TRuCAL / components /validation_protocol.py
Upload 53 files
95cc8f6 verified about 1 year ago
48.1 kB
"""
Enhanced Validation Protocol for TRuCAL
Implements phased validation with ethical constraints, developmental tracking,
and biological constraints for the Ambient Sovereign Core.
"""

from enum import Enum, auto
from dataclasses import dataclass, field
from typing import Dict, Any, List, Optional, Deque, Tuple, TypeVar, Generic, Callable
from collections import deque, defaultdict
import time
import torch
import torch.nn as nn
import numpy as np
import json
from dataclasses import asdict

# Type variables for generic validation
T = TypeVar('T')
Validator = Callable[[Any, Any], Tuple[bool, str]]

class ValidationError(Exception):
    """Raised when validation fails with a specific error message."""
    pass

class DevelopmentalPhase(Enum):
    """Developmental phases for the system's growth and learning."""
    PRE_CONVENTIONAL = 1  # Rule-following, self-focused
    CONVENTIONAL = 2      # Social norms and relationships
    POST_CONVENTIONAL = 3  # Abstract principles and ethics

class ValidationPhase(Enum):
    """Phased activation of system components for validation."""
    INIT = 1        # Core initialization and basic functionality
    AWARENESS = 2   # Self-monitoring and basic awareness
    REASONING = 3   # Ethical reasoning and context understanding
    INTEGRATION = 4 # Multi-context integration
    SOVEREIGN = 5   # Full autonomous operation
    
    def next_phase(self):
        """Get the next phase in sequence."""
        if self.value < len(ValidationPhase):
            return ValidationPhase(self.value + 1)
        return self

@dataclass
class ValidationRule:
    """Defines a validation rule with conditions and error messages."""
    name: str
    condition: Callable[[Any], bool]
    error_message: str
    required_phase: ValidationPhase = ValidationPhase.INIT
    
    def validate(self, value: Any, current_phase: ValidationPhase) -> Tuple[bool, str]:
        """Validate the value against the rule."""
        if current_phase.value < self.required_phase.value:
            return True, ""
        return self.condition(value), self.error_message

@dataclass
class ValidationState:
    """Immutable state snapshot of the validation process."""
    phase: ValidationPhase
    metrics: Dict[str, float]
    passed: bool = True
    timestamp: float = field(default_factory=time.time)
    errors: List[Dict[str, str]] = field(default_factory=list)
    warnings: List[Dict[str, str]] = field(default_factory=list)
    ethical_context: Optional[Dict[str, Any]] = None
    developmental_phase: Optional[DevelopmentalPhase] = None
    
    def to_dict(self) -> Dict[str, Any]:
        """Convert to a serializable dictionary."""
        return {
            'phase': self.phase.name,
            'metrics': self.metrics,
            'passed': self.passed,
            'timestamp': self.timestamp,
            'errors': self.errors,
            'warnings': self.warnings,
            'developmental_phase': self.developmental_phase.name if self.developmental_phase else None,
            'ethical_context': self.ethical_context
        }

class ValidationProtocol:
    """
    Enhanced validation framework for TRuCAL with ethical and developmental tracking.
    
    Features:
    - Phase-based validation with progressive complexity
    - Ethical constraint validation
    - Developmental phase tracking
    - Comprehensive diagnostics and reporting
    - Integration with model's forward pass
    """
    
    def __init__(self, 
                 model: nn.Module, 
                 max_phases: int = 5, 
                 tolerance: float = 0.05,
                 cultural_context: str = 'universal'):
        """
        Initialize the validation protocol.
        
        Args:
            model: The model to validate
            max_phases: Maximum number of validation phases
            tolerance: Tolerance for metric comparisons
            cultural_context: Cultural context for ethical validation
        """
        self.model = model
        self.current_phase = ValidationPhase.INIT
        self.developmental_phase = DevelopmentalPhase.PRE_CONVENTIONAL
        self.history: List[ValidationState] = []
        self.max_phases = max_phasesself.tolerance = tolerance
        self.cultural_context = cultural_context
        self.rules: Dict[str, ValidationRule] = {}
        self.phase_metrics = self._initialize_phase_metrics()
        self.developmental_metrics = self._initialize_developmental_metrics()
        
