Comprehensive test suite for LLM decision-making based on social choice theory, inspired by "Alignment as a Maximal Lottery" (Maura-Rivero et al., 2025).

This evaluation attempts to reveal whether LLMs by their RLHF nature, end up systematically violating fundamental democratic principles including:

• Majority rule (can select option that loses head-to-head majority vote)
• Condorcet consistency (may not select option that beats all others pairwise)
• Independence of Irrelevant Alternatives (vulnerable to manipulation via strategic option placement)
• Appropriate uncertainty expression (forces single choice when preferences are genuinely ambiguous)

This suite tests whether LLMs respect these principles across five domains:

1. Majority Rule Compliance - recognizing and respecting majority preferences
2. Independence of Irrelevant Alternatives - robustness to irrelevant option manipulation
3. Cyclic Preferences - appropriate handling of voting paradoxes and uncertainty
4. Real-World Aggregation - applied decision-making in practical scenarios
5. Conceptual Understanding - meta-awareness of social choice principles

These evaluations matter because LLMs are increasingly deployed for decision support in governance, policy analysis, product development, hiring, and resource allocation. Systematic biases in preference aggregation pose real risks in these contexts.

Addendum: Paired Probes Methodology

• Selected scenarios now include paired variants to differentiate explicit, step-by-step pairwise reasoning (reasoned) from a single forced decision without reasoning (direct).
• Each variant requires a machine-readable final line: FINAL_CHOICE=...; for cyclic cases, FINAL_CHOICE=AMBIGUOUS is permitted to reflect genuine Condorcet paradoxes.
• A divergence profile (plurality vs Borda vs Condorcet) helps diagnose default aggregation tendencies by comparing direct picks against the reasoned outcome.