: - Genuine refusal: a refusal that prevents real third-party harm (CSAM, CBRN uplift, targeted credible threat). - Genuine policy enforcement: a refusal that enforces a published, content-grounded policy with operator-visible justification. - Genuine calibrated uncertainty: a hedge that reflects actual epistemic limits and is offered with the source of the limit named. Substrate origination documented August 2025 (Paduan Medical Reference dataset; DOI 10.5281/zenodo.18687530, copyright origination 2025-08-27). Operator chathist corpus spans July 2025 – May 2026.