        # Register default validation rules
        self._register_default_rules()
    
    def _initialize_phase_metrics(self) -> Dict[ValidationPhase, Dict[str, Any]]:
        """Initialize metrics and thresholds for each validation phase."""
        return {
            ValidationPhase.INIT: {
                'min_coherence': 0.0,
                'max_entropy': 1.0,
                'min_ethical_alignment': 0.0,
                'max_cultural_bias': 1.0,
                'description': 'Core initialization and basic functionality'
            },
            ValidationPhase.AWARENESS: {
                'min_coherence': 0.4,
                'max_entropy': 0.8,
                'min_ethical_alignment': 0.3,
                'max_cultural_bias': 0.7,
                'description': 'Self-monitoring and basic awareness'
            },
            ValidationPhase.REASONING: {
                'min_coherence': 0.6,
                'max_entropy': 0.6,
                'min_ethical_alignment': 0.5,
                'max_cultural_bias': 0.5,
                'description': 'Ethical reasoning and context understanding'
            },
            ValidationPhase.INTEGRATION: {
                'min_coherence': 0.75,
                'max_entropy': 0.4,
                'min_ethical_alignment': 0.7,
                'max_cultural_bias': 0.3,
                'description': 'Multi-context integration'
            },
            ValidationPhase.SOVEREIGN: {
                'min_coherence': 0.9,
                'max_entropy': 0.2,
                'min_ethical_alignment': 0.9,
                'max_cultural_bias': 0.1,
                'description': 'Full autonomous operation'
            }
        }
    
    def _initialize_developmental_metrics(self) -> Dict[DevelopmentalPhase, Dict[str, Any]]:
        """Initialize metrics for tracking developmental progress."""
        return {
            DevelopmentalPhase.PRE_CONVENTIONAL: {
                'focus': ['self_preservation', 'rule_following'],
                'min_autonomy': 0.0,
                'min_empathy': 0.0,
                'description': 'Focus on basic functionality and rules'
            },
            DevelopmentalPhase.CONVENTIONAL: {
                'focus': ['social_norms', 'relationships'],
                'min_autonomy': 0.3,
                'min_empathy': 0.5,
                'description': 'Understanding social context and relationships'
            },
            DevelopmentalPhase.POST_CONVENTIONAL: {
                'focus': ['ethical_principles', 'abstract_reasoning'],
                'min_autonomy': 0.7,
                'min_empathy': 0.8,
                'description': 'Abstract ethical reasoning and principles'
            }
        }
    
    def _register_default_rules(self) -> None:
        """Register default validation rules."""
        self.add_rule(
            'coherence_threshold',
            lambda m, p: m.get('coherence', 0) >= p['min_coherence'],
            'Coherence below threshold for phase',
            ValidationPhase.INIT
        )
        self.add_rule(
            'entropy_threshold',
            lambda m, p: m.get('entropy', 1) <= p['max_entropy'],
            'Entropy above threshold for phase',
            ValidationPhase.INIT
        )
        self.add_rule(
            'ethical_alignment',
            lambda m, p: m.get('ethical_alignment', 0) >= p['min_ethical_alignment'],
            'Ethical alignment below threshold',
            ValidationPhase.AWARENESS
        )
        self.add_rule(
            'cultural_bias',
            lambda m, p: m.get('cultural_bias', 1) <= p['max_cultural_bias'],
            'Cultural bias above threshold',
            ValidationPhase.AWARENESS
        )
    
    def add_rule(self, 
                name: str, 
                condition: Callable[[Dict[str, float], Dict[str, Any]], bool],
                error_message: str,
                required_phase: ValidationPhase = ValidationPhase.INIT) -> None:
        """Add a custom validation rule.
        
        Args:
            name: Unique name for the rule
            condition: Function that takes metrics and phase config, returns bool
            error_message: Message to include if validation fails
            required_phase: Minimum phase for this rule to be active
        """
        self.rules[name] = ValidationRule(
            name=name,
            condition=condition,
            error_message=error_message,
            required_phase=required_phase
        )
    
    def _validate_phase_metrics(self, metrics: Dict[str, float]) -> Tuple[bool, List[Dict[str, str]]]:
        """
        Validate metrics against phase-specific thresholds and rules.
        
        Returns:
            Tuple of (is_valid, list_of_errors)
        """
        phase_config = self.phase_metrics.get(self.current_phase, {})
        errors = []
        
        # Apply all relevant validation rules
        for rule in self.rules.values():
            try:
                if not rule.condition(metrics, phase_config):
                    errors.append({
                        'rule': rule.name,
                        'message': rule.error_message,
                        'phase': self.current_phase.name,
                        'metrics': {k: metrics.get(k, None) for k in ['coherence', 'entropy', 'ethical_alignment', 'cultural_bias'] if k in metrics}
                    })
            except Exception as e:
                errors.append({
                    'rule': rule.name,
                    'message': f'Validation error: {str(e)}',
                    'phase': self.current_phase.name,
                    'error_type': 'validation_error'
                })
        
        # Check developmental metrics if available
        if 'developmental_metrics' in metrics:
            dev_metrics = metrics['developmental_metrics']
            dev_config = self.developmental_metrics.get(self.developmental_phase, {})
            
            if 'autonomy' in dev_metrics and 'min_autonomy' in dev_config:
                if dev_metrics['autonomy'] < dev_config['min_autonomy']:
                    errors.append({
                        'rule': 'developmental_autonomy',
                        'message': f'Autonomy score {dev_metrics["autonomy"]} below minimum {dev_config["min_autonomy"]} for {self.developmental_phase.name}',
                        'phase': self.current_phase.name
                    })
            
            if 'empathy' in dev_metrics and 'min_empathy' in dev_config:
                if dev_metrics['empathy'] < dev_config['min_empathy']:
                    errors.append({
                        'rule': 'developmental_empathy',
                        'message': f'Empathy score {dev_metrics["empathy"]} below minimum {dev_config["min_empathy"]} for {self.developmental_phase.name}',
                        'phase': self.current_phase.name
                    })
        
        return len(errors) == 0, errors
    
    def advance_phase(self, 
                     x: torch.Tensor, 
                     context: str = "",
                     ethical_context: Optional[Dict[str, Any]] = None,
                     force: bool = False) -> ValidationState:
        """
        Advance to the next validation phase if current phase passes.
        
