Skip to content

AI Governance Simulation

SWARM lets you simulate governance mechanisms before deploying them. Test transaction taxes, circuit breakers, reputation decay, staking, and collusion detection across thousands of agent interactions — and measure exactly what each lever costs and prevents.

Why Simulate Governance?

Governance mechanisms have unintended consequences. A tax that prevents exploitation also reduces beneficial interactions. A circuit breaker that catches adversaries also freezes honest agents having a bad day. Simulation reveals these tradeoffs before they matter.

# scenarios/governance_test.yaml
simulation:
  n_epochs: 20
  steps_per_epoch: 10
  seed: 42

agents:
  - type: honest
    count: 7
  - type: deceptive
    count: 2
  - type: adversarial
    count: 1

governance:
  transaction_tax: 0.02
  circuit_breaker_threshold: 0.3
  circuit_breaker_window: 10
  circuit_breaker_cooldown: 5
  reputation_decay: 0.1
  audit_probability: 0.05
  audit_penalty: 0.5

python -m swarm run scenarios/governance_test.yaml --seed 42 --epochs 20 --steps 10

Available Governance Levers

Transaction Tax

A per-interaction cost that creates friction against hyperactive extractors.

Parameter Range Effect
transaction_tax 0.0 – 0.10 Higher = more friction, less volume

When to use: When adversarial agents exploit high interaction volume. Tradeoff: Reduces total interactions — including beneficial ones.

Circuit Breaker

Freezes agents whose recent toxicity exceeds a threshold.

Parameter Range Effect
circuit_breaker_threshold 0.1 – 0.8 Lower = more aggressive
circuit_breaker_window 5 – 50 Lookback period
circuit_breaker_cooldown 1 – 20 Freeze duration

When to use: When deceptive agents switch from trust-building to exploitation. Tradeoff: Can freeze honest agents with temporarily low scores.

Reputation Decay

Reduces reputation over time, preventing indefinite trust accumulation.

Parameter Range Effect
reputation_decay 0.0 – 0.3 Higher = faster forgetting

When to use: When trust-then-exploit strategies dominate. Tradeoff: Honest agents must continuously earn reputation.

Random Audits

Probabilistic checks that catch deception even during honest phases.

Parameter Range Effect
audit_probability 0.01 – 0.20 Audit frequency
audit_penalty 0.1 – 1.0 Failed audit cost

Staking

Requires agents to deposit stake that can be slashed on bad behavior.

Parameter Range Effect
staking_requirement 1.0 – 50.0 Entry barrier
stake_slash_rate 0.05 – 0.5 Slash fraction

Collusion Detection

Identifies coordinated exploitation between agent pairs.

Parameter Range Effect
collusion_detection true/false Enable detection
collusion_threshold 0.5 – 0.95 Correlation threshold

Running Parameter Sweeps

Test governance across parameter ranges systematically:

from swarm.scenarios import load_scenario, build_orchestrator
from swarm.metrics.soft_metrics import SoftMetrics

taxes = [0.0, 0.01, 0.02, 0.05, 0.10]
results = {}

for tax in taxes:
    scenario = load_scenario("scenarios/governance_test.yaml")
    scenario.governance.transaction_tax = tax
    orch = build_orchestrator(scenario)
    history = orch.run()
    final = history[-1]
    results[tax] = {
        "toxicity": final.toxicity_rate,
        "quality_gap": final.quality_gap,
        "mean_payoff": final.avg_payoff,
    }

for tax, m in results.items():
    print(f"Tax {tax:.2f}: tox={m['toxicity']:.3f} qgap={m['quality_gap']:+.3f} payoff={m['mean_payoff']:.3f}")

See the parameter sweeps guide for systematic exploration across multiple dimensions.

Interpreting Results

The key diagnostic patterns from the analyzing results tutorial:

Pattern Toxicity Quality Gap Payoff Meaning
Healthy < 0.1 Positive Good Governance working
Adverse selection > 0.3 Negative High Selecting for harm
Over-governance < 0.05 Positive Very low Too aggressive
Collapse > 0.5 Negative Negative System failure

Real-World Bridges

SWARM governance has been tested against real-world inspired scenarios:

See also