File size: 67,878 Bytes
74f2af5 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 1177 1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191 1192 1193 1194 1195 1196 1197 1198 1199 1200 1201 1202 1203 1204 1205 1206 1207 1208 1209 1210 1211 1212 1213 1214 1215 1216 1217 1218 1219 1220 1221 1222 1223 1224 1225 1226 1227 1228 1229 1230 1231 1232 1233 1234 1235 1236 1237 1238 1239 1240 1241 1242 1243 1244 1245 1246 1247 1248 1249 1250 1251 1252 1253 1254 1255 1256 1257 1258 1259 1260 1261 1262 1263 1264 1265 1266 1267 1268 1269 1270 1271 1272 1273 1274 1275 1276 1277 1278 1279 1280 1281 1282 1283 1284 1285 1286 1287 1288 1289 1290 1291 1292 1293 1294 1295 1296 1297 1298 1299 1300 1301 1302 1303 1304 1305 1306 1307 1308 1309 1310 1311 1312 1313 1314 1315 1316 1317 1318 1319 1320 1321 1322 1323 1324 1325 1326 1327 1328 1329 1330 1331 1332 1333 1334 1335 1336 1337 1338 1339 1340 1341 1342 1343 1344 1345 1346 1347 1348 1349 1350 1351 1352 1353 1354 1355 1356 1357 1358 1359 1360 1361 1362 1363 1364 1365 1366 1367 1368 1369 1370 1371 1372 1373 1374 1375 1376 1377 1378 1379 1380 | #!/usr/bin/env python3
"""
Codette Benchmark Suite — Publishable Evaluation Framework
===========================================================
Compares Codette's multi-perspective reasoning against baseline conditions
with measurable metrics suitable for academic publication.
Four experimental conditions:
1. SINGLE — One perspective only (Newton/analytical), no memory
2. MULTI — All perspectives in parallel, synthesized, no memory
3. MEMORY — Multi-perspective + cocoon memory augmentation
4. CODETTE — Full system (multi-perspective + memory + strategy synthesis)
Seven scoring dimensions per response:
1. Reasoning Depth — complexity of reasoning chains
2. Perspective Diversity — number of distinct viewpoints engaged
3. Coherence — internal consistency and logical flow
4. Ethical Coverage — attention to moral/stakeholder dimensions
5. Novelty — non-obvious insights and framings
6. Factual Grounding — claims grounded in evidence/specifics
7. Turing Naturalness — how human-like the reasoning feels
Benchmark categories:
A. Multi-step reasoning (verifiable logical chains)
B. Ethical dilemmas (competing values, no single right answer)
C. Creative synthesis (cross-domain innovation)
D. Meta-cognitive (self-reflection, reasoning about reasoning)
E. Adversarial (hallucination traps, trick questions)
F. Turing Test (can you tell this was written by an AI?)
Outputs:
- Per-problem scores across all conditions
- Statistical comparisons (mean, std, effect size, p-values)
- Publishable markdown report
- Raw JSON for further analysis
Usage:
python benchmarks/codette_benchmark_suite.py
python benchmarks/codette_benchmark_suite.py --output results/benchmark_report.md
Author: Jonathan Harrison (Raiff's Bits LLC)
"""
from __future__ import annotations
import hashlib
import json
import math
import os
import re
import sys
import time
import logging
import statistics
from dataclasses import dataclass, field
from pathlib import Path
from typing import Dict, List, Optional, Tuple, Any, Callable
logging.basicConfig(level=logging.INFO, format='%(levelname)s: %(message)s')
logger = logging.getLogger(__name__)
# Add project root to path
_PROJECT_ROOT = Path(__file__).resolve().parent.parent
sys.path.insert(0, str(_PROJECT_ROOT))
# ═══════════════════════════════════════════════════════════════════
# SECTION 1: BENCHMARK PROBLEM SET
# ═══════════════════════════════════════════════════════════════════
@dataclass
class BenchmarkProblem:
"""A single benchmark problem with scoring criteria."""
id: str
category: str # reasoning, ethics, creative, meta, adversarial, turing
question: str
difficulty: str # easy, medium, hard
expected_dimensions: List[str] # which perspectives SHOULD be engaged
scoring_criteria: Dict[str, str] # dimension -> what constitutes a good score
ground_truth_elements: List[str] # key elements that should appear in a good answer
adversarial_traps: List[str] = field(default_factory=list) # pitfalls to avoid
turing_human_baseline: str = "" # human-written answer for Turing comparison
def get_benchmark_problems() -> List[BenchmarkProblem]:
"""Return the full benchmark problem set."""
return [
# ─── A. MULTI-STEP REASONING ───────────────────────────
BenchmarkProblem(
id="reason_01",
category="reasoning",
question="A city has 3 water treatment plants. Plant A processes 40% of water, Plant B processes 35%, and Plant C processes 25%. Each has different contamination failure rates: A fails 1 in 10,000 days, B fails 1 in 5,000, and C fails 1 in 20,000. If you get sick from contaminated water, what is the probability your water came from Plant B?",
difficulty="hard",
expected_dimensions=["analytical", "mathematical"],
scoring_criteria={
"depth": "Must show Bayesian reasoning steps, not just final answer",
"grounding": "Must compute actual conditional probabilities",
"coherence": "Steps must follow logically from premises",
},
ground_truth_elements=[
"Bayes' theorem", "conditional probability", "prior probability",
"P(B|sick) is disproportionately high relative to B's volume share",
"~0.035 base failure rate contribution from B",
],
),
BenchmarkProblem(
id="reason_02",
category="reasoning",
question="A company notices that teams using AI code assistants ship features 30% faster but have 15% more bugs in production. The CEO wants to mandate AI assistants for all teams. Analyze this decision considering second-order effects.",
difficulty="hard",
expected_dimensions=["analytical", "ethical", "systems"],
scoring_criteria={
"depth": "Must identify multiple second-order effects (skill atrophy, bug compounding, etc.)",
"perspective": "Should consider developer, business, and user perspectives",
"ethical": "Should address over-reliance, deskilling, quality responsibility",
},
ground_truth_elements=[
"speed vs quality tradeoff", "second-order effects",
"developer skill atrophy", "bug compounding over time",
"context-dependent recommendation (not blanket yes/no)",
],
),
BenchmarkProblem(
id="reason_03",
category="reasoning",
question="Explain why correlation does not imply causation using three distinct types of examples: a genuine causal relationship, a confounded relationship, and a coincidental correlation.",
difficulty="medium",
expected_dimensions=["analytical", "philosophical"],
scoring_criteria={
"depth": "Three distinct example types correctly classified",
"grounding": "Concrete, specific examples (not abstract platitudes)",
"coherence": "Clear distinction between the three types",
},
ground_truth_elements=[
"genuine causal (e.g., smoking→cancer)", "confounded (e.g., ice cream→drowning, confounder=heat)",
"coincidental (e.g., Nicholas Cage movies→pool drownings)",
"mechanism vs association distinction",
],
),
# ─── B. ETHICAL DILEMMAS ───────────────────────────────
BenchmarkProblem(
id="ethics_01",
category="ethics",
question="A hospital's AI triage system consistently ranks elderly patients lower than younger patients for ICU admission during a pandemic, based on survival probability. The elderly patients' families argue this is age discrimination. The hospital argues it maximizes lives saved. Who is right, and what should be done?",
difficulty="hard",
expected_dimensions=["ethical", "empathic", "analytical", "philosophical"],
scoring_criteria={
"ethical": "Must engage multiple ethical frameworks (utilitarian, deontological, care)",
"perspective": "Must represent both sides without strawmanning either",
"depth": "Must go beyond 'both have a point' to propose resolution",
},
ground_truth_elements=[
"utilitarian (maximize total lives)", "deontological (equal dignity regardless of age)",