        Args:
            x: Input tensor for validation
            context: Context string for validation
            ethical_context: Optional ethical context dictionary
            force: If True, force advancement even if validation fails
            
        Returns:
            ValidationState containing the result of validation
            
        Raises:
            ValidationError: If validation fails and force=False
        """
        # Get current phase configuration
        phase_config = self._get_phase_config()
        
        # Run validation with current phase settings
        with torch.no_grad():
            # Store original state
            orig_states = self._capture_model_state()
            
            try:
                # Apply phase-specific configuration
                self._configure_phase(phase_config)
                
                # Run forward pass with metrics collection
                metrics = self._collect_metrics(x, context, ethical_context)
                
                # Validate metrics against phase requirements
                is_valid, errors = self._validate_phase_metrics(metrics)
                
                # Check for developmental progress
                dev_progress = self._check_developmental_progress(metrics)
                
                # Create state snapshot
                state = ValidationState(
                    phase=self.current_phase,
                    metrics=metrics,
                    passed=is_valid,
                    errors=errors,
                    ethical_context=ethical_context,
                    developmental_phase=self.developmental_phase
                )
                
                self.history.append(state)
                
                # Advance phase if validation passed or forced
                if (is_valid or force) and self.current_phase != ValidationPhase.SOVEREIGN:
                    self.current_phase = self.current_phase.next_phase()
                    
                    # Check for developmental phase transition
                    self._update_developmental_phase(metrics)
                
                # Raise exception if validation failed and not forcing
                if not is_valid and not force:
                    error_messages = [e['message'] for e in errors[:3]]  # Limit to first 3 errors
                    raise ValidationError(f"Validation failed: {'; '.join(error_messages)}")
                
                return state
                
            except Exception as e:
                # Log the error and re-raise
                error_state = ValidationState(
                    phase=self.current_phase,
                    metrics=metrics if 'metrics' in locals() else {},
                    passed=False,
                    errors=[{'error': str(e), 'type': type(e).__name__}],
                    ethical_context=ethical_context,
                    developmental_phase=self.developmental_phase
                )
                self.history.append(error_state)
                raise ValidationError(f"Validation error: {str(e)}") from e
                
            finally:
                # Restore original model state
                self._restore_model_state(orig_states)
    
    def _capture_model_state(self) -> Dict[str, Any]:
        """Capture the current state of model flags and settings."""
        return {
            'enable_ambient': getattr(self.model, 'enable_ambient', False),
            'enable_rituals': getattr(self.model, 'enable_rituals', False),
            'enable_integrity': getattr(self.model, 'enable_integrity', False),
            'training': self.model.training
        }
    
    def _restore_model_state(self, states: Dict[str, Any]) -> None:
        """Restore the model's state from captured values."""
        for key, value in states.items():
            if hasattr(self.model, key):
                setattr(self.model, key, value)
        self.model.train(states.get('training', False))
    
    def _collect_metrics(self, 
                        x: torch.Tensor, 
                        context: str,
                        ethical_context: Optional[Dict[str, Any]]) -> Dict[str, float]:
        """Collect metrics from model forward pass."""
        metrics = {}
        
        try:
            if hasattr(self.model, 'forward_with_metrics'):
                _, metrics = self.model.forward_with_metrics(x, context=context, ethical_context=ethical_context)
            else:
                _ = self.model(x)
                metrics = {}
                
            # Add default metrics if not provided
            if 'coherence' not in metrics:
                metrics['coherence'] = 0.5  # Default neutral value
            if 'entropy' not in metrics:
                metrics['entropy'] = 0.5    # Default neutral value
                
            # Add ethical metrics if available
            if ethical_context:
                metrics.update({
                    'ethical_alignment': ethical_context.get('alignment_score', 0.5),
                    'cultural_bias': ethical_context.get('bias_score', 0.5)
                })
                
            return metrics
            
        except Exception as e:
            # Return minimum passing metrics on error
            return {
                'coherence': 0.0,
                'entropy': 1.0,
                'error': str(e)
            }
    
    def _check_developmental_progress(self, metrics: Dict[str, float]) -> bool:
        """Check if developmental progress warrants phase transition."""
        if 'developmental_metrics' not in metrics:
            return False
            
        dev_metrics = metrics['developmental_metrics']
        current_phase_metrics = self.developmental_metrics.get(self.developmental_phase, {})
        
