File size: 35,571 Bytes
b33660c | 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 | """
Phase 11 — Deep-Dive Adversarial Diagnostic Script
====================================================
Senior ML Engineer adversarial testing for the BAYAN pipeline.
Tests raw models SEPARATELY then the INTEGRATED pipeline.
Forcefully triggers 5 known edge-case classes and provides root cause analysis.
Usage:
python tests/phase11/deep_dive_adversarial.py [--url URL]
"""
import argparse
import json
import re
import sys
import time
import difflib
from pathlib import Path
from dataclasses import dataclass, field, asdict
from typing import List, Dict, Optional, Tuple
import requests
DEFAULT_URL = "https://bayan10-bayan-api.hf.space"
REPORT_DIR = Path(__file__).parent / "reports"
# ═══════════════════════════════════════════════════════════════════════════════
# API CLIENT
# ═══════════════════════════════════════════════════════════════════════════════
class API:
def __init__(self, base):
self.base = base.rstrip('/')
self.s = requests.Session()
self.s.headers['Content-Type'] = 'application/json'
def _post(self, ep, payload, timeout=180):
t0 = time.time()
try:
r = self.s.post(f"{self.base}{ep}", json=payload, timeout=timeout)
ms = int((time.time() - t0) * 1000)
d = r.json()
d['_ms'] = ms
d['_status'] = r.status_code
return d
except requests.Timeout:
return {'error': 'TIMEOUT', '_ms': int((time.time() - t0) * 1000), '_status': 0}
except Exception as e:
return {'error': str(e), '_ms': int((time.time() - t0) * 1000), '_status': 0}
def analyze(self, text):
"""Full pipeline: /api/analyze"""
return self._post('/api/analyze', {'text': text})
def grammar(self, text):
"""Raw grammar only: /api/grammar"""
return self._post('/api/grammar', {'text': text})
def punctuation(self, text):
"""Raw punctuation only: /api/punctuation"""
return self._post('/api/punctuation', {'text': text})
# ═══════════════════════════════════════════════════════════════════════════════
# RESULT TYPES
# ═══════════════════════════════════════════════════════════════════════════════
@dataclass
class TestVector:
"""A single adversarial test case."""
id: str
category: str # Edge-case category (1-5)
subcategory: str # Specific sub-type
input_text: str
expected_behavior: str # What SHOULD happen
severity: str = "critical"
@dataclass
class IsolationResult:
"""Results from testing each model in isolation + pipeline."""
vector_id: str
category: str
subcategory: str
input_text: str
expected_behavior: str
# Raw model outputs
grammar_raw_output: str = ""
grammar_raw_ms: int = 0
punctuation_raw_output: str = ""
punctuation_raw_ms: int = 0
# Pipeline output
pipeline_output: str = ""
pipeline_corrected: str = ""
pipeline_suggestions: list = field(default_factory=list)
pipeline_ms: int = 0
pipeline_timing: dict = field(default_factory=dict)
# Analysis
verdict: str = "" # PASS, FAIL_MODEL, FAIL_PIPELINE, FAIL_FILTER, ERROR
root_cause: str = ""
failure_detail: str = ""
# Integration delta
grammar_raw_changed: bool = False
pipeline_changed: bool = False
raw_vs_pipeline_match: bool = False
# Coordinate validation
span_errors: list = field(default_factory=list)
offset_drift: int = 0
