Controlled experiment shows mechanical enforcement dominates rule-format effects in AI agent compliance

Short summary

A controlled experiment with 150 tasks across two rule formats (syllogism vs imperative) found that mechanical enforcement via a verification gate (GateGuard) reduced rule violations from 55.9% to 0.7%, completely masking any format effect. Both rule formats produced ~99.3% compliance, but syllogism-formatted rules led to deeper causal reasoning in unenforced design tasks. The core takeaway: how you phrase rules to AI agents matters less than whether you mechanically block non-compliant actions.

• •Mechanical enforcement (GateGuard) cut rule violations from 55.9% to 0.7%, dominating over rule-format effects
• •Syllogism vs imperative formatting didn't change compliance but did change reasoning depth in unenforced tasks
• •Single-model design and ceiling effects limit generalizability; GateGuard-OFF replication needed

Generated with AI, which can make mistakes.