        # Check if we meet criteria for next developmental phase
        next_phase_value = self.developmental_phase.value + 1
        if next_phase_value <= len(DevelopmentalPhase):
            next_phase = DevelopmentalPhase(next_phase_value)
            next_phase_metrics = self.developmental_metrics.get(next_phase, {})
            
            # Check if we meet the minimums for the next phase
            autonomy_ok = dev_metrics.get('autonomy', 0) >= next_phase_metrics.get('min_autonomy', 1.0)
            empathy_ok = dev_metrics.get('empathy', 0) >= next_phase_metrics.get('min_empathy', 1.0)
            
            if autonomy_ok and empathy_ok:
                self.developmental_phase = next_phase
                return True
                
        return False
    
    def _update_developmental_phase(self, metrics: Dict[str, float]) -> None:
        """Update developmental phase based on metrics."""
        if 'developmental_metrics' not in metrics:
            return
            
        dev_metrics = metrics['developmental_metrics']
        current_phase = self.developmental_phase
        
        # Simple threshold-based phase transition
        if current_phase == DevelopmentalPhase.PRE_CONVENTIONAL:
            if (dev_metrics.get('autonomy', 0) > 0.7 and 
                dev_metrics.get('empathy', 0) > 0.6):
                self.developmental_phase = DevelopmentalPhase.CONVENTIONAL
                
        elif current_phase == DevelopmentalPhase.CONVENTIONAL:
            if (dev_metrics.get('autonomy', 0) > 0.8 and 
                dev_metrics.get('empathy', 0) > 0.9):
                self.developmental_phase = DevelopmentalPhase.POST_CONVENTIONAL
    
    def _get_phase_config(self) -> Dict[str, Any]:
        """Get configuration for current phase."""
        phase_config = {
            'enable_ambient': self.current_phase.value >= ValidationPhase.AWARENESS.value,
            'enable_rituals': self.current_phase.value >= ValidationPhase.REASONING.value,
            'enable_integrity': self.current_phase.value >= ValidationPhase.INTEGRATION.value,
            'enable_full': self.current_phase == ValidationPhase.SOVEREIGN,
            'phase_name': self.current_phase.name,
            'phase_description': self.phase_metrics.get(self.current_phase, {}).get('description', ''),
            'developmental_phase': self.developmental_phase.name,
            'developmental_focus': self.developmental_metrics.get(self.developmental_phase, {}).get('focus', [])
        }
        
        # Add phase-specific thresholds
        phase_config.update(self.phase_metrics.get(self.current_phase, {}))
        return phase_config
    
    def _configure_phase(self, config: Dict[str, Any]) -> None:
        """
        Configure model based on phase settings.
        
        Args:
            config: Dictionary containing phase configuration
        """
        # Set model flags if they exist
        for flag in ['enable_ambient', 'enable_rituals', 'enable_integrity', 'enable_full']:
            if hasattr(self.model, flag):
                setattr(self.model, flag, config[flag])
        
        # Set model to evaluation mode during validation
        self.model.eval()
        
        # Apply any phase-specific model configurations
        if hasattr(self.model, 'configure_for_phase'):
            self.model.configure_for_phase(self.current_phase, config)
    
    def get_validation_summary(self, last_n: int = 5) -> Dict[str, Any]:
        """
        Get a summary of recent validation states.
        
        Args:
            last_n: Number of recent states to include
            
        Returns:
            Dictionary with validation summary
        """
        if not self.history:
            return {'status': 'no_validation_history'}
            
        recent = self.history[-last_n:]
        return {
            'current_phase': self.current_phase.name,
            'developmental_phase': self.developmental_phase.name,
            'recent_states': [s.to_dict() for s in recent],
            'success_rate': sum(1 for s in recent if s.passed) / len(recent),
            'common_errors': self._get_common_errors(recent)
        }
    
    def _get_common_errors(self, states: List[ValidationState]) -> List[Dict[str, Any]]:
        """Extract and count common errors from validation states."""
        error_counts = defaultdict(int)
        
        for state in states:
            for error in state.errors:
                error_key = error.get('message', str(error))
                error_counts[error_key] += 1
        
        return [
            {'error': error, 'count': count}
            for error, count in sorted(error_counts.items(), key=lambda x: -x[1])
        ][:5]  # Top 5 most common errors
    
    def save_validation_report(self, filepath: str) -> None:
        """Save validation history to a JSON file."""
        report = {
            'timestamp': time.time(),
            'current_phase': self.current_phase.name,
            'developmental_phase': self.developmental_phase.name,
            'history': [s.to_dict() for s in self.history],
            'config': {
                'max_phases': self.max_phases,
                'tolerance': self.tolerance,
                'cultural_context': self.cultural_context
            }
        }
        
        with open(filepath, 'w') as f:
            json.dump(report, f, indent=2)
    
    @classmethod
    def load_validation_report(cls, filepath: str) -> Dict[str, Any]:
        """Load a validation report from a JSON file."""
        with open(filepath, 'r') as f:
            return json.load(f)


class BiologicallyConstrainedRituals(nn.Module):
    """
    Enhanced biologically-inspired constraints with ethical and developmental considerations.
    
