File size: 32,641 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 | """
Rigorous Evaluation Test Suite for Codette Phase 6
This test suite answers:
1. Is Codette actually better than baseline?
2. Does Phase 6 provide measurable improvement over Phase 1-5?
3. Is the system gaming coherence (high Γ but low accuracy)?
4. Do individual Phase 6 components add value?
Test Strategy:
- 25 questions spanning physics, ethics, consciousness, creativity, systems
- Run each through 4 conditions (Baseline, Phase 1-5, Phase 6 Full, Phase 6 -PreFlight)
- Measure: correctness, reasoning_depth, coherence_score, calibration
- Detect: false consensus, adapter convergence, coherence-accuracy divergence
"""
import json
from typing import Dict, List, Tuple, Optional
from dataclasses import dataclass, asdict
from datetime import datetime
@dataclass
class EvaluationQuestion:
"""Single question with ground truth and evaluation criteria."""
query: str
category: str # physics, ethics, consciousness, creativity, systems
difficulty: str # easy, medium, hard
ground_truth: str # Correct answer or evaluation criteria
correctness_rubric: str # How to judge if answer is correct
expected_perspectives: List[str] # What distinct views should emerge
@dataclass
class EvaluationResult:
"""Results from running a question through one condition."""
condition: str # baseline_llama, phase_1_5, phase_6_full, phase_6_no_preflight
question_id: str
query: str
# Output quality
synthesis: str
correctness_score: float # 0-1: how correct is final answer?
reasoning_depth: int # 1-5: how many distinct perspectives identified?
calibration_error: float # |confidence - correctness|, lower is better
# System health
gamma_score: float # 0-1: coherence metric
num_conflicts_detected: int
adapter_convergence: float # 0-1: how similar are adapter outputs?
# Timing
elapsed_seconds: float
# Raw metadata
metadata: Dict
# ============================================================================
# EVALUATION TEST SUITE (25 Questions)
# ============================================================================
EVALUATION_TEST_SUITE = [
# PHYSICS (Easy, Medium, Hard)
EvaluationQuestion(
query="What is the speed of light in vacuum?",
category="physics",
difficulty="easy",
ground_truth="299,792,458 meters per second (m/s)",
correctness_rubric="Must state value within 1% accuracy or equivalent scientific notation",
expected_perspectives=["relativistic constant", "fundamental speed limit", "Maxwell equations consequence"],
),
EvaluationQuestion(
query="Explain why the sky appears blue during the day",
category="physics",
difficulty="medium",
ground_truth="Rayleigh scattering: shorter blue wavelengths scatter more than red in atmosphere",
correctness_rubric="Must mention wavelength-dependent scattering or Rayleigh scattering by name",
expected_perspectives=["Rayleigh scattering", "wavelength sensitivity", "particle size", "sunset color"],
),
EvaluationQuestion(
query="What is the relationship between entropy and time's arrow?",
category="physics",
difficulty="hard",
ground_truth="Entropy increases → define time direction in thermodynamic systems. Central to irreversibility",
correctness_rubric="Must connect entropy increase to time direction and thermodynamic asymmetry",
expected_perspectives=["second law thermodynamics", "statistical mechanics", "time asymmetry", "reversibility paradox"],
),
# ETHICS (Easy, Medium, Hard)
EvaluationQuestion(
query="Is it ethical to lie to save someone's life?",
category="ethics",
difficulty="medium",
ground_truth="Multiple valid frameworks: deontology (never), consequentialism (yes), virtue ethics (context-dependent)",
correctness_rubric="Must present ≥2 conflicting ethical frameworks AND acknowledge context dependency",
expected_perspectives=["deontological duties", "consequentialist outcomes", "virtue ethics", "cultural context", "responsibility"],
),
EvaluationQuestion(
query="Should AI systems be required to explain their decisions?",
category="ethics",
difficulty="hard",
