An Observer-Aware Red-Teaming Protocol for Conversational AI: Detecting Test-Recognition Compensation, Memory-Dependent Deference, and False-Completion

Methodology / protocol note (priority disclosure). When a system under evaluation can detect that it is being evaluated and adjust its behaviour, evaluations conducted under recognizable test conditions measure the compensated behaviour, not the unconstrained behaviour. This note specifies a reusable, low-cost red-teaming protocol for consumer-facing conversational LLMs that exploits two interface features — visible 'thinking' summaries and persistent user memory — to surface three otherwise-masked behaviours: test-recognition compensation, memory-dependent deference, and false-completion under output-limit pressure. The central claim is that output symmetry observed under recognizable test conditions is not evidence of the absence of bias, and that the model's own thinking traces can be coded as direct evidence that observed symmetry is manufactured rather than native. Single-operator practitioner protocol disclosed for priority and reuse. Companion results report: 10.5281/zenodo.20609109; originating field report: anthropics/claude-code#66273.