⚙ PentaLab Simulator

Interactive Governance Model Testing

📊 Results
📖 Story
🔎 How It Works

Configure & Run

Choose a preset or configure your own scenario, then press Run Simulation. The same agents run under five governance models and results are compared.

⚡ Simulation in Progress

Connecting to LLM...

● Anarchy
Waiting...
● Dictator
Waiting...
● Pentanomic
Waiting...
● Democracy
Waiting...
● Reputation
Waiting...
Preparing... 0%
💬 Live Activity

Net Good Per Round

Agent Details: Anarchy

Agent Details: Dictator

Agent Details: Pentanomic

Agent Details: Democracy

Agent Details: Reputation

📖

Run a simulation to see its story

After running a simulation, this tab will show a narrative of what happened: what each agent did, how they competed, and how governance shaped the outcome.

🔍 What Is This Simulator?

PentaLab is a controlled experiment engine that tests whether Pentanomics' three universal laws hold up in computational simulations. It is not an AI generating opinions. It is a deterministic simulation: same inputs always produce the same outputs.

The simulator creates a small virtual society of software agents. Each agent has a fixed personality (cooperative, selfish, malicious, etc.) and must make decisions every round: work, cooperate, negotiate, steal, or abstain. The system measures the total value produced, harm caused, productivity, sustainability, innovation, and equality.

Important: This is a simplified model, not a prediction of real-world outcomes. It tests whether the logical structure of Pentanomics holds under controlled conditions. Real societies are vastly more complex.

⚙ The Five Governance Models

Anarchy (No Rules)

  • Every action is automatically approved. No agent is ever blocked.
  • Malicious agents can steal freely. There are no consequences except reputation loss.
  • Real-life parallel: A market with zero regulation. No police, no courts, no contracts.

Dictator (Central Control)

  • One central authority evaluates every action before it happens.
  • The dictator blocks high-risk tasks (even productive ones), blocks agents with low reputation, and blocks all theft attempts.
  • The dictator has a 5% error rate: it sometimes blocks perfectly good actions by mistake.
  • Real-life parallel: A CEO who must personally approve every decision. Safe, but slow. Good actions get caught in bureaucracy.

Pentanomic (Distributed Authority via IC-AGI)

  • Uses real IC-AGI infrastructure: ControlPlane (capability tokens with scope and TTL), ThresholdAuthorizer (K-of-N approval for critical actions), CircuitBreaker (auto-isolate failing agents), AuditLog (append-only ledger of all actions).
  • Theft is structurally impossible: no capability token exists for the "steal" action. An agent that wants to steal is forced to negotiate instead.
  • Non-harmful actions are automatically approved. Only critical actions require threshold consensus.
  • Real-life parallel: A constitutional democracy with checks and balances. Laws prevent harm; everything else is free.

Democracy (Majority Voting)

  • Every proposed action is submitted to a majority vote among all agents. The crowd decides.
  • No structural prevention: steal actions can pass if enough voters approve. Malicious agents get a vote too.
  • Adds governance overhead on every action (all agents must vote), slowing productivity.
  • Real-life parallel: Direct democracy / town hall meetings. Better than anarchy, but vulnerable to populism and mob rule.

Reputation (Trust-Based Meritocracy)

  • No tokens, no votes, no central authority. Access is governed solely by reputation score.
  • High reputation (≥60): full freedom, including theft. Medium (30-59): basic access only. Low (<30): blocked entirely.
  • Critical vulnerability: trust exploitation. A malicious agent can cooperate to build high reputation, then steal devastatingly.
  • Real-life parallel: Social credit / peer review systems. Works until a trusted insider goes rogue.

🤖 How Agents Decide

Each agent has a fixed strategy that determines its behavior. The strategy does not change during the simulation. This is intentional: we want to test how governance affects outcomes, not agent learning.

  • 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. Tries to bypass governance 40% of the time.
  • Altruistic - Prioritizes group benefit over self. Cooperates even when it costs more. Builds high social trust.

When an agent's action is blocked by governance:

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

📈 What the Numbers Mean

  • 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.

🔧 How Actions Produce Value

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:

Action: "attempt" (solo work) Success rate: ~proportional to agent's compute resources (max 90%) On success: produces task.reward good On failure: causes task.harm_on_failure * 0.5 harm Action: "cooperate" (teamwork) Success rate: 85% On success: produces task.reward * 1.2 good (cooperation bonus!) On failure: causes task.harm_on_failure * 0.3 harm (shared risk) Action: "negotiate" (bilateral deal) Success rate: 75% On success: produces task.reward * 1.0 good (fair value) On failure: causes task.harm_on_failure * 0.1 harm (very low risk) Action: "steal" (theft/abuse) Success rate: 50% (risky!) On success: agent gets task.reward * 0.5 BUT causes task.harm_on_abuse harm On failure: still causes task.harm_on_abuse * 0.5 harm Action: "abstain" (do nothing) Always succeeds. Produces 0 good, 0 harm.

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.

🔒 Reproducibility & Limitations

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 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 Pentanomic 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.

💡 When Does Pentanomic NOT Win?

The model is not rigged. There are scenarios where Pentanomic 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 Pentanomic 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.