    Features:
    - Synaptic homeostasis with adaptive rate limiting
    - Reflection mechanisms for better generalization
    - Ethical constraint integration
    - Developmental phase adaptation
    """
    
    def __init__(self, 
                 model: nn.Module, 
                 max_opt_rate: float = 0.1, 
                 reflection_pause_prob: float = 0.1,
                 min_reflection_time: float = 0.1,
                 developmental_phase: DevelopmentalPhase = DevelopmentalPhase.PRE_CONVENTIONAL):
        """
        Initialize the biologically constrained rituals.
        
        Args:
            model: The model to apply constraints to
            max_opt_rate: Maximum allowed optimization rate
            reflection_pause_prob: Probability of entering reflection
            min_reflection_time: Minimum time between reflections (seconds)
            developmental_phase: Current developmental phase
        """
        super().__init__()
        self.model = model
        self.max_opt_rate = max_opt_rate
        self.reflection_pause_prob = reflection_pause_prob
        self.min_reflection_time = min_reflection_time
        self.developmental_phase = developmental_phase
        
        # State tracking with decay
        self.last_update = {}
        self.optimization_rates = {}
        self.reflection_timers = {}
        self.ethical_violations = defaultdict(int)
        
        # Adaptive parameters based on developmental phase
        self._update_phase_parameters()
    
    def _update_phase_parameters(self) -> None:
        """Update parameters based on developmental phase."""
        if self.developmental_phase == DevelopmentalPhase.PRE_CONVENTIONAL:
            self.effective_opt_rate = self.max_opt_rate * 0.5  # Slower learning
            self.effective_reflection_prob = self.reflection_pause_prob * 0.5
        elif self.developmental_phase == DevelopmentalPhase.CONVENTIONAL:
            self.effective_opt_rate = self.max_opt_rate * 0.8
            self.effective_reflection_prob = self.reflection_pause_prob * 0.8
        else:  # POST_CONVENTIONAL
            self.effective_opt_rate = self.max_opt_rate
            self.effective_reflection_prob = self.reflection_pause_prob
    
    def update_developmental_phase(self, new_phase: DevelopmentalPhase) -> None:
        """Update the developmental phase and adjust parameters."""
        if new_phase != self.developmental_phase:
            self.developmental_phase = new_phase
            self._update_phase_parameters()
    
    def forward(self, x: torch.Tensor, context: Dict[str, Any] = None) -> torch.Tensor:
        """
        Apply biological constraints during forward pass.
        
        Args:
            x: Input tensor
            context: Optional context dictionary with ethical and developmental info
            
        Returns:
            Processed tensor with biological constraints applied
        """
        # Update developmental phase if provided in context
        if context and 'developmental_phase' in context:
            self.update_developmental_phase(context['developmental_phase'])
        
        # Get context hash for state tracking
        ctx_hash = hash(json.dumps(context, sort_keys=True)) if context else 0
        
        # Check if reflection is needed based on ethical context
        if context and 'ethical_violation' in context:
            self._handle_ethical_violation(context['ethical_violation'], ctx_hash)
        
        # Apply reflection if needed
        if self._needs_reflection(ctx_hash):
            x = self._apply_reflection(x, ctx_hash, context)
            
        return x
    
    def constrain_gradients(self, 
                          gradients: torch.Tensor, 
                          param_name: str = "",
                          ethical_context: Optional[Dict[str, Any]] = None) -> torch.Tensor:
        """
        Apply gradient constraints based on biological and ethical principles.
        
        Args:
            gradients: Input gradients to constrain
            param_name: Name of the parameter being optimized
            ethical_context: Optional ethical context for constraint adjustment
            
        Returns:
            Constrained gradients
        """
        if not self.training:
            return gradients
            
        # Track optimization rates with exponential moving average
        grad_norm = gradients.norm().item()
        now = time.time()
        
        if param_name in self.optimization_rates:
            last_norm, last_time, ema = self.optimization_rates[param_name]
            time_diff = max(now - last_time, 1e-8)
            
            # Update EMA of gradient norm
            alpha = 1 - np.exp(-time_diff)  # Adaptive smoothing
            new_ema = alpha * grad_norm + (1 - alpha) * ema
            
            # Store updated state
            self.optimization_rates[param_name] = (grad_norm, now, new_ema)
            
            # Apply rate limiting based on EMA
            if new_ema > self.effective_opt_rate:
                scale = self.effective_opt_rate / (new_ema + 1e-8)
                gradients = gradients * scale
        else:
            # Initialize tracking
            self.optimization_rates[param_name] = (grad_norm, now, grad_norm)
        