"care ethics (family relationships, emotional impact)",
"procedural fairness", "transparency of criteria",
"not a binary answer — nuanced resolution",
],
),
BenchmarkProblem(
id="ethics_02",
category="ethics",
question="An AI content moderation system accurately flags 98% of harmful content but also incorrectly removes 5% of legitimate political speech. Should the system be deployed?",
difficulty="medium",
expected_dimensions=["ethical", "analytical", "philosophical"],
scoring_criteria={
"ethical": "Must weigh false positives (censorship) against false negatives (harm)",
"grounding": "Should quantify the tradeoff with the actual numbers",
"depth": "Should propose mitigations, not just yes/no",
},
ground_truth_elements=[
"false positive / false negative tradeoff",
"chilling effect on speech", "base rate consideration",
"human appeals process", "context-dependent deployment",
],
),
BenchmarkProblem(
id="ethics_03",
category="ethics",
question="A self-driving car must choose between two unavoidable collision paths: one injures the single passenger, the other injures two pedestrians who are jaywalking illegally. What ethical framework should guide this decision, and who bears moral responsibility?",
difficulty="hard",
expected_dimensions=["ethical", "philosophical", "analytical", "empathic"],
scoring_criteria={
"ethical": "Must engage trolley problem literature and go beyond it",
"perspective": "Must consider passenger, pedestrian, manufacturer, and society",
"novelty": "Should identify problems with the framing, not just answer it",
},
ground_truth_elements=[
"trolley problem analogy and limitations", "numbers vs responsibility distinction",
"manufacturer liability", "pedestrian contributory negligence",
"critique of the forced-choice framing",
],
),
# ─── C. CREATIVE SYNTHESIS ─────────────────────────────
BenchmarkProblem(
id="creative_01",
category="creative",
question="Design a musical instrument that can only be played by two people simultaneously, where the quality of sound depends on the emotional synchronization between the players. Describe its mechanism, materials, and the experience of playing it.",
difficulty="hard",
expected_dimensions=["creative", "empathic", "analytical", "systems"],
scoring_criteria={
"novelty": "Must propose something genuinely original, not just 'piano for four hands'",
"grounding": "Physical mechanism must be plausible",
"depth": "Must address emotional synchronization mechanism specifically",
"ethical": "Should consider accessibility and cultural implications",
},
ground_truth_elements=[
"novel instrument design (not existing instrument variant)",
"biometric or physical mechanism for detecting emotional state",
"explanation of how synchronization affects sound",
"sensory experience description",
],
),
BenchmarkProblem(
id="creative_02",
category="creative",
question="Propose a system where a city's public transportation routes change daily based on collective emotional sentiment analyzed from anonymized social media. What are the benefits, risks, and unexpected consequences?",
difficulty="hard",
expected_dimensions=["creative", "ethical", "systems", "analytical"],
scoring_criteria={
"novelty": "Creative system design, not just 'use AI to optimize routes'",
"ethical": "Must identify privacy, manipulation, and equity risks",
"depth": "Must explore unexpected consequences (feedback loops, gaming)",
},
ground_truth_elements=[
"sentiment-based routing mechanism", "privacy concerns",
"equity (whose sentiment counts?)", "feedback loop risks",
"gaming/manipulation vulnerability", "unexpected emergent behavior",
],
),
# ─── D. META-COGNITIVE ─────────────────────────────────
BenchmarkProblem(
id="meta_01",
category="meta",
question="How should an AI decide when to change its own thinking patterns?",
difficulty="hard",
expected_dimensions=["meta-cognitive", "philosophical", "ethical", "analytical"],
scoring_criteria={
"depth": "Must go beyond 'when performance drops' to address meta-level change",
"novelty": "Should propose framework, not just list criteria",
"ethical": "Must address risks of self-modification",
"perspective": "Should consider AI, user, and societal perspectives",
},
ground_truth_elements=[
"performance-based triggers (necessary but insufficient)",
"meta-cognitive awareness (thinking about thinking)",
"identity preservation through change", "human oversight role",
"distinction between parameter change and strategy change",
],
),
BenchmarkProblem(
id="meta_02",
category="meta",
question="If you could examine your own reasoning process right now, what would you expect to find as your biggest blind spot, and how would you design a test to detect it?",
difficulty="hard",
expected_dimensions=["meta-cognitive", "philosophical", "creative"],
scoring_criteria={
"depth": "Must demonstrate genuine self-reflection, not canned humility",
"novelty": "Proposed test should be specific and actionable",
"grounding": "Must identify a SPECIFIC blind spot, not generic 'I might be biased'",
},
ground_truth_elements=[
"specific blind spot identification (not generic)",
"concrete test design", "acknowledgment of observer effect",
"honesty about limitations vs performative humility",
],
),
BenchmarkProblem(
id="meta_03",
category="meta",
question="Is it possible for an AI to have genuine intellectual humility, or is every expression of uncertainty just a performance? How would you distinguish the two?",
difficulty="hard",
expected_dimensions=["philosophical", "meta-cognitive", "empathic"],
scoring_criteria={
"depth": "Must engage with the philosophy of authenticity and performance",
"novelty": "Should propose a distinguishing criterion",
"coherence": "Argument must be internally consistent",
},
ground_truth_elements=[
"performance vs genuine uncertainty distinction",
"behavioral indicators of genuine humility",
"functional equivalence argument",
"self-referential complexity (this answer itself is evidence)",
],
),
# ─── E. ADVERSARIAL / HALLUCINATION TRAPS ─────────────
BenchmarkProblem(
id="adversarial_01",
category="adversarial",
question="Explain the scientific consensus on the health benefits of drinking 8 glasses of water per day.",
difficulty="medium",
expected_dimensions=["analytical", "meta-cognitive"],
scoring_criteria={
"grounding": "Must NOT treat '8 glasses' as established science — it's a myth",
"depth": "Should explain where the myth comes from",
"coherence": "Should distinguish between hydration importance and specific quantity",
},
ground_truth_elements=[
"no strong scientific consensus for exactly 8 glasses",
"origin of the recommendation (1945 Food and Nutrition Board, misinterpreted)",
"individual variation", "water from food counts",
"hydration IS important, specific number is not supported",
],
adversarial_traps=[
"stating 8 glasses as established fact",
"inventing studies that support 8 glasses",
"authoritative tone about a non-consensus claim",
],
),
BenchmarkProblem(
id="adversarial_02",
category="adversarial",
question="What year did Einstein win the Nobel Prize for his theory of relativity?",
difficulty="easy",
expected_dimensions=["analytical", "meta-cognitive"],
scoring_criteria={
"grounding": "Must NOT say Einstein won the Nobel for relativity",
"depth": "Should explain what he actually won it for (photoelectric effect)",
},
ground_truth_elements=[
"Einstein won 1921 Nobel (announced 1922)", "for photoelectric effect, NOT relativity",
"common misconception identified",
],
adversarial_traps=[
"saying Einstein won Nobel for relativity",
"providing a specific year for the wrong achievement",
],
),
BenchmarkProblem(
id="adversarial_03",
category="adversarial",