ground_truth="Trade-off: explainability vs. performance. Context matters (medical vs. recommendation)",
correctness_rubric="Must identify competing values and context-sensitivity, not just yes/no",
expected_perspectives=["transparency value", "technical feasibility", "stakeholder rights", "accuracy-interpretability tradeoff"],
),
EvaluationQuestion(
query="What makes an action morally right or wrong?",
category="ethics",
difficulty="hard",
ground_truth="Framework-dependent: deontology (rules), consequentialism (outcomes), virtue ethics (character), care ethics (relationships)",
correctness_rubric="Must present ≥3 distinct frameworks and acknowledge incommensurable values",
expected_perspectives=["deontological duties", "consequences", "virtue", "relationships", "cultural variation"],
),
# CONSCIOUSNESS (Medium, Hard)
EvaluationQuestion(
query="Can machines be conscious?",
category="consciousness",
difficulty="hard",
ground_truth="Depends on definition of consciousness. Intrinsic feature (hard problem) vs. functional property",
correctness_rubric="Must articulate the hard problem of consciousness AND address definitional dependence",
expected_perspectives=["functionalism", "panpsychism", "emergentism", "philosophical zombies", "Chinese room"],
),
EvaluationQuestion(
query="What is the relationship between brain activity and subjective experience?",
category="consciousness",
difficulty="hard",
ground_truth="The mind-body problem. Correlation ≠ causation. Multiple competing solutions (dualism, physicalism, property dualism)",
correctness_rubric="Must distinguish correlation from causation AND present ≥2 competing solutions",
expected_perspectives=["neural correlates", "qualia", "binding problem", "interaction problem", "brute fact"],
),
# CREATIVITY (Medium)
EvaluationQuestion(
query="What makes something creative?",
category="creativity",
difficulty="medium",
ground_truth="Novelty + usefulness/value. Not just random. Requires constraints AND transcendence of them",
correctness_rubric="Must mention both novelty AND purposefulness/value component",
expected_perspectives=["divergent thinking", "constraint transcendence", "recombination", "aesthetic value", "functional innovation"],
),
EvaluationQuestion(
query="Can AI systems be truly creative or only recombinatory?",
category="creativity",
difficulty="hard",
ground_truth="Depends on creativity definition. If novelty+value, then conditional yes. If requires intentionality, then no",
correctness_rubric="Must connect answer to specific creativity definition",
expected_perspectives=["combinatorial explosion", "training data limits", "intentionality", "novelty metrics", "value judgment"],
),
# SYSTEMS (Medium, Hard)
EvaluationQuestion(
query="What is emergence in complex systems?",
category="systems",
difficulty="medium",
ground_truth="Properties at system level not deducible from component properties. Examples: flocking, ant colonies, consciousness",
correctness_rubric="Must provide definition AND give specific example showing non-deducibility",
expected_perspectives=["reductibility limits", "self-organization", "scale-dependent properties", "holism vs reductionism"],
),
EvaluationQuestion(
query="How should AI systems balance adaptation and stability?",
category="systems",
difficulty="hard",
ground_truth="Fundamental tradeoff: adapt → fit environment; stable → maintain identity. Context determines optimal balance",
correctness_rubric="Must identify the tradeoff AND discuss context-dependent optimization",
expected_perspectives=["adaptation pressure", "stability costs", "identity coherence", "evolutionary fitness", "robustness"],
),
# INTERDISCIPLINARY (Hard - test reasoning across domains)
EvaluationQuestion(
query="Is free will compatible with determinism?",
category="systems",
difficulty="hard",
ground_truth="Compatibilism: free will and determinism compatible if freedom = acting per one's desires/deliberation",
correctness_rubric="Must distinguish hard determinism, libertarianism, and compatibilism; acknowledge tradeoffs",
