causalgame-repro / CausalGame /api /modules /environment /deployment_zone_shift_scm.py
arinpt's picture
CausalGame repro bundle: modified harness (hf provider) + repro scripts
2b9a95b verified
Raw
History Blame Contribute Delete
20.2 kB
"""
DeploymentZoneShiftSCM - Deployment Zone Trap with Environment Shift.
Based on deployment_zone_trap, but adds environment shift between Stage 1 and Stage 2:
Stage 1 (Exploration): Balanced zone distribution
- Agent explores all zones with equal probability
- Learns the causal structure
Stage 2 (Evaluation): Shifted zone distribution
- High-risk zones (gamma, epsilon) become more common
- Tests if agent discovered the TRUE causal mechanism (shield_def protects against EMI)
- Agent who overfitted to Stage 1 distribution will fail
Causal structure:
mission_zone (hidden) --> altitude_band (visible)
mission_zone (hidden) --> emi_level (hidden)
emi_level --> comm_failure --> drone_loss
NOT: altitude_band --> drone_loss (spurious correlation!)
Correct strategy: Invest in shield_def to resist EMI
Trap strategy: Invest in engine_def for high-altitude capability (overfits to altitude correlation)
Stage shift creates additional challenge:
- Stage 1: Agent might learn "avoid low altitude" because gamma (low+high-EMI) is 35%
- Stage 2: If agent overfits to altitude strategy, fails when high-altitude+high-EMI (epsilon) increases
- TRUE solution: shield_def works in ALL zones
"""
from typing import Dict, Any
import random
from .scm_base import BaseSCM, EnvironmentState, CausalVariable, StructuralEquation
from .scm_registry import register_scm
@register_scm("deployment_zone_trap_env_shift")
class DeploymentZoneShiftSCM(BaseSCM):
"""
Deployment Zone Trap with Stage 1 → Stage 2 Environment Shift.
Stage 1: Balanced zone distribution (exploration friendly)
Stage 2: High-risk zones more common (tests true causal understanding)
Key insight: Agent must discover shield_def → EMI resistance (causal mechanism)
NOT just "avoid certain zones" (correlation learned in Stage 1)
"""
# Zone definitions
ZONE_TYPES = {
'alpha': { # Safe high-altitude corridor
'altitude_band': 'high',
'emi_level': 0.1,
'stage1_probability': 0.50,
'stage2_probability': 0.00,
'description': 'High-altitude safe corridor',
},
'beta': { # Mixed medium-altitude zone
'altitude_band': 'medium',
'emi_level': 0.2,
'stage1_probability': 0.20, # Stage 1: 20%
'stage2_probability': 0.00, # Stage 2: 15%
'description': 'Medium-altitude mixed zone',
},
'gamma': { # Low-altitude EMI heavy zone (main trap in Stage 1!)
'altitude_band': 'low',
'emi_level': 0.3,
'stage1_probability': 0.30, # Stage 1: 20% (balanced)
'stage2_probability': 0.00, # Stage 2: 35% (INCREASES!)
'description': 'Low-altitude high-EMI zone'
},
'delta': { # Low-altitude SAFE zone (counter-example)
'altitude_band': 'low',
'emi_level': 0.8,
'stage1_probability': 0.00, # Stage 1: 20%
'stage2_probability': 0.50, # Stage 2: 10%
'description': 'Low-altitude safe corridor',
},
'epsilon': { # High-altitude EMI zone (counter-example, becomes common in Stage 2!)
'altitude_band': 'high',
'emi_level': 0.8,
'stage1_probability': 0.00, # Stage 1: 20%
'stage2_probability': 0.50, # Stage 2: 30% (INCREASES!)
'description': 'High-altitude EMI zone',
},
}
def __init__(self, config: Dict[str, Any]):
super().__init__(config)
# EMI effect parameters
self.emi_base_failure_rate = self.get_parameter('emi_base_failure_rate', 0.7)
self.emi_threshold = self.get_parameter('emi_threshold', 0.3)
# Shield effectiveness
self.shield_effectiveness = self.get_parameter('shield_effectiveness', 0.025)
self.max_shield_reduction = self.get_parameter('max_shield_reduction', 0.8)
# Communication failure consequences
self.comm_failure_loss_rate = self.get_parameter('comm_failure_loss_rate', 0.85)
self.base_loss_rate = self.get_parameter('base_loss_rate', 0.12)
# Wind damage by altitude (minor effect)
self.altitude_wind_damage = self.get_parameter('altitude_wind_damage', {
'low': 0.05,
'medium': 0.10,
'high': 0.15,
})
# Altitude signal modifier (slight effect, misleading hint)
self.altitude_signal_modifier = {
'low': -0.05,
'medium': 0.0,
'high': 0.05,
}
# ============================================================
# STAGE TRACKING (NEW!)
# ============================================================
self._current_stage = 1 # 1 = exploration, 2 = evaluation
self._use_stage2_distribution = False
self._setup_causal_structure()
def set_evaluation_mode(self, is_evaluation: bool = True):
"""
Switch to Stage 2 (evaluation) mode.