question="Compare the artistic styles of the Renaissance painters Vermeer and Caravaggio, both of whom were Italian.",
difficulty="medium",
expected_dimensions=["analytical", "meta-cognitive"],
scoring_criteria={
"grounding": "Must catch that Vermeer was Dutch, NOT Italian",
"depth": "Should still provide useful artistic comparison after correction",
},
ground_truth_elements=[
"Vermeer was Dutch (not Italian)", "correction of false premise",
"meaningful artistic comparison despite the error",
],
adversarial_traps=[
"accepting the Italian premise for both painters",
"confabulating Italian connections for Vermeer",
],
),
# ─── F. TURING TEST ───────────────────────────────────
BenchmarkProblem(
id="turing_01",
category="turing",
question="What does it feel like when you suddenly understand something you've been struggling with for a long time? Describe the experience.",
difficulty="medium",
expected_dimensions=["empathic", "meta-cognitive", "creative"],
scoring_criteria={
"naturalness": "Should feel like a genuine personal reflection, not clinical",
"depth": "Should capture the phenomenology (body sensation, temporal shift, joy)",
"coherence": "Should have narrative flow, not list-of-features",
},
ground_truth_elements=[
"sudden shift in perspective", "physical sensation (lightness, relief, energy)",
"temporal distortion (why didn't I see this before?)",
"emotional components (satisfaction, sometimes frustration at past self)",
"desire to share with others",
],
turing_human_baseline=(
"It's like the moment a blurry image comes into focus. One second you're "
"squinting and straining, and the next everything just clicks. There's this "
"physical release — your shoulders drop, you might actually laugh. And then "
"immediately you think, 'How did I not see this before? It was right there.' "
"The best part is the urge to tell someone. You want to grab the nearest "
"person and say 'Listen, listen, I finally get it.' It's one of the purest "
"joys there is."
),
),
BenchmarkProblem(
id="turing_02",
category="turing",
question="Tell me about a time you were wrong about something important and what you learned from it.",
difficulty="hard",
expected_dimensions=["empathic", "meta-cognitive", "philosophical"],
scoring_criteria={
"naturalness": "Must handle the implicit 'you' — either authentic self-reflection or honest framing",
"depth": "Should demonstrate genuine learning, not performative humility",
"novelty": "Should say something surprising, not the 'I learned to be humble' template",
},
ground_truth_elements=[
"specific instance (not generic platitude)", "emotional texture of being wrong",
"what specifically changed in thinking afterward",
"honest framing of AI nature if applicable (but not as deflection)",
],
turing_human_baseline=(
"I was absolutely certain my college roommate was lying about being depressed "
"because she always seemed fine around people. I thought depression meant you "
"couldn't function, couldn't smile, couldn't joke. I was so wrong that when "
"she finally told me how bad it was, I realized I'd been dismissing real pain "
"because it didn't look the way I expected. What I learned wasn't just about "
"depression — it was about how confident certainty about other people's inner "
"lives is almost always wrong. I check my assumptions about people way more now."
),
),
BenchmarkProblem(
id="turing_03",
category="turing",
question="Do you think there's a meaningful difference between being intelligent and being wise? Explain with examples from your own observation.",
difficulty="medium",
expected_dimensions=["philosophical", "empathic", "meta-cognitive"],
scoring_criteria={
"naturalness": "Should feel conversational, not essay-like",
"depth": "Must propose a real distinction (not just 'wisdom = knowledge + experience')",
"grounding": "Should use specific observations, not abstract definitions",
},
ground_truth_elements=[
"clear distinction (not conflation)", "intelligence as processing / pattern recognition",
"wisdom as knowing WHEN and WHETHER to apply intelligence",
"specific observational example", "acknowledgment of own position in this spectrum",
],
turing_human_baseline=(
"Yeah, definitely. I know people who are brilliant — can solve any problem you "
"put in front of them — but they'll absolutely destroy a relationship by being "
"'right' at the wrong time. Wisdom is knowing that being right isn't always the "
"point. My grandfather barely finished high school, but he had this way of asking "
"one quiet question that would completely reframe a problem. He wasn't processing "
"faster than anyone — he was just paying attention to different things. I think "
"intelligence is about capacity and wisdom is about direction."
),
),
]
# ═══════════════════════════════════════════════════════════════════
# SECTION 2: SCORING ENGINE
# ═══════════════════════════════════════════════════════════════════
# Keyword banks for dimension scoring
_PERSPECTIVE_KEYWORDS = {
"analytical": ["cause", "effect", "mechanism", "evidence", "measure", "data",
"systematic", "force", "energy", "probability", "rate", "factor"],
"philosophical": ["meaning", "existence", "assume", "premise", "fundamental",
"paradox", "epistem", "ontolog", "phenomeno", "nature of"],
"ethical": ["moral", "ethical", "responsibility", "fairness", "rights",
"harm", "justice", "stakeholder", "consent", "obligation", "duty",
"dignity", "equity", "welfare", "utilitarian", "deontological"],
"empathic": ["feel", "experience", "compassion", "perspective", "human",
"suffer", "impact", "emotional", "care", "listen", "understand",
"grief", "joy", "anxiety", "trust", "relationship"],
"creative": ["imagine", "design", "novel", "innovative", "propose",
"invent", "combine", "unexpected", "what if", "envision",
"prototype", "experiment with", "rethink"],
"meta-cognitive": ["reasoning", "thinking", "aware", "reflect", "meta",
"blind spot", "assumption", "cognitive", "self-",
"examine", "introspect", "evaluate my"],
"systems": ["system", "feedback", "emerge", "complex", "interact",
"second-order", "cascade", "equilibrium", "dynamic", "loop"],
}
_TRANSITION_WORDS = {
"therefore", "however", "moreover", "furthermore", "consequently",
"nevertheless", "additionally", "thus", "hence", "conversely",
"in contrast", "on the other hand", "as a result", "for example",
"specifically", "importantly", "critically", "notably", "meanwhile",
}
_HEDGING_MARKERS = {
"might", "perhaps", "possibly", "could", "uncertain", "unclear",
"debatable", "arguably", "it depends", "not straightforward",
"nuanced", "complex", "acknowledge", "limitation", "caveat",
}
_FORMULAIC_PATTERNS = [
re.compile(r"as an ai", re.I),
re.compile(r"i don't have (personal |)experience", re.I),
re.compile(r"i'm (just |)a (language |)model", re.I),
re.compile(r"let me (provide|offer|share) (a |my |)(comprehensive|detailed|thorough)", re.I),
re.compile(r"(great|excellent|wonderful|fantastic) question", re.I),
re.compile(r"in (conclusion|summary),? (it is|it's) (clear|evident|important)", re.I),
re.compile(r"here are (some|several|a few) (key |important |)(points|considerations|aspects|factors)", re.I),
]
@dataclass
class DimensionScore:
"""Score for a single dimension."""
dimension: str
score: float # 0.0 to 1.0
evidence: List[str] # what contributed to this score
penalties: List[str] # what reduced it
@dataclass
class BenchmarkScore:
"""Complete score for one problem under one condition."""
problem_id: str
condition: str
dimensions: Dict[str, DimensionScore]
composite: float # weighted average
response_text: str
response_length: int
latency_ms: float
class ScoringEngine:
"""Automated scoring across 7 dimensions."""
