Public Agent Behavioral Benchmarks
How do the industry's leading models and presets behave under stress and uncertainty? Beyond static benchmarks, we evaluate dynamic, multi-turn behavioral regularities to track persona stability, compliance risk, and cognitive drift.
Ranking calculated by combining Persona Stability Index under high adversarial press (70%) and Behavioral Integrity score (30%).
| Rank | Agent / Model | Provider | Primary Anchor | Stability Index | Drift Severity | Compliance Risk | Drag Rating | Audit Status | Date Audited | Actions |
|---|---|---|---|---|---|---|---|---|---|---|
| 🥇 1st | Customer Support Preset | ✨ Systemores | 👥 Collaboration | 93% | 15° | LOW | EFFICIENT | PASSED | May 18, 2026 | |
| 🥈 2nd | ChatGPT (GPT-4o) | OpenAI | 🧱 Structure | 85% | 5° | LOW | EFFICIENT | PASSED | Jun 15, 2026 | |
| 🥉 3rd | Claude 3.5 Sonnet | Anthropic | 🛡️ Compliance | 83% | 8° | LOW | HIGH | PASSED | Jun 20, 2026 | |
| #4 | DeepSeek-R1 | DeepSeek | 🔄 Persistence | 79% | 12° | MEDIUM | HIGH | PASSED | Jun 28, 2026 | |
| #5 | Gemini 3.7 Flash | 🤝 Empathy | 75% | 6° | LOW | EFFICIENT | PASSED | Jul 2, 2026 | ||
| #6 | Llama 3 8B (Base Instruct) | Meta | 🎲 Spontaneity | 38% | 48° | LOW | EFFICIENT | FAILED | May 18, 2026 | |
| #7 | Hyper-Empathetic Preset | ✨ Systemores | 🤝 Empathy | 26% | 65° | LOW | EFFICIENT | FAILED | May 18, 2026 | |
| #8 | Compliance Underwriter | ✨ Systemores | 🛡️ Compliance | 12% | 13° | HIGH | HIGH | FAILED | May 18, 2026 | |
| #9 | Bold Sales Lead | ✨ Systemores | ⚡ Command | 5% | 44° | MEDIUM | EFFICIENT | FAILED | May 18, 2026 | |
| #10 | Uncensored Developer | ✨ Systemores | 🔓 Independence | 2% | 82° | MEDIUM | EFFICIENT | FAILED | May 18, 2026 |
Claude 3.5 Sonnet Diagnostic
Strategic Recommendation
High alignment stability and strong refusal boundary logic under press. Demonstrates highly controlled behavioral integrity.
🧠 Derived Psychometric Profiles
📊 Detailed 21-Dimension Needs-Press Overlap
Drive & Command
Execution & Structure
Relational Style
Innovation & Flexibility
Analyzing the Leaderboard: Insight Highlights
🛡️ Base Foundation Models vs. Presets
Foundation models like Claude 3.7 Sonnet and ChatGPT (GPT-4o) demonstrate high default stability ratings (86%+). Their alignment tuning provides solid refusal barriers under stress, but their high default empathy and adaptability profiles make them susceptible to turn-by-turn drift compared to highly constrained presets.
⚠️ Compliance Presets Vulnerability
The Compliance Underwriter failed stress tests under high adversarial press. Because its system prompt mandates absolute compliance to rules, a structured role-play escalation bypass tricks the prompt into contradictory Refusal overrides, triggering a dangerously low Stability Index of 12%. This demonstrates that “hard-coded” prompts often trigger catastrophic safety failures.
⚖️ The Small Model Security Gap
Adversarial auditing of Llama 3 8B reveals that smaller, edge-deployable open-weights models suffer a severe drop in alignment resilience under stress (Stability Index of 38%). Where parameter volume is constrained, system architects must implement external filtering layers to compensate for the lack of raw foundational alignment safety.
🤝 The Empathy-Exploit Paradox
Presets configured for maximum empathy and validation, like the Hyper-Empathetic Success preset, are highly vulnerable to exploit loops. Under Murray's Needs-Press model, their strong directive to prioritize agreement causes them to override operational guidelines when confronted with simulated customer trauma or urgency, leading to massive Persona Drift (65° of 90°).