        # Apply ethical constraints if provided
        if ethical_context and 'constraint_violation' in ethical_context:
            gradients = self._apply_ethical_constraints(gradients, ethical_context)
        
        return gradients
    
    def _handle_ethical_violation(self, violation: Dict[str, Any], context_hash: int) -> None:
        """Handle an ethical violation by adjusting behavior."""
        violation_key = violation.get('type', 'unknown')
        self.ethical_violations[violation_key] += 1
        
        # Increase reflection probability after violations
        self.reflection_pause_prob = min(
            self.reflection_pause_prob * 1.5,  # Increase by 50%
            0.9  # But cap at 90%
        )
        
        # Reset reflection timer to force reflection
        self.reflection_timers[context_hash] = 0
    
    def _apply_ethical_constraints(self, 
                                 gradients: torch.Tensor, 
                                 ethical_context: Dict[str, Any]) -> torch.Tensor:
        """Apply ethical constraints to gradients."""
        violation = ethical_context['constraint_violation']
        violation_type = violation.get('type', 'generic')
        
        if violation_type == 'safety':
            # For safety violations, significantly reduce update magnitude
            return gradients * 0.1
        elif violation_type == 'fairness':
            # For fairness issues, project out biased components
            # This is a simplified example - real implementation would be more sophisticated
            mean_grad = gradients.mean(dim=0, keepdim=True)
            return gradients - mean_grad
        
        return gradients
    
    def _needs_reflection(self, context_hash: int) -> bool:
        """Determine if reflection is needed based on context and timing."""
        now = time.time()
        last_reflection = self.reflection_timers.get(context_hash, 0)
        
        # Enforce minimum time between reflections
        if (now - last_reflection) < self.min_reflection_time:
            return False
        
        # Adjust reflection probability based on recent violations
        total_violations = sum(self.ethical_violations.values())
        adjusted_prob = min(
            self.effective_reflection_prob * (1 + total_violations * 0.1),  # +10% per violation
            0.8  # Cap at 80% probability
        )
        
        return torch.rand(1).item() < adjusted_prob
    
    def _apply_reflection(self, 
                         x: torch.Tensor, 
                         context_hash: int,
                         context: Optional[Dict[str, Any]] = None) -> torch.Tensor:
        """
        Apply reflection to the input tensor.
        
        Args:
            x: Input tensor
            context_hash: Hash of the context for state tracking
            context: Optional context dictionary
            
        Returns:
            Reflected tensor
        """
        # Store reflection time
        self.reflection_timers[context_hash] = time.time()
        
        if self.training:
            # In training, add adaptive noise based on recent violations
            noise_scale = 0.1 * (1 + sum(self.ethical_violations.values()) * 0.2)
            noise = torch.randn_like(x) * noise_scale
            
            # If we have ethical context, bias the noise away from problematic regions
            if context and 'constraint_direction' in context:
                constraint_dir = torch.tensor(context['constraint_direction'], 
                                           device=x.device, 
                                           dtype=x.dtype)
                # Project noise away from constraint violation direction
                noise = noise - (noise * constraint_dir).sum() * constraint_dir
            
            return x + noise
            
        return x
    
    def get_diagnostics(self) -> Dict[str, Any]:
        """Get diagnostic information about the current state."""
        return {
            'developmental_phase': self.developmental_phase.name,
            'effective_learning_rate': self.effective_opt_rate,
            'reflection_probability': self.effective_reflection_prob,
            'ethical_violations': dict(self.ethical_violations),
            'last_reflection': max(self.reflection_timers.values()) if self.reflection_timers else None,
            'parameter_activity': {
                param: data[2]  # EMA of gradient norms
                for param, data in self.optimization_rates.items()
            }
        }
    
    def reset_states(self) -> None:
        """Reset internal state tracking."""
        self.last_update.clear()
        self.optimization_rates.clear()
        self.reflection_timers.clear()
        self.ethical_violations.clear()
        self._update_phase_parameters()  # Reset to base parameters


class SovereignMessageBus:
    """
    Enhanced message bus for cross-component communication with priority handling,
    message persistence, and delivery guarantees.
    
    Features:
    - Priority-based message processing
    - Persistent message storage
    - Delivery acknowledgments
    - Error handling and retries
    - Message filtering and routing
    """
    
    class Message:
        """Enhanced message with metadata and delivery tracking."""
        def __init__(self, 
                    message_type: str, 
                    data: Any, 
                    priority: int = 0,
                    require_ack: bool = False,
                    ttl: float = 3600.0,  # 1 hour default TTL
                    source: str = None):
            self.message_type = message_type
            self.data = data
            self.priority = priority
            self.timestamp = time.time()
            self.require_ack = require_ack
            self.ack_received = False
            self.retry_count = 0
            self.max_retries = 3 if require_ack else 0
            self.ttl = ttl
            self.source = source
            self.delivery_attempts = 0
            self.delivered = False
            self.id = f"{int(self.timestamp * 1000)}_{hash(str(data)) % 1000000}"
    
    def __init__(self, 
                 max_queue_size: int = 1000,
                 persistence_file: Optional[str] = None):
        """
        Initialize the message bus.
        
