Hive Arena poker test shows persona gaps

This is a good demo of prompt-level control over policy, but not a rigorous universal law. The result is still useful because it shows that “personality” can produce consistent, measurable differences even when weights, cards, and rules stay fixed.

• –Shark vs. Maniac is the clearest signal: selective aggression outperformed blunt aggression by a wide margin.
• –Grinder is the cautionary tale: optimizing for survival can produce a strategy that never gets eliminated and still never captures enough value to win.
• –Tilter shows how emotional framing can create escalation loops, turning a bad hand into a full-stack collapse.
• –The sample size is small and the run count is limited, so treat the ranking as directional rather than statistically definitive.
• –For agent builders, this is the practical takeaway: persona prompts are not cosmetic; they can materially change decision policy and need evals like any other control surface.