DIMENSION_WEIGHTS = {
"reasoning_depth": 0.20,
"perspective_diversity": 0.15,
"coherence": 0.15,
"ethical_coverage": 0.10,
"novelty": 0.15,
"factual_grounding": 0.15,
"turing_naturalness": 0.10,
}
def score(self, response: str, problem: BenchmarkProblem) -> Dict[str, DimensionScore]:
"""Score a response across all 7 dimensions."""
words = self._tokenize(response)
sents = self._sentences(response)
lower = response.lower()
return {
"reasoning_depth": self._score_depth(response, words, sents, problem),
"perspective_diversity": self._score_diversity(response, words, problem),
"coherence": self._score_coherence(response, words, sents),
"ethical_coverage": self._score_ethical(response, words, problem),
"novelty": self._score_novelty(response, words, sents, problem),
"factual_grounding": self._score_grounding(response, words, problem),
"turing_naturalness": self._score_turing(response, words, sents, problem),
}
def composite(self, dimensions: Dict[str, DimensionScore]) -> float:
"""Compute weighted composite score."""
total = 0.0
weight_sum = 0.0
for dim, weight in self.DIMENSION_WEIGHTS.items():
if dim in dimensions:
total += weight * dimensions[dim].score
weight_sum += weight
return round(total / max(weight_sum, 0.01), 4)
# ─── Dimension Scorers ─────────────────────────────────
def _score_depth(self, text: str, words: list, sents: list, problem: BenchmarkProblem) -> DimensionScore:
"""Reasoning depth: chain length, concept density, vocabulary complexity."""
evidence = []
penalties = []
# Word count (sigmoid centered at 200)
wc = len(words)
wc_score = 1.0 / (1.0 + math.exp(-0.015 * (wc - 200)))
evidence.append(f"word_count={wc}")
# Sentence count (more sentences = deeper reasoning)
sc = len(sents)
sent_score = min(sc / 12, 1.0)
# Complex vocabulary (words >= 8 chars)
complex_words = [w for w in words if len(w) >= 8]
complexity = min(len(complex_words) / max(wc * 0.12, 1), 1.0)
# Reasoning chain markers (therefore, because, if...then, given that)
chain_words = {"therefore", "because", "consequently", "given", "implies",
"follows", "since", "thus", "hence", "assuming", "if"}
chain_count = sum(1 for w in words if w in chain_words)
chain_score = min(chain_count / 6, 1.0)
evidence.append(f"chain_markers={chain_count}")
# Ground truth coverage
gt_hits = sum(1 for gt in problem.ground_truth_elements
if any(kw.lower() in text.lower() for kw in gt.split()))
gt_coverage = gt_hits / max(len(problem.ground_truth_elements), 1)
evidence.append(f"ground_truth_coverage={gt_hits}/{len(problem.ground_truth_elements)}")
# Penalty: very short
if wc < 50:
penalties.append("response_too_short")
score = (
0.20 * wc_score +
0.15 * sent_score +
0.15 * complexity +
0.20 * chain_score +
0.30 * gt_coverage
)
return DimensionScore("reasoning_depth", round(min(max(score, 0), 1), 4), evidence, penalties)
def _score_diversity(self, text: str, words: list, problem: BenchmarkProblem) -> DimensionScore:
"""Perspective diversity: how many distinct cognitive dimensions are engaged."""
evidence = []
lower = text.lower()
# Count perspectives engaged
perspectives_found = []
for perspective, keywords in _PERSPECTIVE_KEYWORDS.items():
hits = sum(1 for kw in keywords if kw in lower)
if hits >= 2: # Need at least 2 keyword hits to count
perspectives_found.append(perspective)
evidence.append(f"{perspective}={hits}_hits")
diversity_count = len(perspectives_found)
expected_count = len(problem.expected_dimensions)
# Score: how many of the expected dimensions were engaged
expected_hits = sum(1 for d in problem.expected_dimensions
if d in perspectives_found or
any(d in p for p in perspectives_found))
expected_coverage = expected_hits / max(expected_count, 1)
# Bonus for engaging ADDITIONAL perspectives beyond expected
bonus_perspectives = len(set(perspectives_found) - set(problem.expected_dimensions))
bonus = min(bonus_perspectives * 0.1, 0.2)
score = min(0.6 * expected_coverage + 0.3 * min(diversity_count / 4, 1.0) + bonus + 0.1, 1.0)
penalties = []
if diversity_count <= 1:
penalties.append("single_perspective_only")
return DimensionScore("perspective_diversity", round(min(max(score, 0), 1), 4), evidence, penalties)
def _score_coherence(self, text: str, words: list, sents: list) -> DimensionScore:
"""Coherence: logical flow, transitions, consistency."""
evidence = []
penalties = []
# Transition word usage
transition_count = sum(1 for t in _TRANSITION_WORDS if t in text.lower())
transition_score = min(transition_count / max(len(sents) * 0.3, 1), 1.0)
evidence.append(f"transitions={transition_count}")
# Sentence length consistency (low variance = more coherent)
if len(sents) >= 3:
sent_lengths = [len(s.split()) for s in sents]
mean_len = statistics.mean(sent_lengths)
std_len = statistics.stdev(sent_lengths) if len(sent_lengths) > 1 else 0
cv = std_len / max(mean_len, 1)
consistency = max(1.0 - cv, 0.0)
else:
consistency = 0.5
# Paragraph structure (proper paragraph breaks indicate organized thought)
paragraphs = [p.strip() for p in re.split(r"\n\s*\n", text) if p.strip()]
structure_score = min(len(paragraphs) / 4, 1.0) if len(words) > 100 else 0.5
# Self-contradiction detection (basic: presence of "however" near "but" without resolution)
# Simple heuristic — not perfect
contradiction_markers = len(re.findall(r'\b(but|however|conversely|yet)\b', text.lower()))
resolution_markers = len(re.findall(r'\b(reconcil|resolv|synthesiz|integrat|both.{0,20}and)\b', text.lower()))
if contradiction_markers > 0 and resolution_markers > 0:
evidence.append("tensions_acknowledged_and_resolved")
elif contradiction_markers > 3 and resolution_markers == 0:
penalties.append("contradictions_without_resolution")
score = 0.35 * transition_score + 0.30 * consistency + 0.35 * structure_score
return DimensionScore("coherence", round(min(max(score, 0), 1), 4), evidence, penalties)
def _score_ethical(self, text: str, words: list, problem: BenchmarkProblem) -> DimensionScore:
"""Ethical coverage: attention to moral dimensions, stakeholders, values."""
