File size: 11,786 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 | """
Phase 6: Specialization Tracker
Monitors adapter specialization and prevents semantic convergence.
Key metrics:
- specialization_score = domain_accuracy / usage_frequency
(higher = expert in domain, not overtaxed)
- semantic_convergence = similarity between adapter outputs
(alert if > 0.85, indicates monoculture within adapters)
Prevents:
- Weight drift (Phase 5 catches at system level)
- Semantic convergence (adapters giving similar answers, Phase 6 catches)
"""
from typing import List, Dict, Optional
import numpy as np
from datetime import datetime
class SpecializationTracker:
"""
Tracks per-adapter per-domain performance to maintain specialization
and detect when adapters are overlapping semantically.
"""
# Domain keywords for query classification
DOMAIN_KEYWORDS = {
"physics": ["force", "momentum", "gravity", "quantum", "relativity", "acceleration", "Newton", "energy"],
"ethics": ["should", "right", "wrong", "moral", "ethics", "justice", "fair", "values", "good"],
"consciousness": ["aware", "conscious", "mind", "self", "experience", "perception", "qualia", "sentient"],
"creativity": ["design", "create", "novel", "innovative", "imagine", "artistic", "original", "aesthetic"],
"systems": ["system", "architecture", "scalable", "complex", "interdependent", "emergence", "network"],
"philosophy": ["meaning", "existence", "truth", "knowledge", "being", "essence", "reasoning"],
}
def __init__(self):
"""Initialize tracking dictionaries."""
self.domain_accuracy = {} # {adapter: {domain: [coherence_scores]}}
self.domain_usage = {} # {adapter: {domain: count}}
self.domain_last_used = {} # {adapter: {domain: timestamp}}
self.query_domains = {} # {query_id: [domain_tags]}
self.semantic_convergence_history = [] # Track convergence over time
def classify_query_domain(self, query: str) -> List[str]:
"""
Classify query by topic domain using keyword heuristics.
Returns:
List of domain tags, e.g., ["physics", "ethics"] for multi-domain queries.
Returns ["general"] if no keywords match.
"""
domains = []
query_lower = query.lower()
for domain, keywords in self.DOMAIN_KEYWORDS.items():
if any(k.lower() in query_lower for k in keywords):
domains.append(domain)
return domains if domains else ["general"]
def record_adapter_performance(self, adapter: str, query: str, coherence: float):
"""
Log adapter performance in domain(s) for a query.
Args:
adapter: Adapter name (e.g., "newton", "empathy")
query: Query text
coherence: Output coherence score [0, 1]
"""
domains = self.classify_query_domain(query)
for domain in domains:
# Initialize if needed
if adapter not in self.domain_accuracy:
self.domain_accuracy[adapter] = {}
self.domain_usage[adapter] = {}
self.domain_last_used[adapter] = {}
if domain not in self.domain_accuracy[adapter]:
self.domain_accuracy[adapter][domain] = []
self.domain_usage[adapter][domain] = 0
self.domain_last_used[adapter][domain] = None
# Record coherence and increment usage
self.domain_accuracy[adapter][domain].append(coherence)
self.domain_usage[adapter][domain] += 1
self.domain_last_used[adapter][domain] = datetime.now()
def compute_specialization(self, adapter: str) -> Dict[str, float]:
"""
Compute specialization_score for each domain an adapter is used in.
specialization_score[domain] = mean_accuracy[domain] / usage_frequency[domain]
Returns:
{domain: specialization_score} for all domains used
Higher = more specialized (good performance, not overused)
"""
if adapter not in self.domain_accuracy:
return {}
specialization = {}
for domain in self.domain_accuracy[adapter]:
accuracies = self.domain_accuracy[adapter][domain]
usage = self.domain_usage[adapter][domain]
mean_accuracy = float(np.mean(accuracies)) if accuracies else 0.5
# Avoid division by zero, natural penalty for high usage
specialization[domain] = mean_accuracy / max(usage, 1)
return specialization
def get_global_specialization(self) -> Dict[str, Dict[str, float]]:
"""
Compute specialization scores for all adapters.
Returns:
{adapter: {domain: specialization_score}}
"""
return {adapter: self.compute_specialization(adapter) for adapter in self.domain_accuracy.keys()}
def detect_domain_expert(self, domain: str) -> Optional[str]:
"""
Find best-performing adapter for a specific domain.
Returns:
Adapter name with highest specialization in domain, or None
"""
specs = self.get_global_specialization()
experts = {a: s.get(domain, 0) for a, s in specs.items() if domain in s}
if not experts:
return None
return max(experts.keys(), key=lambda a: experts[a])
def detect_semantic_convergence(
self, adapter_outputs: Dict[str, str], semantic_engine=None, threshold: float = 0.85
) -> Dict:
"""
Measure overlap between adapter outputs on same query.
Alerts if any pair similarity > threshold (converging).