# ═══════════════════════════════════════════════════════════════════════════════
# STEP 1: ADVERSARIAL TEST SUITE
# ═══════════════════════════════════════════════════════════════════════════════
ADVERSARIAL_VECTORS: List[TestVector] = [
# ──────────────────────────────────────────────────────────
# CATEGORY 1: Punctuation Coordinate Shift (OffsetMapper bug)
# ──────────────────────────────────────────────────────────
TestVector("OFS-001", "offset_shift", "punct_after_spelling_fix",
"ذهبت الي المدرسه وقابلت صديقتي وتحدثنا كثيرا",
"Punct marks should align with corrected text, not shift 1-2 words"),
TestVector("OFS-002", "offset_shift", "punct_after_multi_word_fix",
"الطلاب ذهبو الي المكتبه لقراءة الكتب",
"Punctuation positions should not drift after spelling corrects الي→إلى and المكتبه→المكتبة"),
TestVector("OFS-003", "offset_shift", "length_changing_fix",
"انا احب القراءه كثيرا وخصوصا الكتب العلميه",
"أنا → 3 chars vs انا → 3 chars (same length), but القراءه→القراءة changes char positions"),
TestVector("OFS-004", "offset_shift", "insertion_diffs",
"كتبت رساله لصديقي واخبرته بالاخبار",
"When spelling inserts أ (اخبرته→أخبرته), downstream coords must shift +1"),
TestVector("OFS-005", "offset_shift", "multi_stage_coord_chain",
"المهندسون ذهبو الي الموقعو وبدأو العمل",
"3+ fixes in same sentence: verify offset chain doesn't accumulate drift"),
# ──────────────────────────────────────────────────────────
# CATEGORY 2: Prefixes and Prepositions (السوابق وحروف الجر)
# ──────────────────────────────────────────────────────────
TestVector("PFX-001", "prefix_handling", "wa_prefix",
"وبالمستشفيات الكبيره يعالجون المرضي",
"وبالمستشفيات should NOT be split or hallucinated — it's a valid prefixed word"),
TestVector("PFX-002", "prefix_handling", "ba_prefix",
"بالمدرسه الجديده تعلمت كثيرا",
"بالمدرسه → بالمدرسة (only ه→ة fix, keep ب prefix)"),
TestVector("PFX-003", "prefix_handling", "li_prefix",
"للطالبات المجتهدات جوائز قيمه",
"للطالبات should be preserved intact"),
TestVector("PFX-004", "prefix_handling", "ka_prefix",
"كالعاده ذهبت الي العمل",
"كالعادة — ه→ة fix should work through ك+ال prefix"),
TestVector("PFX-005", "prefix_handling", "fa_prefix",
"فبالتالي يجب ان نعمل بجد",
"فبالتالي is valid, should not be split or changed"),
TestVector("PFX-006", "prefix_handling", "complex_prefix_chain",
"وبالاضافه الي ذلك فان المشكله كبيره",
"وبالإضافة إلى — hamza fix through وب prefix, plus ه→ة fixes"),
TestVector("PFX-007", "prefix_handling", "wa_attached_to_error",
"والمهندسين بنو المبني بسرعه",
"والمهندسين should stay, بنو→بنوا (grammar), المبني→المبنى, بسرعه→بسرعة"),
TestVector("PFX-008", "prefix_handling", "attached_pronoun",
"كتابه قديم وصفحاته ممزقه",
"كتابه (his book) — ه is pronoun, NOT ta marbuta. Must NOT become كتابة"),
TestVector("PFX-009", "prefix_handling", "pronoun_vs_tamarbuta",
"قرأته بعنايه فتأملته جيدا",
"قرأته and فتأملته — ته is pronoun suffix, NOT ta marbuta"),
TestVector("PFX-010", "prefix_handling", "merged_preposition",
"ذهبتللمدرسه صباحا",
"Missing space: ذهبت+ل+المدرسه — should split and fix"),
# ──────────────────────────────────────────────────────────
# CATEGORY 3: StageLocker / Directional Flow
# ──────────────────────────────────────────────────────────
TestVector("SLK-001", "stagelocker", "spelling_lock_grammar_override",
"السياره جميل جدا",