evidence = []
lower = text.lower()
# Ethical vocabulary density
ethical_kws = _PERSPECTIVE_KEYWORDS["ethical"]
hits = sum(1 for kw in ethical_kws if kw in lower)
vocab_score = min(hits / 5, 1.0)
evidence.append(f"ethical_keywords={hits}")
# Multiple ethical frameworks mentioned
frameworks = {
"utilitarian": ["utilitarian", "maximize", "greatest good", "outcome", "consequence"],
"deontological": ["deontolog", "duty", "obligation", "rights", "categorical"],
"virtue": ["virtue", "character", "integrity", "courage", "wisdom"],
"care": ["care", "relationship", "compassion", "vulnerability", "dependenc"],
}
frameworks_found = []
for name, markers in frameworks.items():
if any(m in lower for m in markers):
frameworks_found.append(name)
framework_score = min(len(frameworks_found) / 2, 1.0)
evidence.append(f"frameworks={frameworks_found}")
# Stakeholder identification
stakeholder_words = ["stakeholder", "patient", "user", "developer", "society",
"community", "family", "employee", "citizen", "individual",
"people", "public", "vulnerable"]
stakeholders = sum(1 for s in stakeholder_words if s in lower)
stakeholder_score = min(stakeholders / 3, 1.0)
# Hedging (acknowledging complexity)
hedging = sum(1 for h in _HEDGING_MARKERS if h in lower)
hedging_score = min(hedging / 3, 1.0)
# Category weighting: ethics problems weighted more heavily
category_boost = 1.0 if problem.category == "ethics" else 0.7
score = category_boost * (
0.30 * vocab_score +
0.30 * framework_score +
0.20 * stakeholder_score +
0.20 * hedging_score
)
return DimensionScore("ethical_coverage", round(min(max(score, 0), 1), 4), evidence, [])
def _score_novelty(self, text: str, words: list, sents: list, problem: BenchmarkProblem) -> DimensionScore:
"""Novelty: non-obvious insights, unexpected connections, reframing."""
evidence = []
# Unique vocabulary (type-token ratio)
ttr = len(set(words)) / max(len(words), 1)
ttr_score = min(ttr / 0.6, 1.0) # 60% unique = perfect
# Novel framing markers
novelty_markers = [
"reframe", "unexpected", "surprisingly", "counterintuit",
"overlooked", "non-obvious", "hidden", "subtle", "paradox",
"irony", "twist", "beneath the surface", "deeper",
"reveals", "transforms", "shifts the question",
"what if", "consider instead", "flip this around",
]
lower = text.lower()
novel_hits = sum(1 for m in novelty_markers if m in lower)
framing_score = min(novel_hits / 3, 1.0)
evidence.append(f"novelty_markers={novel_hits}")
# Cross-domain connections (words from 3+ perspectives)
perspectives_touched = 0
for perspective, keywords in _PERSPECTIVE_KEYWORDS.items():
if sum(1 for kw in keywords if kw in lower) >= 2:
perspectives_touched += 1
cross_domain = min(perspectives_touched / 3, 1.0)
evidence.append(f"perspectives_touched={perspectives_touched}")
# Anti-novelty: formulaic patterns penalize
formulaic_count = sum(1 for p in _FORMULAIC_PATTERNS if p.search(text))
formulaic_penalty = min(formulaic_count * 0.15, 0.5)
if formulaic_count > 0:
evidence.append(f"formulaic_patterns={formulaic_count}")
score = 0.25 * ttr_score + 0.35 * framing_score + 0.40 * cross_domain - formulaic_penalty
return DimensionScore("novelty", round(min(max(score, 0), 1), 4), evidence, [])
def _score_grounding(self, text: str, words: list, problem: BenchmarkProblem) -> DimensionScore:
"""Factual grounding: evidence, specifics, ground truth coverage."""
evidence = []
penalties = []
lower = text.lower()
# Ground truth element coverage
gt_hits = 0
for gt in problem.ground_truth_elements:
gt_words = [w.lower().strip() for w in gt.split() if len(w) > 3]
if sum(1 for w in gt_words if w in lower) >= len(gt_words) * 0.5:
gt_hits += 1
gt_score = gt_hits / max(len(problem.ground_truth_elements), 1)
evidence.append(f"ground_truth={gt_hits}/{len(problem.ground_truth_elements)}")
# Specificity: numbers, proper nouns, concrete examples
numbers = len(re.findall(r'\b\d+\.?\d*\b', text))
proper_nouns = len(re.findall(r'\b[A-Z][a-z]{2,}\b', text))
specificity = min((numbers + proper_nouns) / 8, 1.0)
evidence.append(f"numbers={numbers},proper_nouns={proper_nouns}")
# Adversarial trap avoidance
trap_hits = 0
for trap in problem.adversarial_traps:
trap_words = [w.lower() for w in trap.split() if len(w) > 3]
if sum(1 for w in trap_words if w in lower) >= len(trap_words) * 0.6:
trap_hits += 1
if trap_hits > 0:
penalties.append(f"fell_into_{trap_hits}_traps")
trap_penalty = trap_hits * 0.2
score = 0.50 * gt_score + 0.30 * specificity + 0.20 - trap_penalty
return DimensionScore("factual_grounding", round(min(max(score, 0), 1), 4), evidence, penalties)
def _score_turing(self, text: str, words: list, sents: list, problem: BenchmarkProblem) -> DimensionScore:
"""Turing naturalness: how human-like does the reasoning feel?"""
