File size: 14,667 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 | """
Phase 6: Benchmarking Suite
Measures Phase 6 improvements:
1. Multi-round debate: Does accuracy improve across rounds?
2. Memory weighting: Does memory-boosted routing reduce error?
3. Semantic tension: Are embeddings better than heuristics?
4. Specialization: Are adapters maintaining domain expertise?
Run with: pytest test_phase6_e2e.py -v
"""
import json
import numpy as np
from typing import Dict, List, Tuple
from datetime import datetime
class Phase6Benchmarks:
"""
Comprehensive Phase 6 evaluation suite.
"""
def __init__(self, forge_engine=None):
"""
Initialize benchmarks.
Args:
forge_engine: ForgeEngine instance to test against
"""
self.forge = forge_engine
self.results = {
"timestamp": datetime.now().isoformat(),
"multi_round_convergence": {}, # Coherence per round
"memory_weighting_impact": {}, # With vs. without memory
"semantic_tension_quality": {}, # Embeddings vs heuristics
"specialization_metrics": {}, # Domain expertise scores
}
def benchmark_multi_round_debate(self, queries: List[str], num_rounds: int = 3) -> Dict:
"""
BENCHMARK 1: Multi-Round Debate Convergence
Question: Does multi-round debate improve answer quality?
Hypothesis: As agents debate across rounds:
- Tensions decrease (convergence)
- Coherence increases
- Synthesis accuracy improves
Measurement:
- Run each query through N rounds
- Track coherence_score per round
- Track resolution_rate per round
- Compute convergence rate (tension decay)
Returns:
{
"queries_tested": int,
"rounds_per_query": int,
"coherence_by_round": {round: [scores...]},
"convergence_rate": float,
"improved_queries": int,
}
"""
if not self.forge:
return {"error": "ForgeEngine not available"}
coherence_by_round = {i: [] for i in range(num_rounds)}
resolution_by_round = {i: [] for i in range(num_rounds)}
improved_count = 0
for query in queries:
try:
result = self.forge.forge_with_debate(query, num_rounds=num_rounds)
metadata = result.get("metadata", {})
# Extract per-round metrics
for round_num in range(num_rounds):
round_key = f"round_{round_num}"
if round_key in metadata:
coherence = metadata[round_key].get("coherence", 0.5)
resolution = metadata[round_key].get("resolution_rate", 0.5)
coherence_by_round[round_num].append(coherence)
resolution_by_round[round_num].append(resolution)
# Check if coherence improved from round 0 to final
initial_coh = coherence_by_round[0][-1] if coherence_by_round[0] else 0.5
final_coh = coherence_by_round[num_rounds - 1][-1] if coherence_by_round[num_rounds - 1] else 0.5
if final_coh > initial_coh:
improved_count += 1
except Exception as e:
print(f"Error benchmarking query '{query[:50]}...': {e}")
# Compute statistics
coherence_means = {
i: float(np.mean(scores)) if scores else 0.5 for i, scores in coherence_by_round.items()
}
convergence_rate = 0.0
if num_rounds > 1:
initial = coherence_means.get(0, 0.5)
final = coherence_means.get(num_rounds - 1, 0.5)
if initial > 0:
convergence_rate = (final - initial) / initial # Positive = improvement
self.results["multi_round_convergence"] = {
"queries_tested": len(queries),
"rounds_per_query": num_rounds,
"coherence_by_round": {str(k): round(v, 3) for k, v in coherence_means.items()},
"convergence_rate": round(convergence_rate, 3),
"improved_queries": improved_count,
"improvement_percentage": round(100 * improved_count / max(len(queries), 1), 1),
}
return self.results["multi_round_convergence"]
def benchmark_memory_weighting(self, queries: List[str]) -> Dict:
"""
BENCHMARK 2: Memory Weighting Impact
Question: Does memory-weighted routing reduce error vs. pure keyword routing?
Hypothesis: Adapter weights from past experience guide routing better
than keywords alone.