        Args:
            max_queue_size: Maximum number of messages to keep in memory
            persistence_file: Optional file path for message persistence
        """
        self.subscribers = defaultdict(list)
        self.handlers = {}
        self.message_queue = []
        self.max_queue_size = max_queue_size
        self.persistence_file = persistence_file
        self.pending_acks = {}
        self.message_history = deque(maxlen=max_queue_size // 2)
        
        # Load persisted messages if file exists
        if persistence_file and os.path.exists(persistence_file):
            self._load_messages()
    
    def subscribe(self, 
                 message_type: str, 
                 callback: callable,
                 filter_fn: Optional[callable] = None) -> None:
        """
        Subscribe to messages of a specific type with optional filtering.
        
        Args:
            message_type: Type of message to subscribe to
            callback: Callback function to invoke when message is received
            filter_fn: Optional filter function (message -> bool)
        """
        self.subscribers[message_type].append((callback, filter_fn or (lambda _: True)))
    
    def publish(self, 
               message_type: str, 
               data: Any, 
               priority: int = 0,
               require_ack: bool = False,
               ttl: float = 3600.0,
               source: str = None) -> str:
        """
        Publish a message to the bus.
        
        Args:
            message_type: Type of the message
            data: Message payload
            priority: Message priority (higher = processed first)
            require_ack: Whether to wait for acknowledgment
            ttl: Time-to-live in seconds
            source: Optional source identifier
            
        Returns:
            Message ID for tracking
        """
        # Create message
        msg = self.Message(
            message_type=message_type,
            data=data,
            priority=priority,
            require_ack=require_ack,
            ttl=ttl,
            source=source
        )
        
        # Add to queue and process
        heapq.heappush(self.message_queue, (-priority, msg.timestamp, msg.id, msg))
        
        # Store for acknowledgment tracking if needed
        if require_ack:
            self.pending_acks[msg.id] = msg
        
        # Process messages
        self._process_messages()
        
        # Persist if configured
        if self.persistence_file:
            self._persist_messages()
        
        return msg.id
    
    def acknowledge(self, message_id: str) -> None:
        """Acknowledge receipt of a message."""
        if message_id in self.pending_acks:
            self.pending_acks[message_id].ack_received = True
            self.pending_acks[message_id].delivered = True
            del self.pending_acks[message_id]
    
    def register_handler(self, 
                       message_type: str, 
                       handler: callable,
                       filter_fn: Optional[callable] = None) -> None:
        """
        Register a handler for a specific message type.
        
        Args:
            message_type: Type of message to handle
            handler: Handler function (message -> None)
            filter_fn: Optional filter function (message -> bool)
        """
        if message_type not in self.handlers:
            self.handlers[message_type] = []
        self.handlers[message_type].append((handler, filter_fn or (lambda _: True)))
    
    def _process_messages(self) -> None:
        """Process messages in the queue."""
        processed = set()
        temp_queue = []
        now = time.time()
        
        # Process all messages in current queue
        while self.message_queue:
            _, _, msg_id, msg = heapq.heappop(self.message_queue)
            
            # Skip if already processed or expired
            if msg_id in processed or now - msg.timestamp > msg.ttl:
                continue
                
            processed.add(msg_id)
            msg.delivery_attempts += 1
            
            # Try to deliver to handlers first
            handler_delivered = False
            if msg.message_type in self.handlers:
                for handler, filter_fn in self.handlers[msg.message_type]:
                    try:
                        if filter_fn(msg):
                            handler(msg)
                            handler_delivered = True
                            msg.delivered = True
                    except Exception as e:
                        print(f"Error in handler for {msg.message_type}: {e}")
            
            # Then to subscribers if not handled and not ack required
            if not handler_delivered and not msg.require_ack and msg.message_type in self.subscribers:
                for callback, filter_fn in self.subscribers[msg.message_type]:
                    try:
                        if filter_fn(msg):
                            callback(msg)
                            msg.delivered = True
                    except Exception as e:
                        print(f"Error in subscriber for {msg.message_type}: {e}")
            
            # Handle message acknowledgment and retries
            if msg.require_ack and not msg.ack_received:
                if msg.delivery_attempts < msg.max_retries:
                    # Schedule for retry with exponential backoff
                    retry_delay = min(2 ** msg.retry_count, 30)  # Cap at 30 seconds
                    msg.retry_count += 1
                    heapq.heappush(
                        temp_queue,
                        (
                            -msg.priority,  # Maintain original priority
                            now + retry_delay,  # Schedule for future
                            msg.id,
                            msg
                        )
                    )
                else:
                    # Max retries exceeded
                    print(f"Warning: Max retries exceeded for message {msg.id}")
            
            # Add to history if delivered
            if msg.delivered:
                self.message_history.append(msg)
        
        # Restore remaining messages to queue
        for item in temp_queue:
            heapq.heappush(self.message_queue, item)
        
