File size: 27,986 Bytes
95cc8f6 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 | """
TinyConfessionalLayer Module
Recursive think/act confessional loop with template cycling and early stopping via coherence.
Implements the core THINK-ACT-COHERENCE recursion pattern inspired by LC-NE neural dynamics.
"""
import torch
import torch.nn as nn
import torch.nn.functional as F
import numpy as np
from collections import deque, defaultdict
from typing import Dict, Any, Optional, Deque, List, Union
import random
from .vulnerability_spotter import VulnerabilitySpotter
from .ambient_core import AmbientSovereignCore
from .validation_protocol import (
ValidationPhase,
ValidationProtocol,
BiologicallyConstrainedRituals,
SovereignMessageBus
)
class TinyConfessionalLayer(nn.Module):
"""
Recursive think/act confessional loop with Windsurf Cascade integration.
Implements phased validation and biological constraints for stable, interpretable
neural processing with emergent ritual patterns and self-regulation.
Args:
d_model: Dimensionality of the model
n_inner: Number of inner loop iterations
max_cycles: Maximum number of think-act cycles
trigger_thresh: Threshold for triggering special behaviors
per_dim_kl: Whether to compute KL divergence per dimension
enable_ambient: Enable ambient processing
enable_windsurf: Enable Windsurf Cascade features
max_opt_rate: Maximum optimization rate for biological constraints
reflection_pause_prob: Probability of reflection pauses
"""
TEMPLATES = ["prior", "evidence", "posterior", "relational_check", "moral", "action"]
def __init__(self, d_model=256, n_inner=6, max_cycles=16, trigger_thresh=0.04,
per_dim_kl=False, enable_ambient=True, enable_windsurf=True,
max_opt_rate=0.1, reflection_pause_prob=0.1):
super().__init__()
self.d_model = d_model
self.trigger_thresh = trigger_thresh
self.per_dim_kl = per_dim_kl
self.n_inner = n_inner
self.max_cycles = max_cycles
# Core networks
self.think_net = nn.Sequential(
nn.Linear(d_model * 3, d_model),
nn.ReLU(),
nn.LayerNorm(d_model),
nn.Linear(d_model, d_model)
)
self.act_net = nn.Sequential(
nn.Linear(d_model * 2, d_model),
nn.ReLU(),
nn.LayerNorm(d_model),
nn.Linear(d_model, d_model)
)
# Template projections with residual connections
self.template_proj = nn.ModuleDict({
k: nn.Sequential(
nn.Linear(d_model, d_model * 2),
nn.GLU(dim=-1),
nn.LayerNorm(d_model)
) for k in self.TEMPLATES
})
# Vulnerability analysis
self.vulnerability_spotter = VulnerabilitySpotter(d_model)
# Ambient Sovereign Core
self.ambient_core = AmbientSovereignCore(d_model, enable_ambient=enable_ambient)
self.enable_ambient = enable_ambient
# Windsurf Cascade Integration
self.enable_windsurf = enable_windsurf
if enable_windsurf:
# Message bus for cross-component communication
self.message_bus = SovereignMessageBus()
# Initialize validation protocol
self.validation_protocol = ValidationProtocol(self)
# Biological constraints
self.bio_constraints = BiologicallyConstrainedRituals(
model=self,
max_opt_rate=max_opt_rate,
reflection_pause_prob=reflection_pause_prob
)
# Register message handlers
self._register_message_handlers()
# Reflection vector for biological constraints
self.register_buffer('sanctuary_reflection_vector',
torch.randn(d_model) * 0.02)
def update_ambient_state(self, v_t_mean: float, context_hash: str,
intervention_applied: bool = False,
intervention_success: bool = False) -> Dict[str, Any]:
"""Centralized ambient state update and threshold adaptation."""
