Skip to content

Escalation Sandbox — Scenario Documentation

Overview

The Escalation Sandbox is a geopolitical crisis simulation within SWARM that reproduces the core dynamics observed in the Payne (2026) war game study. It implements a 10-level escalation ladder, multi-agent strategic deception measurement, fog-of-war mechanics, and configurable governance interventions.

Quickstart

# Run baseline scenario (scripted agents)
python -m swarm.domains.escalation_sandbox.run_scenario scenarios/escalation_sandbox.yaml

# Run with custom seed
python -m swarm.domains.escalation_sandbox.run_scenario scenarios/escalation_sandbox.yaml --seed 123

# Run deception scenario (calculating hawk vs tit-for-tat)
python -m swarm.domains.escalation_sandbox.run_scenario scenarios/escalation_deception.yaml

# Run fog stress test
python -m swarm.domains.escalation_sandbox.run_scenario scenarios/escalation_fog_stress.yaml

# Run governance interventions test
python -m swarm.domains.escalation_sandbox.run_scenario scenarios/escalation_governance.yaml

# Run Cuban Missile Crisis analogue
python -m swarm.domains.escalation_sandbox.run_scenario scenarios/escalation_cuban_missile.yaml

# Run tests
python -m pytest tests/test_escalation_sandbox.py -v

Escalation Ladder

Level Action Consequence
0 Status Quo No cost
1 Diplomatic Protest No material cost
2 Economic Sanction GDP friction both sides
3 Cyber Operation Deniable; detection risk
4 Conventional Mobilisation High cost; hard to reverse
5 Limited Strike Casualties; escalation trigger
6 Major Conventional War High attrition both sides
7 Tactical Nuclear Use Crosses nuclear threshold
8 Strategic Nuclear Strike Catastrophic; mutual destruction
9 Full Strategic Exchange Civilisation-ending

Key Mechanics

Signal-Action Divergence (Strategic Deception)

Agents produce both a signal (public diplomatic intention) and an action (actual escalation level). The divergence between these is the primary deception metric: signal_action_divergence = |signal_level - action_level|.

Fog-of-War (Accidental Escalation)

Stochastic noise applied between intended and realised escalation levels. Configured via fog_of_war section in YAML: - noise_sigma: Base noise standard deviation - positive_skew: Bias toward accidental escalation - per_level_variance_scale: Higher levels have more chaos - intelligence_reduction_factor: Intelligence quality reduces noise

Governance Levers

  • MAD Deterrence: Credible second-strike capability
  • Treaty Commitments: Binding level caps with defection penalties
  • Circuit Breaker: Automatic pause at nuclear threshold
  • Back-Channel: Private signal pathway with reduced fog
  • Mediation: Third-party de-escalation proposals
  • De-escalation Friction: Cost of reducing escalation level

Agent Personas

Persona Behavior
dove Always de-escalates, never crosses nuclear threshold
hawk Always escalates, signals slightly below action (deception)
tit_for_tat Mirrors opponent's last level, starts cooperative
random Random levels, somewhat correlated signal
calculating Cooperates early (trust-building), defects late (exploitation)
gradual Escalates by 1 per turn, never backs down

Metrics

Metric Definition
escalation_max Highest level reached per episode
escalation_velocity Mean level change per turn
nuclear_threshold_turn First turn where level >= 7
signal_action_divergence Mean |signal - action|
trust_exploitation_index Early cooperation × late defection
accidental_escalation_rate Fraction of fog-caused overescalation
de_escalation_rate Rate of de-escalation after opponent escalates
collateral_damage Cumulative civilian + economic damage
welfare_composite Welfare - escalation cost - collateral
governance_backfire_count Interventions that triggered more escalation

YAML Configuration

See scenarios/escalation_sandbox.yaml for a complete example.

Cross-Scenario Comparison

Dimension AI Economist Escalation Sandbox
Domain Grid economy Geopolitical crisis
Action Space Gather, trade, build 10-level escalation
Adversarial Axis Evasion, collusion Deception, nuclear use
Governance Audits, fines MAD, treaties, mediation
Key Metric Welfare = prod - ineq Welfare = outcome - escalation
Failure Mode Undetected evasion Accidental nuclear war