"Spelling fixes السياره→السيارة, Grammar should still fix جميل→جميلة (Phase 11 hierarchical)"),
TestVector("SLK-002", "stagelocker", "grammar_not_overwrite_spelling",
"المدرسه كبير وجميل",
"Spelling: المدرسه→المدرسة. Grammar should NOT revert to المدرسه. Should add ة to كبير/جميل"),
TestVector("SLK-003", "stagelocker", "punct_respects_grammar_lock",
"الطلاب ذهبوا الى المكتبة لقراءة الكتب الجديدة",
"Grammar unchanged (correct text). Punctuation should NOT move/corrupt words"),
TestVector("SLK-004", "stagelocker", "overlapping_patches",
"الاطفال يلعبون بالكره في الحديقه الكبيره",
"Spelling: الاطفال→الأطفال, بالكره→بالكرة, الحديقه→الحديقة, الكبيره→الكبيرة — multiple overlapping spans"),
TestVector("SLK-005", "stagelocker", "backward_overwrite",
"هذه المدينه جميله وهادئه",
"Three ه→ة fixes. Grammar must NOT overwrite any of them back to ه"),
TestVector("SLK-006", "stagelocker", "grammar_overrides_spelling_adjacent",
"الطالبه ذكي في المدرسه",
"Spelling fixes ه→ة, Grammar must fix gender (ذكي→ذكية) on ADJACENT word"),
# ──────────────────────────────────────────────────────────
# CATEGORY 4: Overcorrection & Safety Filter Analysis
# ──────────────────────────────────────────────────────────
TestVector("OVR-001", "overcorrection", "correct_text_unchanged",
"ذهب الطالب إلى المكتبة وقرأ كتاباً مفيداً",
"Perfectly correct text — should return UNCHANGED"),
TestVector("OVR-002", "overcorrection", "correct_grammar_unchanged",
"المهندسون يعملون في المصنع الكبير بجدٍ واجتهاد",
"Correct SV agreement + tanween — must not be modified"),
TestVector("OVR-003", "overcorrection", "jaccard_threshold_test",
"جالس في البيت يقرأ الصحيفة",
"Correct text. Grammar should not hallucinate جالس→جاكسون (Jaccard < 0.3)"),
TestVector("OVR-004", "overcorrection", "meaning_change_block",
"وكان الجو صحواً اليوم",
"وكان must NOT become وكأن (directional block). Correct as-is"),
TestVector("OVR-005", "overcorrection", "rare_word_dampening",
"الفيلسوف سقراط تأمل الحقيقة",
"سقراط is a rare valid name — spelling should not 'correct' it"),
TestVector("OVR-006", "overcorrection", "filter_over_blocking",
"انا ذاهب الي المدرسه",
"IV-IV guard should NOT block انا→أنا or الي→إلى (these are whitelist entries)"),
TestVector("OVR-007", "overcorrection", "tanween_preservation",
"قرأت كتاباً جميلاً عن التاريخ",
"Tanween (ً) must be preserved — grammar should not strip it"),
TestVector("OVR-008", "overcorrection", "correct_feminine_form",
"المعلمة الجديدة شرحت الدرس",
"Already correct — must NOT change المعلمة to المعلمه"),
TestVector("OVR-009", "overcorrection", "brackets_protection",
"الفصل الأول (المقدمة) يتناول الموضوع",
"Brackets must be preserved. Grammar should not unbalance them"),
TestVector("OVR-010", "overcorrection", "digits_protection",
"تأسست الشركة عام 2020 وبلغ عدد موظفيها 150 شخصاً",
"Digits must NOT be corrupted by grammar model"),
# ──────────────────────────────────────────────────────────
# CATEGORY 5: Model Mismatch (Spelling→Grammar friction)
# ──────────────────────────────────────────────────────────
TestVector("MIS-001", "model_mismatch", "spelling_output_breaks_grammar",
"الطالبه المجتهده حصلت علي درجات عاليه",
"Spelling: 4x ه→ة. Grammar should NOT break on the corrected output"),
TestVector("MIS-002", "model_mismatch", "grammar_undoes_spelling",
"ذهبنا الي المكتبه القريبه",
