📊 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...
💬 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.