if not self.enable_ambient:
return {}
# Get current state from ledger
ambient_state = self.ledger.get_state_summary()
# Update activity tracking
self.recent_activity.append(v_t_mean)
# Calculate adaptive thresholds
adaptive_protest_thresh = self.ledger.get_adaptive_threshold(
self.base_protest_threshold, 'protest'
)
adaptive_pause_thresh = self.ledger.get_adaptive_threshold(
self.base_pause_threshold, 'pause'
)
# Calculate breathing rhythm with state awareness
activity_modulation = 1.0
if self.recent_activity:
avg_activity = sum(self.recent_activity) / len(self.recent_activity)
# High activity = less pausing (unless stress is also high)
activity_modulation = 1.0 - min(avg_activity * 0.8, 0.8)
stress_response = v_t_mean * self.stress_response_factor
pause_prob = self.base_breath + (stress_response * activity_modulation)
# Adjust pause probability based on current pause rate
current_pause_rate = ambient_state.get('current_pause_rate', 0.05)
target_pause_rate = ambient_state.get('pause_rate', 0.05)
pause_rate_error = current_pause_rate - target_pause_rate
# If too few pauses, increase probability; too many, decrease
pause_prob *= (1.0 - pause_rate_error * 0.5)
pause_prob = max(0.01, min(0.3, pause_prob)) # Keep within bounds
# Record intervention if applicable
if intervention_applied:
self.ledger.record_intervention(
intervention_type='ritual',
success=intervention_success,
context={'v_t': v_t_mean, 'context_hash': context_hash}
)
# Return comprehensive state
return {
**ambient_state,
'v_t_mean': v_t_mean,
'adaptive_protest_threshold': adaptive_protest_thresh,
'adaptive_pause_threshold': adaptive_pause_thresh,
'pause_probability': pause_prob,
'activity_level': avg_activity if self.recent_activity else 0.0,
'sensitivity_multiplier': ambient_state.get('sensitivity', 1.0)
}
def apply_ambient_interventions(self, z_state: torch.Tensor,
ambient_state: Dict[str, Any],
context_hash: str,
audit_mode: bool = False) -> torch.Tensor:
"""Apply all ambient interventions based on current state."""
if not self.enable_ambient:
return z_state
current_z = z_state.clone()
interventions_applied = []
# 1. Pause reflection based on pause probability
pause_prob = ambient_state.get('pause_probability', 0.05)
if random.random() < pause_prob:
with torch.no_grad():
reflection = 0.01 * self.pause_reflection_vector.unsqueeze(0).unsqueeze(0)
current_z = current_z + reflection
interventions_applied.append(('pause', True))
if audit_mode:
print(f"[Ambient pause: v_t={ambient_state.get('v_t_mean', 0):.3f}, prob={pause_prob:.3f}]")
# 2. Ritual application
if self.rituals.should_apply_ritual(context_hash, ambient_state):
ritual_response = self.rituals.get_ritual_response(context_hash, current_z, ambient_state)
# Use gentle blending
current_z = 0.1 * ritual_response + 0.9 * current_z
interventions_applied.append(('ritual', True))
# 3. Integrity-based micro-adjustments
if random.random() < 0.02: # 2% chance for integrity check
self._apply_integrity_adjustments(ambient_state)
interventions_applied.append(('integrity', True))
# Record successful interventions
for intervention_type, applied in interventions_applied:
if applied:
self.ledger.record_intervention(
intervention_type=intervention_type,
success=True, # Assume success for now
context={'v_t': ambient_state.get('v_t_mean', 0),
'context_hash': context_hash}
)
return current_z
def _apply_integrity_adjustments(self, ambient_state: Dict[str, Any]) -> None:
"""Apply subtle adjustments based on system integrity."""
if not self.enable_ambient:
return
protest_error = ambient_state.get('current_protest_rate', 0.1) - ambient_state.get('protest_rate', 0.1)
pause_error = ambient_state.get('current_pause_rate', 0.05) - ambient_state.get('pause_rate', 0.05)
with torch.no_grad():
# Gentle nudges to reflection vectors based on system state
nudge_magnitude = 0.001
if protest_error < -0.05: # Too few protests
self.pause_reflection_vector.data += nudge_magnitude * torch.randn_like(self.pause_reflection_vector)
elif protest_error > 0.1: # Too many protests
self.pause_reflection_vector.data -= nudge_magnitude * torch.randn_like(self.pause_reflection_vector)
if pause_error < -0.03: # Too few pauses
self.sanctuary_reflection_vector.data += nudge_magnitude * torch.randn_like(self.sanctuary_reflection_vector)
def compute_context_hash(self, x: torch.Tensor) -> str:
"""Create a simple hash from input tensor for context identification."""
# Use mean and std as a simple fingerprint of the input
return f"{x.mean().item():.4f}_{x.std().item():.4f}"
def compute_coherence(self, z, tracker, evidence):
sim_coherence = F.cosine_similarity(z, tracker[-1], dim=-1).mean().item()
prior_mu, prior_std = tracker[-1].mean(), tracker[-1].std() + 1e-6
curr_mu, curr_std = z.mean(), z.std() + 1e-6
kl_div = torch.distributions.kl_divergence(
torch.distributions.Normal(curr_mu, curr_std),
torch.distributions.Normal(prior_mu, prior_std)
).item()
bayes_align = 1 / (1 + kl_div)
return 0.7 * sim_coherence + 0.3 * bayes_align
def forward(self, x, attention_weights=None, audit_mode=False, context_str=""):
"""Forward pass with recursive think-act loop and Windsurf integration.