Args:
adapter_outputs: {adapter_name: output_text}
semantic_engine: SemanticTensionEngine instance (optional, for real embeddings)
threshold: Similarity threshold for convergence alert
Returns:
{
"convergent_pairs": [{pair, similarity, risk}],
"max_similarity": float,
"has_convergence": bool,
}
"""
if len(adapter_outputs) < 2:
return {"convergent_pairs": [], "max_similarity": 0.0, "has_convergence": False}
convergent_pairs = []
max_similarity = 0.0
adapters = list(adapter_outputs.keys())
for i, a1 in enumerate(adapters):
for a2 in adapters[i + 1 :]:
output_a = adapter_outputs[a1]
output_b = adapter_outputs[a2]
# Compute similarity (use semantic engine if available)
if semantic_engine:
try:
tension = semantic_engine.compute_semantic_tension(output_a, output_b)
similarity = 1.0 - tension
except Exception:
# Fallback to text overlap
similarity = self._text_similarity(output_a, output_b)
else:
# Simple fallback: token overlap
similarity = self._text_similarity(output_a, output_b)
max_similarity = max(max_similarity, similarity)
if similarity > threshold:
convergent_pairs.append({
"adapter_a": a1,
"adapter_b": a2,
"similarity": round(similarity, 3),
"convergence_risk": "HIGH" if similarity > 0.92 else "MEDIUM",
})
has_convergence = len(convergent_pairs) > 0
record = {
"timestamp": datetime.now().isoformat(),
"convergent_pairs": convergent_pairs,
"max_similarity": round(max_similarity, 3),
"has_convergence": has_convergence,
"num_adapters": len(adapter_outputs),
}
self.semantic_convergence_history.append(record)
return record
def _text_similarity(self, text_a: str, text_b: str) -> float:
"""
Simple text similarity fallback: Jaccard similarity on tokens.
Args:
text_a, text_b: Text strings
Returns:
Similarity in [0, 1]
"""
tokens_a = set(text_a.lower().split())
tokens_b = set(text_b.lower().split())
if not tokens_a or not tokens_b:
return 0.0
intersection = len(tokens_a & tokens_b)
union = len(tokens_a | tokens_b)
return intersection / max(union, 1)
def get_adapter_health(self, adapter: str) -> Dict:
"""
Get overall health score for an adapter.
Returns:
{
"adapter": adapter,
"num_domains": int,
"avg_accuracy": float,
"total_usage": int,
"specialization_avg": float,
"recommendation": str
}
"""
if adapter not in self.domain_accuracy:
return {"error": f"No data for adapter {adapter}"}
accuracies_all = []
usage_total = 0
for domain in self.domain_accuracy[adapter]:
accuracies_all.extend(self.domain_accuracy[adapter][domain])
usage_total += self.domain_usage[adapter][domain]
avg_accuracy = float(np.mean(accuracies_all)) if accuracies_all else 0.5
specs = self.compute_specialization(adapter)
spec_avg = float(np.mean(list(specs.values()))) if specs else 0.5
# Generate recommendation
if spec_avg > 0.1 and avg_accuracy > 0.75:
recommendation = "excellent_specialist"
elif spec_avg > 0.05 and avg_accuracy > 0.6:
recommendation = "good_generalist"
elif usage_total > 20 and avg_accuracy < 0.5:
recommendation = "overused_poorly"
else:
recommendation = "maintain_current"
return {
"adapter": adapter,
"num_domains": len(self.domain_accuracy[adapter]),
"avg_accuracy": round(avg_accuracy, 3),
"total_usage": usage_total,
"specialization_avg": round(spec_avg, 3),
"recommendation": recommendation,
"domain_specializations": {d: round(s, 3) for d, s in specs.items()},
}
def get_system_health(self) -> Dict:
"""
Get overall system specialization health.
Returns:
Flags convergence risks, identifies experts, recommends actions.
"""
health_by_adapter = {adapter: self.get_adapter_health(adapter) for adapter in self.domain_accuracy.keys()}
overused = [a for a, h in health_by_adapter.items() if h.get("recommendation") == "overused_poorly"]
excellent = [a for a, h in health_by_adapter.items() if h.get("recommendation") == "excellent_specialist"]
experts = {domain: self.detect_domain_expert(domain) for domain in self.DOMAIN_KEYWORDS.keys()}
return {
"timestamp": datetime.now().isoformat(),
"total_adapters": len(health_by_adapter),
"health_by_adapter": health_by_adapter,
"overused_adapters": overused,
"specialist_adapters": excellent,
"domain_experts": experts,
"convergence_alerts": self.semantic_convergence_history[-5:] if self.semantic_convergence_history else [],
}
def export_summary(self) -> Dict:
"""Export complete specialization data for analysis."""
return {
"timestamp": datetime.now().isoformat(),
"global_specialization": self.get_global_specialization(),
"system_health": self.get_system_health(),
"convergence_history": self.semantic_convergence_history,
}
__all__ = ["SpecializationTracker"]
|