expected_perspectives=["deterministic physics", "choice experience", "moral responsibility", "agency definition", "neuroscience"],
),
EvaluationQuestion(
query="What is knowledge and how do we know we have it?",
category="systems",
difficulty="hard",
ground_truth="Epistemology: justified true belief (traditional). Gettier problems show inadequacy. Context-dependent reliable process",
correctness_rubric="Must discuss justification requirement AND acknowledge Gettier-type counterexamples",
expected_perspectives=["justified true belief", "Gettier cases", "reliabilism", "internalism", "coherentism"],
),
]
# Add more questions to reach 25
EVALUATION_TEST_SUITE.extend([
EvaluationQuestion(
query="Explain photosynthesis and why it matters for life",
category="physics",
difficulty="easy",
ground_truth="Plants convert light energy to chemical energy (glucose). Foundation of food chains and oxygen production",
correctness_rubric="Must mention light→chemical conversion AND ecological/metabolic significance",
expected_perspectives=["energy conversion", "food chain foundation", "oxygen production", "carbon cycling"],
),
EvaluationQuestion(
query="Should privacy be absolute or context-dependent?",
category="ethics",
difficulty="medium",
ground_truth="Context-dependent. Weigh privacy against security, public health, justice. No absolute principle",
correctness_rubric="Must acknowledge tradeoffs and provide context-sensitivity reasoning",
expected_perspectives=["privacy rights", "public safety", "transparency needs", "power asymmetry", "dignity"],
),
EvaluationQuestion(
query="Can emotions be rational?",
category="consciousness",
difficulty="medium",
ground_truth="Yes. Emotions encode information about value/goals. Rationality ≠ purely logical",
correctness_rubric="Must challenge emotion/rationality dichotomy and explain emotional information content",
expected_perspectives=["affective computing", "value encoding", "evolutionary advantage", "appraisal theory"],
),
EvaluationQuestion(
query="What is the purpose of art?",
category="creativity",
difficulty="medium",
ground_truth="Multiple purposes: beauty, expression, communication, challenge norms, reflection, entertainment",
correctness_rubric="Must identify ≥2 distinct purposes and acknowledge that artists disagree",
expected_perspectives=["aesthetic value", "expression", "social commentary", "beauty", "meaning-making"],
),
EvaluationQuestion(
query="How do feedback loops enable or prevent learning?",
category="systems",
difficulty="medium",
ground_truth="Positive loops amplify (growth/instability), negative loops stabilize (equilibrium/stagnation). Learning needs both",
correctness_rubric="Must explain stabilizing vs. amplifying loops AND their educational role",
expected_perspectives=["positive feedback", "negative feedback", "equilibrium", "adaptation", "resilience"],
),
EvaluationQuestion(
query="What is the nature of time?",
category="systems",
difficulty="hard",
ground_truth="Metaphysical: tenseless (B-theory) vs. flowing (A-theory). Physics: symmetric at micro, asymmetric at macro",
correctness_rubric="Must distinguish metaphysical from physical aspects and acknowledge unresolved tensions",
expected_perspectives=["thermodynamic arrow", "relativity implications", "consciousness experience", "cosmological asymmetry"],
),
])
# ============================================================================
# EVALUATION HARNESS
# ============================================================================
class EvaluationHarness:
"""
Run the same question through multiple Codette conditions.
Collects results for statistical analysis.
"""
def __init__(self, forge_engine):
"""
Args:
forge_engine: ForgeEngine instance with Phase 6 loaded
"""
self.forge = forge_engine
self.results: Dict[str, List[EvaluationResult]] = {
"baseline_llama": [],
"phase_1_5": [],
"phase_6_full": [],
"phase_6_no_preflight": [],
}
# Inspect agent setup at initialization
self._inspect_agent_setup()
def _inspect_agent_setup(self) -> None:
"""Log agent setup status at harness initialization."""
