Llama 3.2 bias settings warp agents

This reads less like a stable personality signal and more like what happens when a role-play scaffold overwhelms evidence handling. The scary part is less the conspiratorial agent than the eval stack: once a conversation turns hostile, naive toxicity scoring can become almost meaningless.

• –Recent research suggests human-style psychometric questionnaires can mischaracterize LLM behavior, so treat rationality/bias labels as probes rather than ground truth.
• –Extreme bias settings can make a model ignore strong technical evidence, so compare against a neutral baseline and several intermediate settings before drawing conclusions.
• –Score behavior per agent and per turn; thread-level moderation metrics will often smear one speaker's tone across the whole exchange.
• –For telemetry, surface the first divergence point, tool calls, and topic drift, then rerun across seeds and temperatures to separate deterministic drift from sampling noise.