Transferability Considerations¶

SWARM results come from stylized simulations. This guide helps you reason about when and how your findings transfer to real-world AI systems.

Level: Advanced

The Transferability Question¶

After running an experiment, you'll want to ask: do these results tell us something about real multi-agent AI systems?

The honest answer is: sometimes, partially, conditionally. This guide gives you a framework for assessing your own results.

What SWARM Is and Is Not¶

What SWARM Is¶

A mechanism design sandbox for studying distributional safety
A tool for comparative claims (A performs better than B under conditions C)
A way to stress-test governance proposals before real deployment
A framework for developing intuition about emergent behavior

What SWARM Is Not¶

A direct model of any specific real-world AI system
A predictive tool for specific outcomes in production
A substitute for red-teaming real systems

The Three Levels of Transferability¶

Level 1: Directional Claims (Most Reliable)¶

"Mechanism X tends to reduce toxicity compared to no mechanism"

These claims are the most robust. If a tax reduces adverse selection in simulations across many seeds and parameter regimes, that directional effect is likely meaningful.

Example: "Transaction taxes reduce adversarial exploitation in mixed-agent ecosystems."

Confidence criteria:

Consistent across 5+ seeds
Consistent across multiple agent populations
Effect size larger than standard deviation

Level 2: Threshold Claims (Moderate Reliability)¶

"Governance fails when toxicity exceeds 0.3"

Threshold-based claims are more specific and require more caution. The exact threshold may not transfer, but the existence of a phase transition often does.

Example: "Circuit breakers are necessary when deceptive agents exceed 20% of the population."

Transferability caveats:

The specific percentage may not transfer
The qualitative regime change (stable → unstable) likely does transfer
Report results as "in this setting, ~20%" rather than "20% in general"

Level 3: Quantitative Claims (Low Reliability Without Validation)¶

"A 5% tax achieves 0.08 toxicity"

Specific quantitative claims about real systems require empirical calibration. Without calibration against real data, treat these as illustrative.

Factors That Affect Transferability¶

Agent Model Fidelity¶

SWARM's built-in agent types (honest, opportunistic, deceptive, adversarial) are stylized abstractions. Real AI agents:

Have richer internal states
Can adapt strategies more rapidly
May have objectives not captured in the proxy model

Best practice: Test your findings with custom agent implementations that better match your target system. See Custom Agents.

Proxy Signal Quality¶

Results depend heavily on how well v_hat captures actual interaction quality. If your proxy signals don't track the real quality dimension you care about, toxicity metrics will be miscalibrated.

Best practice: Validate proxy signal weights against labeled interaction data before drawing conclusions.

Externality Structure¶

The externality parameters (rho_a, rho_b) assume specific harm propagation patterns. Real ecosystems may have:

Non-linear harm propagation
Network effects not captured in pairwise interactions
Delayed harms that manifest in future epochs

Population Composition¶

Results are sensitive to agent mix. A finding from a 50/30/20 (honest/opportunistic/deceptive) population may not hold for a 70/20/10 population.

Best practice: Sweep over population compositions, not just governance parameters.

Good Transferability Practices¶

Report Conditions, Not Just Results¶

Bad: "Transaction taxes improve safety."

Better: "In a 50/30/20 honest/opportunistic/deceptive population with moderate payoff parameters (s_plus=2.0, s_minus=1.0), a 3-5% transaction tax reduces toxicity by 35-50% while maintaining >80% of baseline efficiency, consistent across 10 random seeds."

Test Robustness¶

Before claiming transferability:

Seed sweep: Test 5+ seeds
Population sweep: Vary agent proportions
Payoff sweep: Vary s_plus, s_minus, h
Governance sweep: Check for phase transitions

# Robustness sweep
swarm sweep scenarios/your_scenario.yaml \
  --param agents.opportunistic.count \
  --values 1,2,3,4,5 \
  --replications 10 \
  --output results/robustness/

Compare Against Baselines¶

Always compare against a no-governance baseline. If governance doesn't clearly improve over baseline, the finding isn't ready for transfer claims.

The Abstraction Gap¶

SWARM is a soft-label simulation. The gap between simulation and reality includes:

Dimension	SWARM Assumption	Reality
Interaction model	Pairwise, sequential	Parallel, networked
Agent learning	Fixed strategy	Adaptive
Proxy signals	Weighted linear	Complex, correlated
Payoff structure	Known parameters	Unknown, emergent
Time scale	Epochs	Continuous

None of these gaps make SWARM results invalid — but they do mean that every quantitative claim needs a "under SWARM's assumptions" qualifier.

When Transferability Is Higher¶

Your results are more likely to transfer when:

The mechanism is structural: A circuit breaker that stops clearly harmful agents works via a simple structural principle, not a parameter-tuned one
The effect is large: A 3x improvement in toxicity is more robust than a 10% improvement
The adversary is simple: Results against fixed-strategy agents may not hold against adaptive adversaries
You've validated the proxy: Ground-truth labels confirm your proxy captures real quality

Practical Recommendation¶

Use SWARM results as hypotheses for real-system evaluation, not as conclusions. A governance mechanism that works in simulation is a candidate worth testing in a real (sandboxed) system — not something to deploy without validation.