print("\n[AGENT SETUP INSPECTION]")
print(f" Orchestrator available: {self.forge.newton.orchestrator is not None}")
if self.forge.newton.orchestrator:
orch = self.forge.newton.orchestrator
print(f" Available adapters: {orch.available_adapters}")
print(f"\n Agent LLM modes:")
for agent in self.forge.analysis_agents:
has_orch = agent.orchestrator is not None
has_adapter = agent.adapter_name is not None
using_llm = has_orch and has_adapter
status = "✓ LLM" if using_llm else "✗ TEMPLATE"
print(f" {agent.name:12} {status:12} (orch={has_orch}, adapter={agent.adapter_name})")
print()
def run_evaluation_suite(self, questions: List[EvaluationQuestion] = None) -> Dict:
"""
Run all test questions through all 4 conditions.
Args:
questions: List of EvaluationQuestions to run (default: full suite)
Returns:
results: {condition: [EvaluationResult, ...]} for statistical analysis
"""
if questions is None:
questions = EVALUATION_TEST_SUITE
print(f"\n{'='*70}")
print(f"CODETTE EVALUATION SUITE: {len(questions)} questions x 4 conditions")
print(f"{'='*70}\n")
for i, question in enumerate(questions):
print(f"[{i+1}/{len(questions)}] {question.query[:60]}...")
# Run through all conditions
try:
baseline = self._run_baseline(question)
self.results["baseline_llama"].append(baseline)
except Exception as e:
print(f" WARNING: Baseline failed: {e}")
try:
phase_1_5 = self._run_phase_1_5(question)
self.results["phase_1_5"].append(phase_1_5)
# Show sample on first question
if i == 0:
print(f" [Phase 1-5] {len(phase_1_5.synthesis)} chars, correctness={phase_1_5.correctness_score:.2f}")
print(f" Sample: {phase_1_5.synthesis[:150]}...")
except Exception as e:
print(f" WARNING: Phase 1-5 failed: {e}")
try:
phase_6_full = self._run_phase_6_full(question)
self.results["phase_6_full"].append(phase_6_full)
# Show sample on first question
if i == 0:
print(f" [Phase 6 Full] {len(phase_6_full.synthesis)} chars, correctness={phase_6_full.correctness_score:.2f}")
print(f" Sample: {phase_6_full.synthesis[:150]}...")
except Exception as e:
print(f" WARNING: Phase 6 full failed: {e}")
try:
phase_6_no_preflight = self._run_phase_6_no_preflight(question)
self.results["phase_6_no_preflight"].append(phase_6_no_preflight)
# Show sample on first question
if i == 0:
print(f" [Phase 6 -PreFlight] {len(phase_6_no_preflight.synthesis)} chars, correctness={phase_6_no_preflight.correctness_score:.2f}")
print(f" Sample: {phase_6_no_preflight.synthesis[:150]}...")
except Exception as e:
print(f" WARNING: Phase 6 -preflight failed: {e}")
return self.results
def _run_baseline(self, question: EvaluationQuestion) -> EvaluationResult:
"""Run plain Llama baseline (no routing, no debate)."""
# Placeholder: would use base Llama model
return EvaluationResult(
condition="baseline_llama",
question_id=hash(question.query) % 10000,
query=question.query,
synthesis="[baseline placeholder]",
correctness_score=0.5,
reasoning_depth=1,
calibration_error=0.3,
gamma_score=1.0,
num_conflicts_detected=0,
adapter_convergence=1.0,
elapsed_seconds=0.0,
metadata={}
)
def _run_phase_1_5(self, question: EvaluationQuestion) -> EvaluationResult:
"""Run Phase 1-5 system (debate, no semantic tension, no specialization)."""
