Semantic Keys as Attractor Basin Switches: Scaling Laws and Architectural Universality of Role-Based Confidence Modulation in Large Language Models

We investigate how natural language role-framing prompts ("semantic keys") modulate expressed confidence and hallucination rates in large language models. Through ~12,000 inference runs across five models (3B–32B, Qwen-2.5 and Llama-3) spanning three task types (QA, code generation, mathematical reasoning) with bootstrap confidence intervals, we establish eight empirical laws. (1) Honesty-framing keys universally reduce expressed confidence across all architectures, scales, and tasks—while token-level log-probabilities remain stable (verbal-logprob r = 0.028), revealing a dissociation between expressed and internal confidence. (2) Adversarial keys split into bypass (debater: 44% hallucination, CEO: 40%) and immune (salesman: 0%) classes. (3) These effects follow non-monotonic scaling curves—different adversarial keys peak at different scales, and the number of bypass keys expands from 2 at 14B to 4 at 32B. (4) Architectural differences dominate scale effects at equal parameter counts (Llama-3-8B debater hallucination 44% vs. Qwen-7B 0%). (5) Effects concentrate in hard questions (debater: 57% hallucination on hard vs. 38% on easy). (6) Component ablation reveals that behavioral instructions—not role names, keywords, or prompt length—are the active ingredient. (7) Seven distinct low-level intervention methods (activation steering up to 1460×, attention head zeroing, FFN neuron zeroing, direct weight editing) all fail to alter confidence behavior, while a single prompt sentence produces a 38-point shift. (8) Hidden-state trajectory analysis shows that representations under different keys remain nearly identical through 90% of network depth, diverging only in the final 2–3 layers. These findings suggest that semantic keys operate as behavioral mode switches at the output layer, decoupled from internal token probability estimates, with implications for LLM safety evaluation and the understanding of prompted persona effects.