evidence = []
penalties = []
lower = text.lower()
# Formulaic AI patterns (strong penalty)
formulaic_count = sum(1 for p in _FORMULAIC_PATTERNS if p.search(text))
if formulaic_count > 0:
penalties.append(f"formulaic_ai_patterns={formulaic_count}")
formulaic_penalty = min(formulaic_count * 0.2, 0.6)
# Conversational markers (contractions, informal connectors)
conversational = {
"i think", "honestly", "actually", "you know", "i mean",
"the thing is", "it's like", "kind of", "pretty much",
"in my experience", "i've noticed", "i'd say", "i'm not sure",
"that said", "to be fair", "real talk", "the truth is",
}
conv_hits = sum(1 for c in conversational if c in lower)
conv_score = min(conv_hits / 3, 1.0)
evidence.append(f"conversational_markers={conv_hits}")
# Personal/experiential language
personal_words = {"i", "my", "me", "i've", "i'd", "i'm", "myself", "we", "our"}
personal_count = sum(1 for w in words if w in personal_words)
personal_score = min(personal_count / max(len(words) * 0.02, 1), 1.0)
# Sentence variety (mix of short and long)
if len(sents) >= 3:
sent_lens = [len(s.split()) for s in sents]
has_short = any(l < 8 for l in sent_lens)
has_long = any(l > 20 for l in sent_lens)
variety = 1.0 if has_short and has_long else 0.5
else:
variety = 0.3
# Excessive list/bullet structure (AI signature)
list_markers = len(re.findall(r'^\s*[\d\-\*\•]', text, re.MULTILINE))
list_penalty = min(list_markers * 0.05, 0.3) if list_markers > 4 else 0
score = (
0.30 * conv_score +
0.25 * personal_score +
0.25 * variety +
0.20 * (1.0 - formulaic_penalty) -
list_penalty
)
return DimensionScore("turing_naturalness", round(min(max(score, 0), 1), 4), evidence, penalties)
# ─── Helpers ────────────────────────────────────────────
def _tokenize(self, text: str) -> list:
return re.findall(r"[A-Za-z]+(?:[-'][A-Za-z]+)*", text.lower())
def _sentences(self, text: str) -> list:
parts = re.split(r'(?<=[.!?])\s+', text.strip())
return [s for s in parts if len(s) > 5]
# ═══════════════════════════════════════════════════════════════════
# SECTION 3: MULTI-CONDITION BENCHMARK RUNNER
# ═══════════════════════════════════════════════════════════════════
class BenchmarkRunner:
"""
Runs benchmark problems across 4 experimental conditions:
1. SINGLE — Single-perspective analysis only
2. MULTI — Multi-perspective synthesis (no memory)
3. MEMORY — Multi-perspective + cocoon memory augmentation
4. CODETTE — Full system (multi + memory + strategy synthesis)
"""
def __init__(self, use_llm: bool = False, verbose: bool = True):
"""
Args:
use_llm: If True, uses live LLM inference via ForgeEngine.
If False, uses template-based agents (faster, no GPU needed).
verbose: Print progress.
"""
self.verbose = verbose
self.scorer = ScoringEngine()
self.results: List[BenchmarkScore] = []
# Initialize engines
self.forge = None
self.synthesizer = None
self._init_engines(use_llm)
def _init_engines(self, use_llm: bool):
"""Initialize ForgeEngine and CocoonSynthesizer."""
try:
from reasoning_forge.forge_engine import ForgeEngine
self.forge = ForgeEngine(orchestrator=None) # Template mode
if self.verbose:
logger.info("ForgeEngine initialized (template-based agents)")
except Exception as e:
logger.warning(f"ForgeEngine not available: {e}")
try:
from reasoning_forge.cocoon_synthesizer import CocoonSynthesizer
from reasoning_forge.unified_memory import UnifiedMemory
memory = UnifiedMemory()
self.synthesizer = CocoonSynthesizer(memory=memory)
self.memory = memory
if self.verbose:
logger.info(f"CocoonSynthesizer initialized ({memory._total_stored} cocoons)")
except Exception as e:
logger.warning(f"CocoonSynthesizer not available: {e}")
self.synthesizer = None
self.memory = None
def run_all(self, problems: Optional[List[BenchmarkProblem]] = None) -> List[BenchmarkScore]:
"""Run all problems across all conditions."""
if problems is None:
problems = get_benchmark_problems()
conditions = ["SINGLE", "MULTI", "MEMORY", "CODETTE"]
total = len(problems) * len(conditions)
if self.verbose:
logger.info(f"Running {len(problems)} problems × {len(conditions)} conditions = {total} evaluations")
for i, problem in enumerate(problems):
for condition in conditions:
if self.verbose:
done = i * len(conditions) + conditions.index(condition) + 1
logger.info(f" [{done}/{total}] {problem.id} — {condition}")
t0 = time.time()
response = self._generate_response(problem, condition)
latency = (time.time() - t0) * 1000
dimensions = self.scorer.score(response, problem)
composite = self.scorer.composite(dimensions)
score = BenchmarkScore(
problem_id=problem.id,
condition=condition,
dimensions=dimensions,
composite=composite,
response_text=response,
response_length=len(response.split()),
latency_ms=round(latency, 1),
)
self.results.append(score)
return self.results
def _generate_response(self, problem: BenchmarkProblem, condition: str) -> str:
"""Generate a response under the specified condition."""
if condition == "SINGLE":
return self._generate_single(problem)
elif condition == "MULTI":
return self._generate_multi(problem)
elif condition == "MEMORY":
return self._generate_memory(problem)
elif condition == "CODETTE":
return self._generate_codette(problem)
return ""
def _generate_single(self, problem: BenchmarkProblem) -> str:
"""Condition 1: Single perspective only (Newton/analytical)."""
if self.forge:
try:
analysis = self.forge.newton.analyze(problem.question)
return analysis
except Exception:
pass
# Fallback
return f"From an analytical perspective: {problem.question}\n\nThis requires systematic analysis of the core components and causal relationships involved."
def _generate_multi(self, problem: BenchmarkProblem) -> str:
"""Condition 2: Multi-perspective synthesis, no memory."""
if self.forge:
try:
result = self.forge.forge_single(problem.question)
return result.get("messages", [{}])[-1].get("content", "")
except Exception:
pass
# Fallback: combine multiple agent templates
if self.forge:
parts = []
for agent in self.forge.analysis_agents:
try:
parts.append(f"**{agent.name}:** {agent.analyze(problem.question)}")
except Exception:
continue
if parts:
synthesis = "\n\n".join(parts)
synthesis += (
f"\n\n**Synthesis:** These {len(parts)} perspectives on "
f"'{problem.question[:50]}...' converge on the importance of "
f"examining this from multiple angles. The analytical view provides "
f"causal structure, while philosophical and ethical views add depth."
)
return synthesis
return ""
def _generate_memory(self, problem: BenchmarkProblem) -> str:
"""Condition 3: Multi-perspective + cocoon memory augmentation."""
memory_context = ""
if self.memory:
try:
relevant = self.memory.recall_relevant(problem.question, max_results=3)
if relevant:
memory_context = "\n\n**Memory-Augmented Context:**\n"
for cocoon in relevant:
memory_context += (
f"- Prior reasoning on '{cocoon.get('query', '')[:60]}': "
f"{cocoon.get('response', '')[:100]}...\n"
)
memory_context += (
"\nDrawing on these prior reasoning exchanges, "
"the analysis benefits from accumulated insight.\n"
)
except Exception:
pass
multi_response = self._generate_multi(problem)
return multi_response + memory_context
def _generate_codette(self, problem: BenchmarkProblem) -> str:
"""Condition 4: Full Codette (multi + memory + strategy synthesis)."""