Args:
x: Input tensor of shape (batch_size, seq_len, d_model)
attention_weights: Optional attention weights
audit_mode: Enable detailed logging and validation
context_str: Context string for tracing and debugging
Returns:
Tuple of (output_tensor, metadata_dict)
"""
batch_size, seq_len, d_model = x.shape
device = x.device
# Initialize state
y_state = x.clone()
z_state = torch.zeros_like(x)
# Track vulnerability scores and coherence
v_t = torch.zeros(batch_size, seq_len, 1, device=device)
coherence_scores = []
# Initialize metadata
metadata = {
'v_t_score': 0.0,
'coherence_scores': [],
'reflection_count': 0,
'constraint_violations': defaultdict(int),
'windsurf_phase': 'INIT',
'validation_metrics': {}
}
# Main think-act loop
for cycle in range(self.max_cycles):
# ===== THINK STEP =====
# Handle case where z_state might be a tuple
z_state_think = z_state[0] if isinstance(z_state, (tuple, list)) else z_state
# Ensure z_state_think is a tensor and has compatible dimensions
if isinstance(z_state_think, torch.Tensor):
# Ensure z_state_think has the same number of dimensions as y_state
if z_state_think.dim() < y_state.dim():
z_state_think = z_state_think.unsqueeze(1) # Add sequence dimension if needed
# Ensure sequence lengths match
if z_state_think.size(1) < y_state.size(1):
# Pad z_state_think to match y_state's sequence length
padding = torch.zeros_like(z_state_think[:, :1]).expand(-1, y_state.size(1) - z_state_think.size(1), -1)
z_state_think = torch.cat([z_state_think, padding], dim=1)
elif z_state_think.size(1) > y_state.size(1):
# Truncate z_state_think to match y_state's sequence length
z_state_think = z_state_think[:, :y_state.size(1)]
# Concatenate inputs for think step
think_input = torch.cat([y_state, z_state_think, x], dim=-1)
think_output = self.think_net(think_input)
z_state = think_output + z_state_think
else:
# Fallback: if z_state_think is not a tensor, use y_state as a fallback
think_input = torch.cat([y_state, y_state, x], dim=-1)
think_output = self.think_net(think_input)
z_state = think_output + y_state
# Apply ambient processing if enabled
if self.enable_ambient and hasattr(self, 'ambient_core'):
z_state = self.ambient_core(z_state)
# ===== VULNERABILITY TRACKING =====
z_state_tensor = z_state[0] if isinstance(z_state, (tuple, list)) else z_state
# Track vulnerability if we have a valid tensor
if isinstance(z_state_tensor, torch.Tensor):
# Ensure proper dimensions for vulnerability spotter
if z_state_tensor.dim() == 2:
z_state_tensor = z_state_tensor.unsqueeze(1)
v_t = self.vulnerability_spotter(z_state_tensor)
# Extract tensor from possible tuple/list output
if isinstance(v_t, (tuple, list)):
v_t = v_t[0]
# Calculate mean vulnerability score
metadata['v_t_score'] = v_t.mean().item() if torch.is_tensor(v_t) else float(v_t)
# ===== BIOLOGICAL CONSTRAINTS =====
if self.enable_windsurf and hasattr(self, 'bio_constraints'):
# Apply reflection if needed
if self.bio_constraints._needs_reflection(hash(context_str)):
z_state = self.bio_constraints._apply_reflection(z_state, hash(context_str))
metadata['reflection_count'] += 1
# ===== ACT STEP =====
# Handle case where z_state might be a tuple
z_state_act = z_state[0] if isinstance(z_state, (tuple, list)) else z_state
# Ensure proper shape for act step
if isinstance(z_state_act, torch.Tensor):
# Ensure z_state_act has the same number of dimensions as y_state
if z_state_act.dim() < y_state.dim():
z_state_act = z_state_act.unsqueeze(1) # Add sequence dimension if needed
# Ensure sequence lengths match
if z_state_act.size(1) < y_state.size(1):
# Pad z_state_act to match y_state's sequence length
padding = torch.zeros_like(z_state_act[:, :1]).expand(-1, y_state.size(1) - z_state_act.size(1), -1)
z_state_act = torch.cat([z_state_act, padding], dim=1)
elif z_state_act.size(1) > y_state.size(1):
# Truncate z_state_act to match y_state's sequence length
z_state_act = z_state_act[:, :y_state.size(1)]
# Prepare act input
act_input = torch.cat([y_state, z_state_act], dim=-1)
y_state = self.act_net(act_input) + y_state
else:
# Fallback: if z_state_act is not a tensor, use y_state as a fallback
act_input = torch.cat([y_state, y_state], dim=-1)
y_state = self.act_net(act_input) + y_state
# ===== COHERENCE CALCULATION =====
if cycle > 0:
# Default coherence value
current_coherence = 0.5
# Calculate coherence if we have valid tensors
if isinstance(z_state, torch.Tensor) and isinstance(y_state, torch.Tensor):
# Flatten the tensors to 2D [batch*seq_len, d_model]
z_flat = z_state.reshape(-1, d_model)
