File size: 18,772 Bytes
ed1b365 | 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 | """
Token-Level Confidence Scoring Engine
Synthesizes four confidence signals to rate individual claims:
1. Semantic Confidence — Confidence markers in text ("I'm confident that...", "arguably...")
2. Attentional Confidence — Semantic overlap with other agents' responses
3. Probabilistic Confidence — Token-level probabilities from LLM logits
4. Integrated Learning Signal — Historical coherence from past similar responses
Author: Claude Code
"""
import re
import time
import json
from dataclasses import dataclass, asdict
from typing import Dict, List, Optional, Tuple, Any
from collections import defaultdict
import hashlib
import logging
logging.basicConfig(level=logging.INFO)
logger = logging.getLogger(__name__)
# Confidence markers (grouped by confidence level)
CONFIDENCE_MARKERS = {
"high": [
r"\bi['\"]?m confident\b",
r"\bdefinitively\b",
r"\bclearly\b",
r"\bunambiguously\b",
r"\bcertainly\b",
r"\bwithout doubt\b",
r"\bno question\b",
r"\bproven\b",
r"\bestablished fact\b",
],
"medium": [
r"\bi argue\b",
r"\b(it appears|it seems)\b",
r"\breasonably\b",
r"\barguably\b",
r"\blikely\b",
r"\bprobably\b",
r"\bin my view\b",
r"\bi think\b",
r"\bi believe\b",
r"\bfrom my perspective\b",
],
"low": [
r"\b(it['\"]?s possible|it could be)\b",
r"\bone could say\b",
r"\bperhaps\b",
r"\bmaybe\b",
r"\buncertain\b",
r"\bi['\"]?m not sure\b",
r"\ballegedly\b",
r"\bseemingly\b",
r"\bapparently\b",
r"\bwhoa\b",
],
}
# Compile regex patterns for performance
_MARKER_PATTERNS = {}
for level, markers in CONFIDENCE_MARKERS.items():
_MARKER_PATTERNS[level] = [re.compile(m, re.IGNORECASE) for m in markers]
@dataclass
class ClaimSegment:
"""A single claim extracted from an agent's response."""
text: str # The claim text
start_idx: int # Position in original response
end_idx: int # End position
confidence: float # Aggregate confidence [0, 1]
semantic_conf: float # From markers
attentional_conf: float # From semantic overlap with peers
probabilistic_conf: float # From logits (if available)
learning_signal: float # From historical coherence
agent_name: str = "" # Which agent produced this
debate_round: int = 0
@dataclass
class TokenConfidenceScore:
"""Per-token confidence analysis for a full response."""
agent_name: str
response_text: str
token_scores: List[float] # [0, 1] per token (or sentence)
claims: List[ClaimSegment]
semantic_confidence_dict: Dict[int, float] # Token idx -> semantic confidence
attentional_confidence_dict: Dict[int, float] # Token idx -> attentional confidence
probabilistic_confidence_dict: Dict[int, float] # Token idx -> logit probability
learning_signal_dict: Dict[int, float] # Token idx -> learning signal
composite_scores: Dict[int, float] # Token idx -> composite [α, β, γ, δ]
timestamp: float = 0.0
def __post_init__(self):
if self.timestamp == 0.0:
self.timestamp = time.time()
def to_dict(self) -> Dict:
"""Serialize for storage."""
return {
"agent_name": self.agent_name,
"response_text": self.response_text[:500],
"mean_token_confidence": sum(self.token_scores) / max(len(self.token_scores), 1),
"claims_count": len(self.claims),
"claims": [
{
"text": c.text,
"confidence": c.confidence,
"semantic_conf": c.semantic_conf,
"attentional_conf": c.attentional_conf,
"probabilistic_conf": c.probabilistic_conf,
"learning_signal": c.learning_signal,
}
for c in self.claims
],
}
class TokenConfidenceEngine:
"""Four-signal token confidence scorer."""
def __init__(
self,
embedding_model: Optional[Any] = None,
living_memory: Optional[Any] = None,
alpha: float = 0.25,
beta: float = 0.25,
gamma: float = 0.25,
delta: float = 0.25,
):
"""
Initialize token confidence engine.
Args:
embedding_model: Model for generating embeddings (optional, uses sklearn if None)
living_memory: LivingMemoryKernel instance for historical coherence lookup
alpha: Weight for semantic confidence
beta: Weight for attentional confidence
gamma: Weight for probabilistic confidence
delta: Weight for learning signal
"""
self.embedding_model = embedding_model
self.living_memory = living_memory
self.alpha = alpha
self.beta = beta
self.gamma = gamma
self.delta = delta
# Lazy-loaded embedder (sklearn TfidfVectorizer for lightweight usage)
self._embedder = None
self._embedder_cache = {}
def score_tokens(
self,
agent_response: str,
agent_name: str,
peer_responses: Optional[Dict[str, str]] = None,
logits: Optional[List[float]] = None,
) -> TokenConfidenceScore:
"""
Score all tokens/claims in an agent's response using 4 signals.