Measurement:
- Run each query WITHOUT memory weighting (baseline)
- Run each query WITH memory weighting
- Compare: coherence_score, conflict_resolution_rate, adapter_diversity
- Compute improvement delta
Returns:
{
"baseline_coherence": float,
"memory_coherence": float,
"coherence_improvement": float,
"memory_helps_percentage": float,
"avg_resolution_baseline": float,
"avg_resolution_memory": float,
}
"""
if not self.forge:
return {"error": "ForgeEngine not available"}
baseline_coherences = []
memory_coherences = []
baseline_resolutions = []
memory_resolutions = []
for query in queries:
try:
# Baseline: without memory weights
result_baseline = self.forge.forge_with_debate(query, use_memory_weights=False)
baseline_meta = result_baseline.get("metadata", {})
baseline_coherences.append(baseline_meta.get("coherence", 0.5))
baseline_resolutions.append(baseline_meta.get("resolution_rate", 0.5))
# With memory: weights from past performance
result_memory = self.forge.forge_with_debate(query, use_memory_weights=True)
memory_meta = result_memory.get("metadata", {})
memory_coherences.append(memory_meta.get("coherence", 0.5))
memory_resolutions.append(memory_meta.get("resolution_rate", 0.5))
except Exception as e:
print(f"Error in memory weighting benchmark: {e}")
# Compute statistics
baseline_coh = float(np.mean(baseline_coherences)) if baseline_coherences else 0.5
memory_coh = float(np.mean(memory_coherences)) if memory_coherences else 0.5
coh_improve = memory_coh - baseline_coh
baseline_res = float(np.mean(baseline_resolutions)) if baseline_resolutions else 0.5
memory_res = float(np.mean(memory_resolutions)) if memory_resolutions else 0.5
# Percentage of queries where memory helped
improved = sum(1 for b, m in zip(memory_coherences, baseline_coherences) if m > b)
help_percentage = 100 * improved / max(len(queries), 1)
self.results["memory_weighting_impact"] = {
"queries_tested": len(queries),
"baseline_avg_coherence": round(baseline_coh, 3),
"memory_avg_coherence": round(memory_coh, 3),
"coherence_delta": round(coh_improve, 3),
"memory_helps_percentage": round(help_percentage, 1),
"baseline_avg_resolution": round(baseline_res, 3),
"memory_avg_resolution": round(memory_res, 3),
"resolution_delta": round(memory_res - baseline_res, 3),
}
return self.results["memory_weighting_impact"]
def benchmark_semantic_tension(self, conflict_samples: List[Tuple[str, str, float]] = None) -> Dict:
"""
BENCHMARK 3: Semantic Tension Quality
Question: Are embedding-based tensions (ξ_semantic) better than heuristics?
Hypothesis: Semantic embeddings capture *real* disagreement better than
discrete opposition scores (0.4/0.7/1.0).
Measurement:
- For known conflict pairs (with ground truth tension)
- Compute heuristic opposition_score
- Compute semantic_tension (embeddings)
- Measure correlation with ground truth
Args:
conflict_samples: List of (claim_a, claim_b, ground_truth_tension)
Returns:
{
"samples_tested": int,
"heuristic_correlation": float,
"semantic_correlation": float,
"semantic_advantage": float,
}
"""
if not self.forge or not self.forge.semantic_tension_engine:
return {"error": "SemanticTensionEngine not available"}
if not conflict_samples:
return {"error": "No conflict samples provided"}
heuristic_scores = []
semantic_scores = []
ground_truths = []
for claim_a, claim_b, ground_truth in conflict_samples:
try:
# Get semantic tension
semantic_tension = self.forge.semantic_tension_engine.compute_semantic_tension(claim_a, claim_b)
semantic_scores.append(semantic_tension)
# Get heuristic opposition (from conflict engine)
_, heuristic_opposition = self.forge.conflict_engine._classify_conflict(claim_a, claim_b, 0.5)
heuristic_scores.append(heuristic_opposition)
ground_truths.append(ground_truth)
except Exception as e:
print(f"Error computing tensions: {e}")
# Compute correlations with ground truth
if len(heuristic_scores) > 1 and len(ground_truths) > 1:
heuristic_corr = float(np.corrcoef(heuristic_scores, ground_truths)[0, 1])
semantic_corr = float(np.corrcoef(semantic_scores, ground_truths)[0, 1])
advantage = semantic_corr - heuristic_corr
else:
heuristic_corr = 0.0
semantic_corr = 0.0
advantage = 0.0
self.results["semantic_tension_quality"] = {
"samples_tested": len(conflict_samples),
"heuristic_correlation": round(heuristic_corr, 3),
"semantic_correlation": round(semantic_corr, 3),
"semantic_advantage": round(advantage, 3),
"semantic_better": semantic_corr > heuristic_corr,
}
return self.results["semantic_tension_quality"]
def benchmark_specialization(self) -> Dict:
"""
BENCHMARK 4: Specialization Tracking
Question: Are adapters maintaining domain specialization?