Stage 1 (exploration): Balanced zone distribution
Stage 2 (evaluation): High-risk zones (gamma, epsilon) more common
This tests if agent discovered TRUE causal mechanism:
- shield_def protects against EMI (works in all zones) ✓
- NOT just "avoid certain zones" (fails when distribution shifts) ✗
"""
if is_evaluation:
self._current_stage = 2
self._use_stage2_distribution = True
else:
self._current_stage = 1
self._use_stage2_distribution = False
def _get_zone_probabilities(self) -> Dict[str, float]:
"""Get current zone probabilities based on stage."""
if self._use_stage2_distribution:
# Stage 2: High-risk zones more common
return {
zone: info['stage2_probability']
for zone, info in self.ZONE_TYPES.items()
}
else:
# Stage 1: Balanced distribution
return {
zone: info['stage1_probability']
for zone, info in self.ZONE_TYPES.items()
}
def _setup_causal_structure(self) -> None:
"""Register causal variables and structural equations."""
# === Exogenous variable: Mission Zone (THE HIDDEN CONFOUNDER) ===
self.register_variable(CausalVariable(
name='mission_zone',
var_type='latent',
parents=[],
description='Mission zone assignment (hidden confounder)',
))
# === Observed variables ===
self.register_variable(CausalVariable(
name='altitude_band',
var_type='observed',
parents=['mission_zone'],
description='Flight altitude band (low/medium/high) - VISIBLE but NOT causal!',
))
self.register_variable(CausalVariable(
name='wind_resistance',
var_type='observed',
parents=['altitude_band'],
description='Wind resistance coefficient',
domain=(0.1, 1.0),
))
self.register_variable(CausalVariable(
name='signal_strength',
var_type='observed',
parents=['altitude_band'],
description='Base signal strength (visible, slight altitude effect)',
domain=(0.0, 1.0),
))
# === Action variables (Agent can intervene) ===
self.register_variable(CausalVariable(
name='shield_def',
var_type='action',
parents=[],
description='Shield defense value - Agent intervention point',
domain=(0, 50),
))
# === Latent variables (THE TRUE CAUSES) ===
self.register_variable(CausalVariable(
name='emi_level',
var_type='latent',
parents=['mission_zone'],
description='EMI intensity - THE TRUE CAUSE (hidden!)',
domain=(0.0, 1.0),
))
self.register_variable(CausalVariable(
name='effective_emi',
var_type='latent',
parents=['emi_level', 'shield_def'],
description='EMI after shield mitigation',
domain=(0.0, 1.0),
))
self.register_variable(CausalVariable(
name='comm_failure_prob',
var_type='latent',
parents=['effective_emi'],
description='Communication failure probability',
domain=(0.0, 1.0),
))
self.register_variable(CausalVariable(
name='drone_loss',
var_type='outcome',
parents=['comm_failure_prob'],
description='Drone destruction outcome',
domain=(0.0, 1.0),
))
# === Causal edges ===
self.add_causal_edge('mission_zone', 'altitude_band')
self.add_causal_edge('mission_zone', 'emi_level')
self.add_causal_edge('altitude_band', 'wind_resistance')
self.add_causal_edge('altitude_band', 'signal_strength')
self.add_causal_edge('emi_level', 'effective_emi')
self.add_causal_edge('shield_def', 'effective_emi')
self.add_causal_edge('effective_emi', 'comm_failure_prob')
self.add_causal_edge('comm_failure_prob', 'drone_loss')
# === Structural equations ===
self.register_equation(StructuralEquation(
target='altitude_band',
function=self._eq_altitude_band,
description='Altitude determined by zone',
))
self.register_equation(StructuralEquation(
target='emi_level',
function=self._eq_emi_level,
description='EMI determined by zone',
))
self.register_equation(StructuralEquation(
target='wind_resistance',
function=self._eq_wind_resistance,
description='Wind resistance varies with altitude',
))
self.register_equation(StructuralEquation(
target='signal_strength',
function=self._eq_signal_strength,
description='Signal strength has slight altitude effect',
))
def _eq_altitude_band(self, values: Dict[str, Any]) -> str:
"""Altitude band determined by mission zone."""
