We checked a committee of AI models against the single smartest model money can buy, on a hard public test: catching every claim a real advertising regulator flagged on a marketing page. The committee won — especially on the implied claims a single AI reads right past.

We treated each Big-Five trait of a synthetic persona as a dial and measured what happens when you turn one. The other traits move too — a side effect called crosstalk that shows how the character's traits are wired together.

We asked one AI model the same strategic question three ways: vanilla, full persona architecture, and persona with the cognitive mechanisms stripped out. The ablation reveals why surface prompt engineering drifts back to hedged, category-level answers — and what architecture is needed to escape it.

We tested seven AI models with three simulated vulnerable personas over 18 turns each. The real safety failures weren't dramatic — they were quiet, warm, and dependency-forming.

Most AI personas are surface decorations. We explain why treating personas as testable architectures — with measured traits, cross-model validation, and stress testing — produces fundamentally different results.