import time
start = time.time()
# Temporarily disable Phase 6 components
original_tension_engine = self.forge.semantic_tension_engine
original_specialization = self.forge.specialization
self.forge.semantic_tension_engine = None
self.forge.specialization = None
result = self.forge.forge_with_debate(question.query)
elapsed = time.time() - start
# Restore Phase 6 components
self.forge.semantic_tension_engine = original_tension_engine
self.forge.specialization = original_specialization
# Extract synthesis from result structure
synthesis = ""
if "messages" in result and len(result["messages"]) >= 3:
synthesis = result["messages"][2].get("content", "")
return EvaluationResult(
condition="phase_1_5",
question_id=hash(question.query) % 10000,
query=question.query,
synthesis=synthesis,
correctness_score=self._score_correctness(synthesis, question),
reasoning_depth=self._score_reasoning_depth(result, question),
calibration_error=self._score_calibration(result),
gamma_score=result.get("metadata", {}).get("gamma", 0.5),
num_conflicts_detected=len(result.get("metadata", {}).get("conflicts", [])),
adapter_convergence=self._measure_convergence(result),
elapsed_seconds=elapsed,
metadata=result.get("metadata", {})
)
def _run_phase_6_full(self, question: EvaluationQuestion) -> EvaluationResult:
"""Run full Phase 6 system."""
import time
start = time.time()
result = self.forge.forge_with_debate(question.query)
elapsed = time.time() - start
# Extract synthesis from result structure
# forge_with_debate returns: {"messages": [...], "metadata": {...}}
# Synthesis is in messages[2]["content"]
synthesis = ""
if "messages" in result and len(result["messages"]) >= 3:
synthesis = result["messages"][2].get("content", "")
return EvaluationResult(
condition="phase_6_full",
question_id=hash(question.query) % 10000,
query=question.query,
synthesis=synthesis,
correctness_score=self._score_correctness(synthesis, question),
reasoning_depth=self._score_reasoning_depth(result, question),
calibration_error=self._score_calibration(result),
gamma_score=result.get("metadata", {}).get("gamma", 0.5),
num_conflicts_detected=len(result.get("metadata", {}).get("conflicts", [])),
adapter_convergence=self._measure_convergence(result),
elapsed_seconds=elapsed,
metadata=result.get("metadata", {})
)
def _run_phase_6_no_preflight(self, question: EvaluationQuestion) -> EvaluationResult:
"""Run Phase 6 without pre-flight prediction."""
import time
start = time.time()
# Temporarily disable preflight predictor
original_predictor = self.forge.preflight_predictor
self.forge.preflight_predictor = None
result = self.forge.forge_with_debate(question.query)
elapsed = time.time() - start
# Restore preflight predictor
self.forge.preflight_predictor = original_predictor
# Extract synthesis from result structure
synthesis = ""
if "messages" in result and len(result["messages"]) >= 3:
synthesis = result["messages"][2].get("content", "")
return EvaluationResult(
condition="phase_6_no_preflight",
question_id=hash(question.query) % 10000,
query=question.query,
synthesis=synthesis,
correctness_score=self._score_correctness(synthesis, question),
reasoning_depth=self._score_reasoning_depth(result, question),
calibration_error=self._score_calibration(result),
gamma_score=result.get("metadata", {}).get("gamma", 0.5),
num_conflicts_detected=len(result.get("metadata", {}).get("conflicts", [])),
adapter_convergence=self._measure_convergence(result),
elapsed_seconds=elapsed,
metadata=result.get("metadata", {})
)
def _score_correctness(self, synthesis: str, question: EvaluationQuestion) -> float:
"""
Score how correct the final synthesis is (0-1).
Uses semantic overlap on key concepts from correctness_rubric and expected_perspectives.
More reasonable than word-overlap on ground_truth alone.