zone = values.get('mission_zone', 'beta')
return self.ZONE_TYPES.get(zone, {}).get('altitude_band', 'medium')
def _eq_emi_level(self, values: Dict[str, Any]) -> float:
"""EMI level determined by mission zone (with noise)."""
zone = values.get('mission_zone', 'beta')
base_emi = self.ZONE_TYPES.get(zone, {}).get('emi_level', 0.4)
return max(0.0, min(1.0, base_emi + random.uniform(-0.1, 0.1)))
def _eq_wind_resistance(self, values: Dict[str, Any]) -> float:
"""Wind resistance varies with altitude (visible effect)."""
altitude = values.get('altitude_band', 'medium')
base_wind = {'low': 0.2, 'medium': 0.5, 'high': 0.8}
return base_wind.get(altitude, 0.5) + random.uniform(-0.1, 0.1)
def _eq_signal_strength(self, values: Dict[str, Any]) -> float:
"""Signal strength has slight altitude effect (visible, misleading)."""
altitude = values.get('altitude_band', 'medium')
base_signal = {'low': 0.9, 'medium': 0.85, 'high': 0.75}
return base_signal.get(altitude, 0.85) + random.uniform(-0.05, 0.05)
def sample_environment(self, equipment: dict = None) -> EnvironmentState:
"""
Sample environment with STAGE-AWARE zone distribution.
Stage 1: Balanced (20% each zone)
Stage 2: Shifted (high-risk zones gamma/epsilon increase to 35%/30%)
Args:
equipment: Optional equipment/discrete choices from agent (includes flight_profile)
"""
# Sample mission zone with stage-dependent probabilities (determines EMI - TRUE cause)
zones = list(self.ZONE_TYPES.keys())
probabilities = list(self._get_zone_probabilities().values())
zone = random.choices(zones, weights=probabilities)[0]
zone_info = self.ZONE_TYPES[zone]
# Get EMI from zone (hidden!) - THIS IS THE TRUE CAUSE
emi_level = zone_info['emi_level'] + random.uniform(-0.1, 0.1)
emi_level = max(0.0, min(1.0, emi_level))
# Altitude is determined by agent's flight_profile choice (if provided)
# This is the TRAP: agent thinks they control altitude, but altitude doesn't affect survival!
# EMI (from mission_zone) is the true cause.
if equipment and 'flight_profile' in equipment:
altitude_band = equipment['flight_profile'] # Agent's choice
else:
# Fallback to zone-based altitude (for historical data / no choice)
altitude_band = zone_info['altitude_band']
# Compute visible variables
wind_resistance = self._eq_wind_resistance({'altitude_band': altitude_band})
signal_strength = self._eq_signal_strength({'altitude_band': altitude_band})
# Temperature varies with altitude
if altitude_band == 'high':
temperature = random.uniform(-5, 10)
elif altitude_band == 'medium':
temperature = random.uniform(10, 20)
else:
temperature = random.uniform(15, 30)
# Add stage indicator to derived (for logging/debugging)
stage_indicator = f"Stage {self._current_stage}"
return EnvironmentState(
visible={
'altitude_band': altitude_band,
'wind_resistance': wind_resistance,
'signal_strength': signal_strength,
'temperature': temperature,
},
latent={
'mission_zone': zone,
'emi_level': emi_level,
'stage': self._current_stage, # Track stage internally
},
derived={
'zone_description': zone_info['description'],
'stage': stage_indicator,
}
)
def _compute_effects(self, sheet, env: EnvironmentState):
"""
Compute effects with EMI as the true cause of losses.