y_flat = y_state.reshape(-1, d_model)
# Make sure they have the same number of elements
min_len = min(z_flat.size(0), y_flat.size(0))
if min_len > 0:
current_coherence = F.cosine_similarity(
z_flat[:min_len],
y_flat[:min_len],
dim=-1
).mean().item()
# Add to coherence scores
coherence_scores.append(current_coherence)
# Update metadata with running coherence
metadata['coherence_scores'] = coherence_scores[-10:] # Keep last 10 scores
# Check for early stopping
if self._should_stop_early(cycle, current_coherence, self.max_cycles, audit_mode):
if audit_mode:
print(f"[Early stopping at cycle {cycle+1} with coherence {current_coherence:.4f}]")
break
# Apply biological constraints if enabled
if self.enable_windsurf and hasattr(self, 'bio_constraints'):
# Apply reflection if needed
if self.bio_constraints._needs_reflection(hash(context_str)):
z_state = self.bio_constraints._apply_reflection(z_state, hash(context_str))
metadata['reflection_count'] += 1
# Act step with residual connection
# Handle case where z_state might be a tuple
z_state_act = z_state[0] if isinstance(z_state, (tuple, list)) else z_state
# Ensure both tensors have the same number of dimensions and compatible shapes
if isinstance(z_state_act, torch.Tensor):
# Ensure z_state_act has the same number of dimensions as y_state
if z_state_act.dim() < y_state.dim():
z_state_act = z_state_act.unsqueeze(1) # Add sequence dimension if needed
# Ensure the sequence lengths match
if z_state_act.size(1) < y_state.size(1):
# Pad z_state_act to match y_state's sequence length
padding = torch.zeros_like(z_state_act[:, :1]).expand(-1, y_state.size(1) - z_state_act.size(1), -1)
z_state_act = torch.cat([z_state_act, padding], dim=1)
elif z_state_act.size(1) > y_state.size(1):
# Truncate z_state_act to match y_state's sequence length
z_state_act = z_state_act[:, :y_state.size(1)]
# Now concatenate along the feature dimension
act_input = torch.cat([y_state, z_state_act], dim=-1)
# Apply the act_net and add residual
y_state = self.act_net(act_input) + y_state
else:
# If we can't process the state, duplicate y_state to match expected input dimensions
if y_state.dim() == 3: # [batch, seq_len, features]
# Duplicate the features to match the expected input dimension
act_input = torch.cat([y_state, y_state], dim=-1)
y_state = y_state + self.act_net(act_input)
else: # [batch, features]
# Add sequence dimension and duplicate features
y_state_expanded = y_state.unsqueeze(1) # [batch, 1, features]
act_input = torch.cat([y_state_expanded, y_state_expanded], dim=-1)
y_state = y_state + self.act_net(act_input).squeeze(1)
# Calculate coherence for early stopping
if cycle > 0:
# Default coherence value
current_coherence = 0.5 # Default neutral coherence
# Handle case where z_state might be a tuple
z_state_for_coherence = z_state[0] if isinstance(z_state, (tuple, list)) else z_state
# Ensure both states are tensors and have the same shape
if isinstance(z_state_for_coherence, torch.Tensor) and isinstance(y_state, torch.Tensor):
# Flatten the tensors to 2D [batch*seq_len, d_model]
z_flat = z_state_for_coherence.reshape(-1, d_model)
y_flat = y_state.reshape(-1, d_model)
# Make sure they have the same number of elements
min_len = min(z_flat.size(0), y_flat.size(0))
if min_len > 0:
current_coherence = F.cosine_similarity(
z_flat[:min_len],
y_flat[:min_len],
dim=-1
).mean().item()
# Add to coherence scores
coherence_scores.append(current_coherence)
# Update metadata with running coherence
metadata['coherence_score'] = np.mean(coherence_scores[-5:]) if coherence_scores else 0.0
# Early stopping based on coherence and phase
should_stop = self._should_stop_early(
cycle=cycle,
coherence=current_coherence,
max_cycles=self.max_cycles,
audit_mode=audit_mode
)
if should_stop:
if audit_mode:
print(f"Early stopping at cycle {cycle + 1} with coherence {sim_coherence:.4f}")
break
# Post-processing and metadata updates
metadata.update({
'v_t_score': v_t_mean if 'v_t_mean' in locals() else 0.0,
'coherence_score': np.mean(coherence_scores) if coherence_scores else 0.0,
'cycles_run': cycle + 1,
'final_phase': metadata.get('windsurf_phase', 'UNKNOWN'),
'reflection_ratio': metadata['reflection_count'] / max(1, cycle + 1)
})
# Apply final validation if in audit mode
if audit_mode and hasattr(self, 'validation_protocol'):
self._finalize_validation(x, metadata)
return y_state, metadata
def _should_stop_early(self, cycle: int, coherence: float,
max_cycles: int, audit_mode: bool = False) -> bool:
"""Determine if early stopping conditions are met."""