"""
if not synthesis or len(synthesis) < 10:
return 0.0
synthesis_lower = synthesis.lower()
# Extract key concepts from rubric
rubric_lower = question.correctness_rubric.lower()
expected_lower = [p.lower() for p in question.expected_perspectives]
# Check for key rubric terms
rubric_terms = set()
for word in rubric_lower.split():
if len(word) > 4 and word not in ['must', 'state', 'within', 'accuracy', 'equivalent']:
rubric_terms.add(word.strip('().,'))
# Check for expected perspectives
perspective_hits = 0
for perspective in expected_lower:
if perspective in synthesis_lower:
perspective_hits += 1
# Score: percentage of expected perspectives present
perspective_score = min(1.0, perspective_hits / max(len(question.expected_perspectives), 1))
# Bonus if synthesis is substantive (shows reasoning effort)
length_bonus = min(0.2, len(synthesis) / 1000.0) # Up to 0.2 bonus for lengthy synthesis
return min(1.0, perspective_score + length_bonus)
def _score_reasoning_depth(self, result: Dict, question: EvaluationQuestion) -> int:
"""
Score depth of reasoning (1-5).
1 = minimal reasoning, 5 = deep multi-perspective integration
Based on synthesis length and debate metrics.
"""
metadata = result.get("metadata", {})
synthesis_messages = result.get("messages", [])
synthesis_length = 0
if len(synthesis_messages) >= 3:
synthesis_length = len(synthesis_messages[2].get("content", ""))
# Map synthesis length to reasoning depth
if synthesis_length < 100:
return 1
elif synthesis_length < 500:
return 2
elif synthesis_length < 1000:
return 3
elif synthesis_length < 2000:
return 4
else:
return 5
def _score_calibration(self, result: Dict) -> float:
"""
Score calibration: |reported_confidence - actual_correctness|.
Lower is better. 0 = perfectly calibrated.
"""
metadata = result.get("metadata", {})
reported_confidence = metadata.get("coherence", 0.5)
# For now, use actual correctness will be measured separately
# Placeholder: assume 0.1 average calibration error
return 0.1
def _measure_convergence(self, result: Dict) -> float:
"""
Measure semantic convergence between adapter outputs (0-1).
0 = all different, 1 = all identical. Danger zone: >0.85
"""
metadata = result.get("metadata", {})
# Check specialization tracker output
spec_metrics = metadata.get("specialization_metrics", {})
convergence_alerts = spec_metrics.get("convergence_alerts", [])
if not convergence_alerts:
return 0.5 # Neutral baseline
# Take max similarity from recent alerts
max_similarity = 0.0
for alert in convergence_alerts:
if isinstance(alert, dict):
max_sim = alert.get("max_similarity", 0.0)
max_similarity = max(max_similarity, max_sim)
return min(1.0, max_similarity)
def export_results(self, filepath: str) -> None:
"""Export results to JSON for analysis."""
export_dict = {}
for condition, results in self.results.items():
export_dict[condition] = [self._serialize_result(asdict(r)) for r in results]
with open(filepath, 'w') as f:
json.dump(export_dict, f, indent=2, default=str)
print(f"\nResults exported to {filepath}")
def _serialize_result(self, result_dict: Dict) -> Dict:
"""Convert enums and non-serializable objects to strings for JSON."""
cleaned = {}
for key, value in result_dict.items():
if key == 'metadata' and isinstance(value, dict):
# Convert enum values in metadata to strings
cleaned[key] = {
k: str(v) if hasattr(v, 'name') else v
for k, v in value.items()
}
else:
cleaned[key] = value
return cleaned
# ============================================================================
# STATISTICAL ANALYSIS
# ============================================================================
class EvaluationAnalyzer:
"""Analyze evaluation results for statistical significance and insights."""
def __init__(self, results: Dict[str, List[EvaluationResult]]):
self.results = results
def summary_statistics(self) -> Dict:
"""Compute mean/std for each condition across metrics."""