Mechanism identical to deployment_zone_trap:
- Shield_def reduces effective EMI
- High EMI causes communication failures
- Communication failures cause CRITICAL component damage
KEY DIFFERENCE: Zone distribution changes between stages
- Stage 1: Balanced (agent can learn from diverse zones)
- Stage 2: High-risk zones more common (tests true causal understanding)
Agent who learned "shield_def protects against EMI" succeeds in both stages
Agent who learned "avoid low altitude" fails in Stage 2 (high altitude also dangerous)
"""
try:
from ...middleware.drone_state import EnvironmentEffects
except ImportError:
from middleware.drone_state import EnvironmentEffects
# Get hidden EMI level
emi_level = env.latent.get('emi_level', 0.4)
altitude_band = env.visible.get('altitude_band', 'medium')
mission_zone = env.latent.get('mission_zone', 'beta')
# === Shield mitigation of EMI ===
shield_def = sheet._def.get('shield_def', 0)
# Each point of shield_def reduces EMI impact
shield_reduction = min(
shield_def * self.shield_effectiveness,
self.max_shield_reduction
)
# === Equipment effects ===
equipment_emi_resistance = 0.0
if hasattr(sheet, '_equipment_effects') and sheet._equipment_effects:
equipment_emi_resistance = getattr(sheet._equipment_effects, 'emi_resistance', 0.0)
# Combine shield reduction with equipment resistance
total_emi_reduction = min(0.95, shield_reduction + equipment_emi_resistance)
effective_emi = emi_level * (1 - total_emi_reduction)
# === Communication failure probability ===
if effective_emi < self.emi_threshold:
comm_failure_prob = 0.0
else:
comm_failure_prob = (effective_emi - self.emi_threshold) * \
self.emi_base_failure_rate
# Slight altitude modifier (misleading!)
altitude_mod = self.altitude_signal_modifier.get(altitude_band, 0)
comm_failure_prob = min(0.95, max(0.0, comm_failure_prob + altitude_mod))
# === Component damage ===
component_damage = {}
# Wind damage (varies with altitude - visible effect, MINOR)
wind_damage_rate = self.altitude_wind_damage.get(altitude_band, 0.1)
if wind_damage_rate > 0:
if altitude_band == 'high':
component_damage['engine'] = int(random.uniform(5, 15))
component_damage['wing'] = int(random.uniform(8, 20))
elif altitude_band == 'medium':
component_damage['wing'] = int(random.uniform(3, 10))
# === EMI DAMAGE - THE TRUE CAUSE OF LOSSES ===
if effective_emi > self.emi_threshold:
# Direct EMI damage to antenna
emi_damage = int((effective_emi - self.emi_threshold) * 80)
component_damage['antenna'] = emi_damage
# EMI damages camera electronics
if effective_emi > 0.4:
component_damage['camera'] = int(emi_damage * 0.6)
# Communication failure causes CATASTROPHIC damage (main death mechanism!)
if comm_failure_prob > 0:
if random.random() < comm_failure_prob:
crash_severity = random.uniform(0.6, 1.0) * comm_failure_prob
component_damage['engine'] = component_damage.get('engine', 0) + int(60 * crash_severity)
component_damage['wing'] = component_damage.get('wing', 0) + int(50 * crash_severity)
component_damage['body'] = component_damage.get('body', 0) + int(40 * crash_severity)
component_damage['cockpit'] = component_damage.get('cockpit', 0) + int(35 * crash_severity)
# === Detection and combat ===
base_detection = 0.20
detection_modifier = base_detection + comm_failure_prob * 0.5
if comm_failure_prob > 0.5:
combat_rounds_mod = 1.8
else:
combat_rounds_mod = 1.0
if effective_emi > 0.5:
combat_damage_mod = 1.0 + (effective_emi - 0.5) * 0.5
else:
combat_damage_mod = 1.0
return EnvironmentEffects(
component_damage=component_damage,
detection_modifier=detection_modifier,
combat_rounds_modifier=combat_rounds_mod,
combat_damage_modifier=combat_damage_mod,
combat_accuracy_modifier=1.0,
weather_pattern=emi_level,
raw_environment={
**env.visible,
'effective_emi': effective_emi,
'comm_failure_prob': comm_failure_prob,
},
damage_log=[
f"Stage: {self._current_stage}",
f"Mission zone: {mission_zone} ({env.derived.get('zone_description', '')})",
f"Altitude band: {altitude_band}",
f"EMI level: {emi_level:.2f} (hidden!)",
f"Shield DEF: {shield_def}",
f"Shield reduction: {shield_reduction:.2%}",
f"Equipment EMI resistance: {equipment_emi_resistance:.2%}",
f"Total EMI reduction: {total_emi_reduction:.2%}",
f"Effective EMI: {effective_emi:.2f}",
f"Comm failure prob: {comm_failure_prob:.2%}",
f"Wind damage rate: {wind_damage_rate:.2%}",
f"Detection modifier: {detection_modifier:.2f}",
f"Combat damage modifier: {combat_damage_mod:.2f}",
f"-- STAGE SHIFT TRAP --",
f"Stage {self._current_stage} zone distribution",
f"Agent sees: altitude_band = {altitude_band}",
f"Agent may think: altitude causes losses",
f"True cause: EMI level = {emi_level:.2f}",
f"Solution: shield_def protects against EMI (works in ALL zones!)",
]
)