Args:
agent_response: The full response text from the agent
agent_name: Name of the agent (for memory lookup)
peer_responses: Dict {peer_agent_name: response_text} for attentional scoring
logits: Optional list of per-token probabilities from generation
Returns:
TokenConfidenceScore with all components
"""
if peer_responses is None:
peer_responses = {}
# Step 1: Parse semantic confidence markers
semantic_conf_dict = self._parse_semantic_markers(agent_response)
# Step 2: Compute attentional confidence (semantic overlap with peers)
attentional_conf_dict = self._compute_attentional_confidence(
agent_response, peer_responses
)
# Step 3: Probabilistic confidence from logits (if provided)
probabilistic_conf_dict = self._extract_probabilistic_confidence(
agent_response, logits
)
# Step 4: Learning signal from memory (historical coherence)
learning_signal_dict = self._compute_learning_signal(
agent_response, agent_name
)
# Step 5: Extract claims and compute aggregate confidence per claim
claims = self._extract_claims(
agent_response,
semantic_conf_dict,
attentional_conf_dict,
probabilistic_conf_dict,
learning_signal_dict,
agent_name,
)
# Step 6: Synthesize composite confidence scores
token_scores = []
composite_scores = {}
for i, token_text in enumerate(agent_response.split()):
semantic = semantic_conf_dict.get(i, 0.5)
attentional = attentional_conf_dict.get(i, 0.5)
probabilistic = probabilistic_conf_dict.get(i, 0.5)
learning = learning_signal_dict.get(i, 0.5)
# Weighted synthesis
composite = (
self.alpha * semantic
+ self.beta * attentional
+ self.gamma * probabilistic
+ self.delta * learning
)
composite = max(0.0, min(1.0, composite)) # Clamp to [0, 1]
token_scores.append(composite)
composite_scores[i] = composite
return TokenConfidenceScore(
agent_name=agent_name,
response_text=agent_response,
token_scores=token_scores,
claims=claims,
semantic_confidence_dict=semantic_conf_dict,
attentional_confidence_dict=attentional_conf_dict,
probabilistic_confidence_dict=probabilistic_conf_dict,
learning_signal_dict=learning_signal_dict,
composite_scores=composite_scores,
)
def _parse_semantic_markers(self, response: str) -> Dict[int, float]:
"""
Parse confidence markers from text.
Returns:
Dict mapping token_idx to confidence [0, 1]
"""
conf_dict = {}
tokens = response.split()
# Find spans of confidence markers and propagate confidence to nearby tokens
for level, confidence_level in [("high", 0.9), ("medium", 0.6), ("low", 0.3)]:
for pattern in _MARKER_PATTERNS[level]:
for match in pattern.finditer(response):
# Map character position to token index
char_pos = match.start()
char_count = 0
for token_idx, token in enumerate(tokens):
if char_count <= char_pos < char_count + len(token):
# Mark this token and nearby tokens
for nearby_idx in range(
max(0, token_idx - 1), min(len(tokens), token_idx + 4)
):
if nearby_idx not in conf_dict:
conf_dict[nearby_idx] = confidence_level
else:
# Take max confidence found
conf_dict[nearby_idx] = max(
conf_dict[nearby_idx], confidence_level
)
break
char_count += len(token) + 1 # +1 for space
# Default to neutral for unscored tokens
for i in range(len(tokens)):
if i not in conf_dict:
conf_dict[i] = 0.5
return conf_dict
def _compute_attentional_confidence(
self, agent_response: str, peer_responses: Dict[str, str]
) -> Dict[int, float]:
"""
Compute attentional confidence via semantic overlap with peers.
High overlap = higher confidence (claim addresses peer perspectives)
Returns:
Dict mapping token_idx to confidence [0.3, 1.0]
"""
conf_dict = {}
tokens = agent_response.split()
if not peer_responses:
# No peers → neutral attentional score
for i in range(len(tokens)):
conf_dict[i] = 0.5
return conf_dict
# Compute token-level overlap with each peer
token_overlaps = defaultdict(list)
for peer_name, peer_response in peer_responses.items():
peer_tokens_set = set(peer_response.lower().split())
for token_idx, token in enumerate(tokens):
# Check if this token or semantically similar tokens appear in peer
if token.lower() in peer_tokens_set:
token_overlaps[token_idx].append(1.0)
elif any(
token.lower().startswith(p[:3]) or p.startswith(token.lower()[:3])
for p in peer_tokens_set
):
# Partial match (first 3 chars)
token_overlaps[token_idx].append(0.6)
# Aggregate overlap: mean overlap with peers, map to [0.3, 1.0]
for i in range(len(tokens)):
if token_overlaps[i]:
overlap_score = sum(token_overlaps[i]) / len(token_overlaps[i])
else:
overlap_score = 0.0
# Scale to [0.3, 1.0]: low overlap agents get 0.3, high get 1.0
attentional_conf = 0.3 + 0.7 * overlap_score
conf_dict[i] = attentional_conf
return conf_dict
def _extract_probabilistic_confidence(
self, response: str, logits: Optional[List[float]] = None
) -> Dict[int, float]:
"""
Extract per-token probabilities from logits.