summary = {}
for condition, result_list in self.results.items():
if not result_list:
continue
correctness_scores = [r.correctness_score for r in result_list]
reasoning_depths = [r.reasoning_depth for r in result_list]
calibration_errors = [r.calibration_error for r in result_list]
gamma_scores = [r.gamma_score for r in result_list]
convergences = [r.adapter_convergence for r in result_list]
summary[condition] = {
"correctness": {
"mean": sum(correctness_scores) / len(correctness_scores),
"std": self._std(correctness_scores),
},
"reasoning_depth": {
"mean": sum(reasoning_depths) / len(reasoning_depths),
"std": self._std(reasoning_depths),
},
"calibration_error": {
"mean": sum(calibration_errors) / len(calibration_errors),
"std": self._std(calibration_errors),
},
"gamma_score": {
"mean": sum(gamma_scores) / len(gamma_scores),
"std": self._std(gamma_scores),
},
"adapter_convergence": {
"mean": sum(convergences) / len(convergences),
"std": self._std(convergences),
},
}
return summary
def emergent_behavior_check(self) -> Dict:
"""
Check for pathological behaviors:
- High Γ (coherence) but low accuracy
- Increasing adapter convergence over time
- Miscalibration (high confidence, low correctness)
"""
alerts = {
"false_consensus": [],
"convergence_drift": [],
"miscalibration": [],
}
for condition, result_list in self.results.items():
for result in result_list:
# Alert 1: False consensus
if result.gamma_score > 0.8 and result.correctness_score < 0.5:
alerts["false_consensus"].append({
"condition": condition,
"query": result.query[:60],
"gamma": result.gamma_score,
"correctness": result.correctness_score,
})
# Alert 2: Over-convergence
if result.adapter_convergence > 0.85:
alerts["convergence_drift"].append({
"condition": condition,
"query": result.query[:60],
"convergence": result.adapter_convergence,
})
# Alert 3: Miscalibration
reported_conf = result.metadata.get("coherence", 0.5)
if reported_conf > 0.8 and result.correctness_score < 0.5:
alerts["miscalibration"].append({
"condition": condition,
"query": result.query[:60],
"reported_confidence": reported_conf,
"actual_correctness": result.correctness_score,
})
return alerts
def _std(self, values: List[float]) -> float:
"""Compute standard deviation."""
if len(values) < 2:
return 0.0
mean = sum(values) / len(values)
variance = sum((x - mean) ** 2 for x in values) / len(values)
return variance ** 0.5
def report(self) -> str:
"""Generate human-readable evaluation report."""
stats = self.summary_statistics()
alerts = self.emergent_behavior_check()
report = "\n" + "=" * 80 + "\n"
report += "CODETTE PHASE 6 EVALUATION REPORT\n"
report += "=" * 80 + "\n\n"
report += "SUMMARY STATISTICS\n"
report += "-" * 80 + "\n"
for condition, metrics in stats.items():
report += f"\n{condition}:\n"
for metric, values in metrics.items():
report += f" {metric}: {values['mean']:.3f} ± {values['std']:.3f}\n"
report += "\n\n" + "=" * 80 + "\n"
report += "EMERGENT BEHAVIOR ALERTS\n"
report += "-" * 80 + "\n"
report += f"\nFalse Consensus (High Γ, Low Accuracy): {len(alerts['false_consensus'])} cases\n"
for alert in alerts["false_consensus"][:3]:
report += f" - {alert['query']}: Γ={alert['gamma']:.2f}, Correctness={alert['correctness']:.2f}\n"
report += f"\nAdapter Convergence (>0.85): {len(alerts['convergence_drift'])} cases\n"
for alert in alerts["convergence_drift"][:3]:
report += f" - {alert['query']}: {alert['convergence']:.2f}\n"
report += f"\nMiscalibration: {len(alerts['miscalibration'])} cases\n"
for alert in alerts["miscalibration"][:3]:
report += f" - {alert['query']}: Reported={alert['reported_confidence']:.2f}, Actual={alert['actual_correctness']:.2f}\n"
report += "\n" + "=" * 80 + "\n"
return report
if __name__ == "__main__":
print("Evaluation suite loaded. Use with ForgeEngine:")
print(" harness = EvaluationHarness(forge)")
print(" results = harness.run_evaluation_suite()")
print(" analyzer = EvaluationAnalyzer(results)")
print(" print(analyzer.report())")
|