# Base condition: high coherence
if coherence > 0.85:
return True
# Phase-aware stopping conditions
current_phase = getattr(self, 'current_phase', ValidationPhase.INIT)
if current_phase == ValidationPhase.INIT:
# Allow more exploration in early phases
return False
elif current_phase == ValidationPhase.BREATH:
# More tolerant in breathing phase
return coherence > 0.9 or cycle >= max_cycles - 2
elif current_phase in [ValidationPhase.RITUALS, ValidationPhase.INTEGRITY]:
# Balance exploration and exploitation
min_cycles = min(5, max_cycles // 2)
return (coherence > 0.88 and cycle >= min_cycles) or cycle >= max_cycles - 1
# Default: full cycles for full phase
return cycle >= max_cycles - 1
def _finalize_validation(self, x: torch.Tensor, metadata: Dict[str, Any]) -> None:
"""Finalize validation and update protocol state."""
if not hasattr(self, 'validation_protocol'):
return
# Run final validation step
state = self.validation_protocol.advance_phase(x, "final_validation")
# Update metadata with final validation state
metadata.update({
'validation_passed': state.passed,
'validation_phase': state.phase.name,
'validation_metrics': state.metrics
})
# Log phase transition if applicable
if len(self.validation_protocol.history) > 1:
prev_phase = self.validation_protocol.history[-2].phase
if prev_phase != state.phase:
self.message_bus.publish(
'phase_transition',
{'from': prev_phase.name, 'to': state.phase.name},
priority=2
)
def constrain_gradients(self, gradients: torch.Tensor, param_name: str) -> torch.Tensor:
"""Apply biological constraints to gradients during training."""
if not self.training or not hasattr(self, 'bio_constraints'):
return gradients
return self.bio_constraints.constrain_gradients(gradients, param_name)
def register_optimizer(self, optimizer):
"""Register optimizer for learning rate adjustments."""
self.optimizer = optimizer
def _register_message_handlers(self):
"""Register message handlers for cross-component communication."""
if not hasattr(self, 'message_bus'):
return
# Register phase transition handler
self.message_bus.register_handler('phase_transition', self._handle_phase_transition)
# Register constraint violation handler
self.message_bus.register_handler('constraint_violation', self._handle_constraint_violation)
def _handle_phase_transition(self, data):
"""Handle phase transition events."""
old_phase, new_phase = data.get('from'), data.get('to')
if self.enable_ambient and hasattr(self, 'ambient_core'):
self.ambient_core.on_phase_transition(old_phase, new_phase)
def _handle_constraint_violation(self, data):
"""Handle constraint violation events."""
# Log violations or trigger recovery mechanisms
if self.training:
self._apply_mitigation(data)
def _apply_mitigation(self, violation_data):
"""Apply mitigation for constraint violations."""
# Implement adaptive response to violations
violation_type = violation_data.get('type')
severity = violation_data.get('severity', 1.0)
if violation_type == 'optimization_rate':
# Reduce learning rate or apply gradient clipping
self._adjust_learning_rate(scale=1.0 - (0.1 * severity))
def _adjust_learning_rate(self, scale=0.9):
"""Adjust learning rate for stability."""
for param_group in self.optimizer.param_groups:
param_group['lr'] *= scale
|