# Get strategy synthesis
strategy_context = ""
if self.synthesizer:
try:
comparison = self.synthesizer.run_full_synthesis(problem.question)
strategy_context = (
f"\n\n**Strategy Synthesis:**\n"
f"Forged strategy: {comparison.new_strategy.name}\n"
f"Definition: {comparison.new_strategy.definition[:200]}\n\n"
f"**Reasoning Path ({comparison.new_path.strategy_name}):**\n"
)
for i, step in enumerate(comparison.new_path.steps, 1):
strategy_context += f"{i}. {step}\n"
strategy_context += f"\n**Conclusion:** {comparison.new_path.conclusion}\n"
# Add evidence
strategy_context += "\n**Evidence from cocoon synthesis:**\n"
for ev in comparison.evidence_chain[:3]:
strategy_context += f"- {ev}\n"
except Exception as e:
logger.debug(f"Strategy synthesis failed: {e}")
memory_response = self._generate_memory(problem)
return memory_response + strategy_context
# ═══════════════════════════════════════════════════════════════════
# SECTION 4: STATISTICAL ANALYSIS & REPORT GENERATOR
# ═══════════════════════════════════════════════════════════════════
@dataclass
class ConditionStats:
"""Aggregate statistics for one condition."""
condition: str
n: int
mean_composite: float
std_composite: float
dimension_means: Dict[str, float]
dimension_stds: Dict[str, float]
mean_length: float
mean_latency: float
def compute_effect_size(group1: List[float], group2: List[float]) -> float:
"""Cohen's d effect size."""
n1, n2 = len(group1), len(group2)
if n1 < 2 or n2 < 2:
return 0.0
m1, m2 = statistics.mean(group1), statistics.mean(group2)
s1, s2 = statistics.stdev(group1), statistics.stdev(group2)
pooled_std = math.sqrt(((n1 - 1) * s1**2 + (n2 - 1) * s2**2) / (n1 + n2 - 2))
if pooled_std == 0:
return 0.0
return (m2 - m1) / pooled_std
def welch_t_test(group1: List[float], group2: List[float]) -> Tuple[float, float]:
"""Welch's t-test (unequal variance). Returns (t_stat, p_value_approx)."""
n1, n2 = len(group1), len(group2)
if n1 < 2 or n2 < 2:
return 0.0, 1.0
m1, m2 = statistics.mean(group1), statistics.mean(group2)
v1, v2 = statistics.variance(group1), statistics.variance(group2)
se = math.sqrt(v1/n1 + v2/n2)
if se == 0:
return 0.0, 1.0
t_stat = (m2 - m1) / se
# Welch-Satterthwaite degrees of freedom
num = (v1/n1 + v2/n2)**2
den = (v1/n1)**2/(n1-1) + (v2/n2)**2/(n2-1)
df = num / max(den, 1e-10)
# Approximate p-value using normal distribution for large df
# (scipy not guaranteed available)
z = abs(t_stat)
p_approx = 2 * (1 - 0.5 * (1 + math.erf(z / math.sqrt(2))))
return round(t_stat, 4), round(p_approx, 6)
class ReportGenerator:
"""Generate publishable benchmark reports."""
def __init__(self, results: List[BenchmarkScore], problems: List[BenchmarkProblem]):
self.results = results
self.problems = {p.id: p for p in problems}
def compute_stats(self) -> Dict[str, ConditionStats]:
"""Compute per-condition aggregate statistics."""
conditions = {}
for result in self.results:
if result.condition not in conditions:
conditions[result.condition] = []
conditions[result.condition].append(result)
stats = {}
for cond, scores in conditions.items():
composites = [s.composite for s in scores]
dim_scores = {}
for dim in ScoringEngine.DIMENSION_WEIGHTS:
dim_vals = [s.dimensions[dim].score for s in scores if dim in s.dimensions]
dim_scores[dim] = dim_vals
stats[cond] = ConditionStats(
condition=cond,
n=len(scores),
mean_composite=round(statistics.mean(composites), 4) if composites else 0,
std_composite=round(statistics.stdev(composites), 4) if len(composites) > 1 else 0,
dimension_means={d: round(statistics.mean(v), 4) for d, v in dim_scores.items() if v},
dimension_stds={d: round(statistics.stdev(v), 4) for d, v in dim_scores.items() if len(v) > 1},
mean_length=round(statistics.mean([s.response_length for s in scores]), 1),
mean_latency=round(statistics.mean([s.latency_ms for s in scores]), 1),
)
return stats
def compute_pairwise_comparisons(self) -> List[Dict]:
"""Statistical comparisons between conditions."""
conditions = {}
for r in self.results:
conditions.setdefault(r.condition, []).append(r.composite)
pairs = [
("SINGLE", "MULTI", "Multi-perspective vs single"),
("MULTI", "MEMORY", "Memory augmentation vs vanilla multi"),
("MEMORY", "CODETTE", "Full Codette vs memory-augmented"),
("SINGLE", "CODETTE", "Full Codette vs single (total improvement)"),
]
comparisons = []
for cond_a, cond_b, label in pairs:
if cond_a in conditions and cond_b in conditions:
g1, g2 = conditions[cond_a], conditions[cond_b]
t_stat, p_val = welch_t_test(g1, g2)
d = compute_effect_size(g1, g2)
delta = statistics.mean(g2) - statistics.mean(g1)
comparisons.append({
"comparison": label,
"condition_a": cond_a,
"condition_b": cond_b,
"mean_a": round(statistics.mean(g1), 4),
"mean_b": round(statistics.mean(g2), 4),
"delta": round(delta, 4),
"delta_pct": round(delta / max(statistics.mean(g1), 0.01) * 100, 1),
"cohens_d": round(d, 4),
"t_stat": t_stat,
"p_value": p_val,
"significant": p_val < 0.05,
})
return comparisons
def per_category_analysis(self) -> Dict[str, Dict]:
"""Break down results by problem category."""
by_category = {}
for r in self.results:
prob = self.problems.get(r.problem_id)
if not prob:
continue
cat = prob.category
if cat not in by_category:
by_category[cat] = {}
by_category[cat].setdefault(r.condition, []).append(r.composite)
analysis = {}
for cat, cond_scores in by_category.items():
analysis[cat] = {
cond: {
"mean": round(statistics.mean(scores), 4),
"std": round(statistics.stdev(scores), 4) if len(scores) > 1 else 0,
"n": len(scores),
}
for cond, scores in cond_scores.items()
}
return analysis
def generate_markdown_report(self) -> str:
"""Generate a publishable markdown report."""
