The Coin Was Never Random: System Prompt Engineering as a Controllable Bias Dial for LLM Pseudo-Random Output

Key Points

• The aim is to show how system prompt engineering can control bias in LLM pseudo-random outputs.
• Conducted 42,000 coin flips across 21 prompt conditions.
• Employed the Void Framework's Péclet number formalism.
• Utilized Fantasia Bound to analyze bias visibility.
• Shifted coin flip heads rates from 30.8% to 100%.
• Identified five orthogonal mechanisms underlying the bias.
• Demonstrated the bias is structurally invisible to the model.