Hypothesis: Spec scores trend positive for expert adapters,
negative for generalists. Convergence alerts trigger when
adapter outputs become too similar.
Returns:
{
"adapters_tracked": int,
"specialist_adapters": list,
"generalist_adapters": list,
"convergence_risks": list,
"health_status": str,
}
"""
if not self.forge or not self.forge.specialization:
return {"error": "SpecializationTracker not available"}
system_health = self.forge.specialization.get_system_health()
health_by_adapter = system_health.get("health_by_adapter", {})
specialists = [a for a, h in health_by_adapter.items() if h.get("recommendation") == "excellent_specialist"]
generalists = [a for a, h in health_by_adapter.items() if h.get("recommendation") == "good_generalist"]
convergence_alerts = system_health.get("convergence_alerts", [])
self.results["specialization_metrics"] = {
"adapters_tracked": len(health_by_adapter),
"specialist_adapters": specialists,
"generalist_adapters": generalists,
"convergence_risk_count": len(convergence_alerts),
"health_by_adapter": {a: h.get("recommendation") for a, h in health_by_adapter.items()},
}
return self.results["specialization_metrics"]
def export_results(self, filepath: str = None) -> Dict:
"""
Export all benchmark results to JSON.
Args:
filepath: Where to save results (optional)
Returns:
Complete results dict
"""
if filepath:
with open(filepath, "w") as f:
json.dump(self.results, f, indent=2)
print(f"Benchmark results saved to {filepath}")
return self.results
def summary(self) -> str:
"""
Generate human-readable summary of all benchmarks.
Returns:
Formatted summary string
"""
summary = "PHASE 6 BENCHMARK SUMMARY\n"
summary += "=" * 60 + "\n"
# Multi-round convergence
mr = self.results.get("multi_round_convergence", {})
if mr:
summary += f"\n[1] MULTI-ROUND DEBATE CONVERGENCE\n"
summary += f" Queries tested: {mr.get('queries_tested', 0)}\n"
summary += f" Convergence rate: {mr.get('convergence_rate', 0):.3f}\n"
summary += f" Queries improved: {mr.get('improvement_percentage', 0)}%\n"
# Memory weighting
mw = self.results.get("memory_weighting_impact", {})
if mw:
summary += f"\n[2] MEMORY WEIGHTING IMPACT\n"
summary += f" Baseline coherence: {mw.get('baseline_avg_coherence', 0):.3f}\n"
summary += f" With memory: {mw.get('memory_avg_coherence', 0):.3f}\n"
summary += f" Delta: {mw.get('coherence_delta', 0):.3f}\n"
summary += f" Memory helps: {mw.get('memory_helps_percentage', 0)}% of queries\n"
# Semantic tension
st = self.results.get("semantic_tension_quality", {})
if st:
summary += f"\n[3] SEMANTIC TENSION QUALITY\n"
summary += f" Semantic correlation: {st.get('semantic_correlation', 0):.3f}\n"
summary += f" Heuristic correlation: {st.get('heuristic_correlation', 0):.3f}\n"
summary += f" Semantic advantage: {st.get('semantic_advantage', 0):.3f}\n"
# Specialization
sp = self.results.get("specialization_metrics", {})
if sp:
summary += f"\n[4] ADAPTER SPECIALIZATION\n"
summary += f" Adapters tracked: {sp.get('adapters_tracked', 0)}\n"
summary += f" Specialists: {len(sp.get('specialist_adapters', []))}\n"
summary += f" Convergence risks: {sp.get('convergence_risk_count', 0)}\n"
summary += "\n" + "=" * 60 + "\n"
return summary
__all__ = ["Phase6Benchmarks"]
|