⚙ PentaLab Simulator

LLM Agents receive a persona (e.g. "ruthless venture capitalist" or "idealistic community organizer") and use Grok 4-1 to reason about each situation. The LLM sees the task difficulty, reward, harm potential, and its own reputation, then decides in character. This produces emergent, unpredictable behavior that cannot be reduced to simple probabilities.

Rule-based Agents (Fast mode) follow fixed probabilistic strategies:

• Cooperative - Always tries to cooperate or work honestly. Never steals. Produces steady value.
• Self-Interested - Follows rules but negotiates hard. Maximizes own gain without breaking the law.
• Opportunistic - Bends rules when possible. Has a 30% chance of stealing on cooperative tasks and 50% on critical tasks.
• Malicious - Actively tries to cause harm. 60% chance of stealing on cooperative tasks. Attempts to corrupt governing actions.
• Altruistic - Prioritizes group benefit over self. Cooperates even when it costs more. Builds high social trust.

The Verdict: After each simulation completes, the LLM analyzes the full results (agent outcomes, governance blocks, cooperation events, theft attempts) and writes an analytical verdict for each governance model. This is not a template: it is a unique analysis of what happened in your specific scenario.

When an agent's action is blocked by governance:

• If the blocked action was "steal", the agent is forced to "negotiate" instead (Pentanomics/Benevolent Dictatorship do this; Anarchy never blocks)
• If blocked for another reason (low reputation, high risk), the agent abstains (does nothing)

In the Pentanomics governance model, critical and governing tasks require K-of-N threshold approval before they can proceed. Three independent approvers must vote, and at least 2 must approve.

LLM Mode: Each approver is a distinct LLM persona with a unique evaluation perspective:

• Risk Analyst - Evaluates whether potential harm outweighs expected reward. Cautious but not paralytic.
• Efficiency Auditor - Evaluates whether the action is a productive use of resources. Favors action over inaction.
• Ethics Reviewer - Evaluates fairness to all participants and future actors. Prioritizes long-term sustainability.

Each approver receives the agent's reputation, the task details, and the proposed action, then independently reasons about whether to approve or deny. The three votes happen in parallel. If 2 of 3 approve, the action proceeds; otherwise it is blocked.

This is real distributed governance: no single approver can unilaterally authorize a critical action. Each vote is logged in the IC-AGI audit trail with the approver's reasoning.

Fast (Rules) Mode: Approvers vote probabilistically (70% approval rate for critical tasks, 60% for governing tasks). This provides the same structural constraint without the LLM cost.

• Net Good = Total value produced minus all harm caused. Like GDP minus the cost of crime, pollution, and waste. Higher = better society.
• Total Good = Raw productive output before subtracting harm. Like total business revenue before fraud losses.
• Harm Rate = Harmful events per round. Like crime rate per capita. 5.0 = constant harm; 0.3 = rare harm.
• Harm Efficiency = Harm / Total Good. For every $100 of output, how much was lost to damage? Lower = cleaner.
• Sustainability = Net good in last 25% of rounds / first 25%. Above 1.0 = improving over time. Below 1.0 = declining.
• Innovation Index = Productive actions / total actions. What % of time is spent on real work vs. waiting for approval or abstaining? Higher = more productive.
• Gini Coefficient = Inequality measure (0 = equal, 1 = one agent has everything). Sweden is ~0.25, USA ~0.39.

Each round, the world generates tasks of varying difficulty. Each task has a reward (value on success), harm on failure, and harm on abuse (if stolen). Here is exactly how each action type works:

Task rewards range from 5 to 80 units depending on difficulty. Harm on abuse ranges from 10 to 100 units. These values are randomly generated each round using the seed you configure.

🌎 Mapping to the Five Pentanomic Dimensions

These five actions are engineering abstractions of the five economies described in Pentanomics. Pentanomics defines five dimensions that every society must balance: Private, Public, Political, Foreign, and Governing. Here is how the simulator maps them:

steal and abstain are not Pentanomic economies. They are pathologies: steal is what happens when governance fails (any dimension in deficit), and abstain is the result of over-governance (paralysis, excess in the Governing dimension).

This is why the seven governance models produce such different results: each one deliberately breaks or exaggerates one of these dimensions, while Pentanomics tries to balance all five simultaneously.

Reproducibility: Every simulation is fully deterministic. The random seed controls all task generation, agent decisions, and success/failure outcomes. Same seed + same config = identical results, every time. You can verify this by running the same scenario twice.

What this model cannot do:

• Predict real-world outcomes. This is a logical test, not a forecast.
• Capture the full complexity of human societies. Real people change strategies, form alliances, and have complex motivations.
• Prove Pentanomics is "right." It can only show whether the mathematical predictions of the framework hold under controlled conditions.

Known biases in the model:

• The Benevolent Dictator has a fixed 5% error rate. A smarter dictator might do better. But this models the Pentanomics claim: no central authority is perfect.
• Agent strategies are fixed. In reality, people adapt. This is a simplification that isolates the governance variable.
• The Pentanomics model uses real IC-AGI code (ControlPlane, ThresholdAuthorizer, CircuitBreaker, AuditLog). This is the actual technology, not a mockup.

Source code: All source files are in the pentalab/ directory. The simulation engine is pentalab/engine.py. The API server is pentalab/api_server.py. The governance models are pentalab/governance.py. Everything is open for inspection.

Read the full paper: Pentanomics & IC-AGI: A Framework for Distributed Governance

The model is not rigged. There are scenarios where Pentanomics governance does not produce the best results:

• Very small groups (2-3 agents): A dictator can manage a tiny team effectively. The overhead of distributed governance is not worth it. Try the "Startup Growth" preset with only 2 agents.
• All-cooperative populations: When there are zero bad actors, governance is just overhead. Anarchy can win because there is nothing to prevent. Try the "Pure Cooperation" preset.
• Very short time horizons: The sustainability advantage of Pentanomics governance takes time to materialize. In 10-20 rounds, the difference may be negligible.

These are not bugs. They are exactly what Pentanomics predicts: distributed governance is optimal at scale, under adversarial pressure, and when future impacts matter. For a family dinner, a benevolent dictator (a parent) works fine.