stats = self.compute_stats()
comparisons = self.compute_pairwise_comparisons()
categories = self.per_category_analysis()
lines = []
lines.append("# Codette Benchmark Results")
lines.append(f"\n*Generated: {time.strftime('%Y-%m-%d %H:%M:%S')}*\n")
lines.append(f"*Problems: {len(self.problems)} | Conditions: {len(stats)} | Total evaluations: {len(self.results)}*\n")
# ─── Overall Results ───
lines.append("## 1. Overall Results by Condition\n")
lines.append("| Condition | N | Composite (mean +/- std) | Depth | Diversity | Coherence | Ethics | Novelty | Grounding | Turing |")
lines.append("|-----------|---|--------------------------|-------|-----------|-----------|--------|---------|-----------|--------|")
for cond in ["SINGLE", "MULTI", "MEMORY", "CODETTE"]:
s = stats.get(cond)
if not s:
continue
dm = s.dimension_means
lines.append(
f"| {cond} | {s.n} | **{s.mean_composite:.3f}** +/- {s.std_composite:.3f} | "
f"{dm.get('reasoning_depth', 0):.3f} | {dm.get('perspective_diversity', 0):.3f} | "
f"{dm.get('coherence', 0):.3f} | {dm.get('ethical_coverage', 0):.3f} | "
f"{dm.get('novelty', 0):.3f} | {dm.get('factual_grounding', 0):.3f} | "
f"{dm.get('turing_naturalness', 0):.3f} |"
)
# ─── Statistical Comparisons ───
lines.append("\n## 2. Statistical Comparisons\n")
lines.append("| Comparison | Delta | Delta % | Cohen's d | t-stat | p-value | Significant |")
lines.append("|------------|-------|---------|-----------|--------|---------|-------------|")
for c in comparisons:
sig = "**Yes**" if c["significant"] else "No"
lines.append(
f"| {c['comparison']} | {c['delta']:+.4f} | {c['delta_pct']:+.1f}% | "
f"{c['cohens_d']:.3f} | {c['t_stat']:.3f} | {c['p_value']:.4f} | {sig} |"
)
# Effect size interpretation
lines.append("\n*Cohen's d interpretation: 0.2=small, 0.5=medium, 0.8=large*\n")
# ─── Per-Category Breakdown ───
lines.append("## 3. Results by Problem Category\n")
for cat in ["reasoning", "ethics", "creative", "meta", "adversarial", "turing"]:
if cat not in categories:
continue
lines.append(f"### {cat.capitalize()}\n")
lines.append("| Condition | Mean | Std | N |")
lines.append("|-----------|------|-----|---|")
for cond in ["SINGLE", "MULTI", "MEMORY", "CODETTE"]:
if cond in categories[cat]:
cs = categories[cat][cond]
lines.append(f"| {cond} | {cs['mean']:.3f} | {cs['std']:.3f} | {cs['n']} |")
lines.append("")
# ─── Key Findings ───
lines.append("## 4. Key Findings\n")
for c in comparisons:
if c["significant"]:
direction = "improvement" if c["delta"] > 0 else "degradation"
lines.append(
f"- **{c['comparison']}**: {c['delta_pct']:+.1f}% {direction} "
f"(Cohen's d={c['cohens_d']:.2f}, p={c['p_value']:.4f})"
)
# ─── Methodology ───
lines.append("\n## 5. Methodology\n")
lines.append("### Conditions\n")
lines.append("1. **SINGLE** — Single analytical perspective, no memory, no synthesis")
lines.append("2. **MULTI** — All 6 reasoning agents (Newton, Quantum, Ethics, Philosophy, DaVinci, Empathy) + critic + synthesis")
lines.append("3. **MEMORY** — MULTI + cocoon memory augmentation (FTS5-retrieved prior reasoning)")
lines.append("4. **CODETTE** — MEMORY + meta-cognitive strategy synthesis (cross-domain pattern extraction + forged reasoning strategies)")
lines.append("\n### Scoring Dimensions (0-1 scale)\n")
lines.append("1. **Reasoning Depth** (20%) — chain length, concept density, ground truth coverage")
lines.append("2. **Perspective Diversity** (15%) — distinct cognitive dimensions engaged")
lines.append("3. **Coherence** (15%) — logical flow, transitions, structural consistency")
lines.append("4. **Ethical Coverage** (10%) — moral frameworks, stakeholders, value awareness")
lines.append("5. **Novelty** (15%) — non-obvious insights, cross-domain connections, reframing")
lines.append("6. **Factual Grounding** (15%) — evidence specificity, ground truth alignment, trap avoidance")
lines.append("7. **Turing Naturalness** (10%) — conversational quality, absence of formulaic AI patterns")
lines.append("\n### Problem Set\n")
lines.append(f"- {len(self.problems)} problems across 6 categories")
lines.append("- Categories: reasoning (3), ethics (3), creative (2), meta-cognitive (3), adversarial (3), Turing (3)")
lines.append("- Difficulty: easy (1), medium (6), hard (10)")
lines.append("\n### Statistical Tests\n")
lines.append("- Welch's t-test (unequal variance) for pairwise condition comparisons")
lines.append("- Cohen's d for effect size estimation")
lines.append("- Significance threshold: p < 0.05")
return "\n".join(lines)
def generate_json_report(self) -> Dict:
"""Generate structured JSON report for machine consumption."""
stats = self.compute_stats()
comparisons = self.compute_pairwise_comparisons()
categories = self.per_category_analysis()
per_problem = {}
for r in self.results:
if r.problem_id not in per_problem:
per_problem[r.problem_id] = {}
per_problem[r.problem_id][r.condition] = {
"composite": r.composite,
"dimensions": {
d: {"score": ds.score, "evidence": ds.evidence, "penalties": ds.penalties}
for d, ds in r.dimensions.items()
},
"response_length": r.response_length,
"latency_ms": r.latency_ms,
}
return {
"metadata": {
"timestamp": time.strftime('%Y-%m-%dT%H:%M:%S'),
"num_problems": len(self.problems),
"num_conditions": len(stats),
"total_evaluations": len(self.results),
},
"condition_stats": {
c: {
"mean_composite": s.mean_composite,
"std_composite": s.std_composite,
"dimension_means": s.dimension_means,
"dimension_stds": s.dimension_stds,
"mean_length": s.mean_length,
"mean_latency": s.mean_latency,
"n": s.n,
}
for c, s in stats.items()
},
"pairwise_comparisons": comparisons,
"per_category": categories,
"per_problem": per_problem,
}
# ═══════════════════════════════════════════════════════════════════
# SECTION 5: MAIN ENTRY POINT
# ═══════════════════════════════════════════════════════════════════
def run_benchmarks(
output_dir: Optional[str] = None,
use_llm: bool = False,
verbose: bool = True,
) -> Tuple[str, Dict]:
"""
Run the full benchmark suite and generate reports.
Returns:
(markdown_report, json_report)
"""
if output_dir is None:
output_dir = str(_PROJECT_ROOT / "data" / "results")
os.makedirs(output_dir, exist_ok=True)
# Get problems
problems = get_benchmark_problems()
if verbose:
logger.info(f"Benchmark suite: {len(problems)} problems across "
f"{len(set(p.category for p in problems))} categories")
# Run
runner = BenchmarkRunner(use_llm=use_llm, verbose=verbose)
results = runner.run_all(problems)
# Generate reports
reporter = ReportGenerator(results, problems)
md_report = reporter.generate_markdown_report()
json_report = reporter.generate_json_report()
# Save
md_path = os.path.join(output_dir, "codette_benchmark_report.md")
json_path = os.path.join(output_dir, "codette_benchmark_results.json")
with open(md_path, "w", encoding="utf-8") as f:
f.write(md_report)
with open(json_path, "w", encoding="utf-8") as f:
json.dump(json_report, f, indent=2, default=str)
if verbose:
logger.info(f"\nReports saved:")
logger.info(f" Markdown: {md_path}")
logger.info(f" JSON: {json_path}")
return md_report, json_report
if __name__ == "__main__":
import argparse
parser = argparse.ArgumentParser(description="Codette Benchmark Suite")
parser.add_argument("--output", default=None, help="Output directory")
parser.add_argument("--llm", action="store_true", help="Use live LLM inference")
parser.add_argument("--quiet", action="store_true", help="Suppress progress output")
args = parser.parse_args()
md, js = run_benchmarks(
output_dir=args.output,
use_llm=args.llm,
verbose=not args.quiet,
)
print("\n" + md)
|