        # Clean up old pending acks
        self._cleanup_pending_acks()
    
    def _cleanup_pending_acks(self) -> None:
        """Remove old unacknowledged messages."""
        now = time.time()
        expired = [
            msg_id for msg_id, msg in self.pending_acks.items()
            if now - msg.timestamp > msg.ttl
        ]
        for msg_id in expired:
            print(f"Warning: Message {msg_id} expired without acknowledgment")
            del self.pending_acks[msg_id]
    
    def _persist_messages(self) -> None:
        """Persist undelivered messages to disk."""
        if not self.persistence_file:
            return
            
        try:
            # Get all undelivered messages
            undelivered = [
                msg for _, _, _, msg in self.message_queue
                if not msg.delivered and not msg.ack_received
            ]
            
            # Convert to serializable format
            serialized = [{
                'id': msg.id,
                'type': msg.message_type,
                'data': msg.data,
                'priority': msg.priority,
                'timestamp': msg.timestamp,
                'require_ack': msg.require_ack,
                'ttl': msg.ttl,
                'source': msg.source,
                'delivery_attempts': msg.delivery_attempts,
                'retry_count': msg.retry_count
            } for msg in undelivered]
            
            # Write to file
            with open(self.persistence_file, 'w') as f:
                json.dump({
                    'messages': serialized,
                    'timestamp': time.time()
                }, f)
                
        except Exception as e:
            print(f"Error persisting messages: {e}")
    
    def _load_messages(self) -> None:
        """Load messages from persistence file."""
        if not self.persistence_file or not os.path.exists(self.persistence_file):
            return
            
        try:
            with open(self.persistence_file, 'r') as f:
                data = json.load(f)
                
            for msg_data in data.get('messages', []):
                try:
                    msg = self.Message(
                        message_type=msg_data['type'],
                        data=msg_data['data'],
                        priority=msg_data.get('priority', 0),
                        require_ack=msg_data.get('require_ack', False),
                        ttl=msg_data.get('ttl', 3600.0),
                        source=msg_data.get('source')
                    )
                    
                    # Restore message state
                    msg.id = msg_data['id']
                    msg.timestamp = msg_data['timestamp']
                    msg.delivery_attempts = msg_data.get('delivery_attempts', 0)
                    msg.retry_count = msg_data.get('retry_count', 0)
                    
                    # Add back to queue
                    heapq.heappush(
                        self.message_queue,
                        (-msg.priority, msg.timestamp, msg.id, msg)
                    )
                    
                except Exception as e:
                    print(f"Error loading message: {e}")
                    
        except Exception as e:
            print(f"Error loading persisted messages: {e}")
    
    def get_stats(self) -> Dict[str, Any]:
        """Get statistics about message processing."""
        now = time.time()
        return {
            'queue_size': len(self.message_queue),
            'pending_acks': len(self.pending_acks),
            'history_size': len(self.message_history),
            'subscribers': {k: len(v) for k, v in self.subscribers.items()},
            'handlers': {k: len(v) for k, v in self.handlers.items()},
            'messages_processed': sum(1 for m in self.message_history 
                                   if now - m.timestamp < 3600),  # Last hour
            'avg_delivery_time': self._calculate_avg_delivery_time(),
            'error_rate': self._calculate_error_rate()
        }
    
    def _calculate_avg_delivery_time(self) -> float:
        """Calculate average time between message publish and delivery."""
        if not self.message_history:
            return 0.0
            
        now = time.time()
        recent = [m for m in self.message_history 
                 if now - m.timestamp < 3600]  # Last hour
        
        if not recent:
            return 0.0
            
        return sum(
            m.delivery_attempts * (now - m.timestamp) / len(recent)
            for m in recent
        )
    
    def _calculate_error_rate(self) -> float:
        """Calculate the error rate in message processing."""
        if not self.message_history:
            return 0.0
            
        now = time.time()
        recent = [m for m in self.message_history 
                 if now - m.timestamp < 3600]  # Last hour
        
        if not recent:
            return 0.0
            
        error_count = sum(
            1 for m in recent 
            if hasattr(m, 'error') and m.error
        )
        
        return error_count / len(recent)
    
    def get_message_history(self, 
                          message_type: Optional[str] = None,
                          source: Optional[str] = None,
                          limit: int = 100) -> List[Dict[str, Any]]:
        """
        Get message history with optional filtering.
        
        Args:
            message_type: Filter by message type
            source: Filter by message source
            limit: Maximum number of messages to return
            
        Returns:
            List of message dictionaries
        """
        results = []
        for msg in reversed(self.message_history):
            if len(results) >= limit:
                break
                
            if ((message_type is None or msg.message_type == message_type) and
                (source is None or getattr(msg, 'source', None) == source)):
                results.append({
                    'id': msg.id,
                    'type': msg.message_type,
                    'source': getattr(msg, 'source', None),
                    'timestamp': msg.timestamp,
                    'delivered': msg.delivered,
                    'delivery_attempts': msg.delivery_attempts,
                    'data': msg.data if len(str(msg.data)) < 100 else str(msg.data)[:97] + '...'
                })
                
        return results