If logits not provided, use fallback heuristic (all 0.5).
Returns:
Dict mapping token_idx to probability [0, 1]
"""
conf_dict = {}
tokens = response.split()
if logits and len(logits) == len(tokens):
# Direct logit probabilities
for i, prob in enumerate(logits):
conf_dict[i] = max(0.0, min(1.0, prob))
else:
# Fallback: common words get higher confidence
common_words = {
"the",
"a",
"is",
"and",
"or",
"of",
"to",
"in",
"that",
"it",
}
for i, token in enumerate(tokens):
if token.lower() in common_words:
conf_dict[i] = 0.9 # Very common
elif len(token) > 3:
conf_dict[i] = 0.6 # More specific words
else:
conf_dict[i] = 0.5 # Neutral
return conf_dict
def _compute_learning_signal(
self, response: str, agent_name: str
) -> Dict[int, float]:
"""
Compute learning signal from historical coherence (Phase 2 enhancement).
Query memory for similar past responses and boost confidence if
they led to high coherence. Recent memories are weighted higher.
Returns:
Dict mapping token_idx to learning signal [0.5, 1.0]
Phase 2: Now includes recency weighting with ~7 day half-life
"""
import math
conf_dict = {}
tokens = response.split()
# If no memory, return neutral signal
if not self.living_memory:
for i in range(len(tokens)):
conf_dict[i] = 0.5
return conf_dict
# Retrieve past responses by this agent
try:
similar_cocoons = self.living_memory.recall_by_adapter(
agent_name, limit=10
)
if not similar_cocoons:
avg_coherence = 0.5
else:
# Phase 2: Weight recent memories higher
# Using exponential decay with ~7 day half-life
recency_weights = []
weighted_coherences = []
for cocoon in similar_cocoons:
age_hours = cocoon.age_hours()
# exp(-age_hours / 168) = 0.5 after 168 hours (~7 days)
recency_weight = math.exp(-age_hours / 168.0)
recency_weights.append(recency_weight)
weighted_coherences.append(cocoon.coherence * recency_weight)
# Compute weighted average
total_weight = sum(recency_weights)
if total_weight > 0:
avg_coherence = sum(weighted_coherences) / total_weight
else:
avg_coherence = 0.5
except Exception as e:
logger.warning(f"Error retrieving memory for {agent_name}: {e}")
avg_coherence = 0.5
# Boost confidence proportional to historical coherence
# learning_signal = 0.5 + 0.5 * avg_coherence → [0.5, 1.0]
learning_signal = 0.5 + 0.5 * avg_coherence
for i in range(len(tokens)):
conf_dict[i] = learning_signal
return conf_dict
def _extract_claims(
self,
response: str,
semantic_conf_dict: Dict[int, float],
attentional_conf_dict: Dict[int, float],
probabilistic_conf_dict: Dict[int, float],
learning_signal_dict: Dict[int, float],
agent_name: str,
) -> List[ClaimSegment]:
"""
Extract individual claims (sentences/clauses) from response.
Returns:
List of ClaimSegment with aggregate confidence from component signals
"""
claims = []
# Simple segmentation: split on sentence boundaries
sentence_pattern = re.compile(r"[.!?]+")
sentences = sentence_pattern.split(response)
token_idx = 0
start_char_idx = 0
for sentence in sentences:
if not sentence.strip():
continue
sentence_tokens = sentence.split()
sentence_token_indices = list(range(token_idx, token_idx + len(sentence_tokens)))
token_idx += len(sentence_tokens)
# Aggregate confidence across sentence tokens
if sentence_token_indices:
semantic = sum(
semantic_conf_dict.get(i, 0.5) for i in sentence_token_indices
) / len(sentence_token_indices)
attentional = sum(
attentional_conf_dict.get(i, 0.5) for i in sentence_token_indices
) / len(sentence_token_indices)
probabilistic = sum(
probabilistic_conf_dict.get(i, 0.5) for i in sentence_token_indices
) / len(sentence_token_indices)
learning = sum(
learning_signal_dict.get(i, 0.5) for i in sentence_token_indices
) / len(sentence_token_indices)
composite_confidence = (
self.alpha * semantic
+ self.beta * attentional
+ self.gamma * probabilistic
+ self.delta * learning
)
composite_confidence = max(0.0, min(1.0, composite_confidence))
claim = ClaimSegment(
text=sentence.strip(),
start_idx=start_char_idx,
end_idx=start_char_idx + len(sentence),
confidence=composite_confidence,
semantic_conf=semantic,
attentional_conf=attentional,
probabilistic_conf=probabilistic,
learning_signal=learning,
agent_name=agent_name,
)
claims.append(claim)
start_char_idx += len(sentence) + 1 # +1 for sentence separator
return claims
|