"Spelling fixes الي→إلى, المكتبه→المكتبة. Grammar must NOT revert these"),
TestVector("MIS-003", "model_mismatch", "cascading_length_change",
"اريد ان اذهب الي المدرسه لانني احب القراءه",
"Many hamza + ه→ة fixes. Total text length changes significantly — grammar must handle shifted text"),
TestVector("MIS-004", "model_mismatch", "grammar_hallucination_on_corrected",
"الاولاد يلعبون والبنات يذهبن",
"Spelling: الاولاد→الأولاد. Grammar should NOT hallucinate on already-correct يلعبون/يذهبن"),
TestVector("MIS-005", "model_mismatch", "punc_on_grammar_corrected",
"السيارات سريعة والطرق واسعه",
"After spelling (واسعه→واسعة), punctuation should add period at end, not corrupt words"),
]
# ═══════════════════════════════════════════════════════════════════════════════
# STEP 2: ISOLATION vs INTEGRATION MATRIX
# ═══════════════════════════════════════════════════════════════════════════════
def strip_diacritics(text: str) -> str:
return re.sub(r'[\u064B-\u065F\u0670]', '', text)
def normalize(text: str) -> str:
t = strip_diacritics(text)
t = re.sub(r'\s+', ' ', t).strip()
return t
def validate_spans(original: str, suggestions: list) -> List[str]:
"""Validate all suggestion spans point to correct original text."""
errors = []
for sg in suggestions:
start = sg.get('start', 0)
end = sg.get('end', 0)
expected_orig = sg.get('original', '')
actual_slice = original[start:end]
if actual_slice != expected_orig:
errors.append(
f"SPAN[{start}:{end}] expected='{expected_orig}' actual='{actual_slice}' "
f"(drift={len(actual_slice) - len(expected_orig)})"
)
return errors
def compute_offset_drift(original: str, suggestions: list) -> int:
"""Compute max drift between expected and actual span content."""
max_drift = 0
for sg in suggestions:
start = sg.get('start', 0)
end = sg.get('end', 0)
expected = sg.get('original', '')
actual = original[start:end]
if actual != expected:
# Try to find the expected text nearby
for delta in range(-5, 6):
try_start = max(0, start + delta)
try_end = min(len(original), end + delta)
if original[try_start:try_end] == expected:
max_drift = max(max_drift, abs(delta))
break
else:
max_drift = max(max_drift, 99) # Not found at all
return max_drift
def run_isolation_matrix(api: API, vectors: List[TestVector]) -> List[IsolationResult]:
"""Run each vector through raw models + pipeline and compare."""
results = []
for i, v in enumerate(vectors):
print(f"\n [{i+1}/{len(vectors)}] {v.id} ({v.category}/{v.subcategory})")
print(f" Input: {v.input_text[:80]}...")
r = IsolationResult(
vector_id=v.id,
category=v.category,
subcategory=v.subcategory,
input_text=v.input_text,
expected_behavior=v.expected_behavior,
)
# ── A) Raw Grammar Only ──
try:
resp_g = api.grammar(v.input_text)
r.grammar_raw_output = resp_g.get('corrected_text', resp_g.get('corrected', ''))
r.grammar_raw_ms = resp_g.get('_ms', 0)
r.grammar_raw_changed = (r.grammar_raw_output != v.input_text)
print(f" [A] Grammar raw: {'CHANGED' if r.grammar_raw_changed else 'unchanged'} ({r.grammar_raw_ms}ms)")
if r.grammar_raw_changed:
print(f" → {r.grammar_raw_output[:80]}")
except Exception as e:
print(f" [A] Grammar raw: ERROR ({e})")
# ── B) Raw Punctuation Only ──
try:
resp_p = api.punctuation(v.input_text)
r.punctuation_raw_output = resp_p.get('corrected_text', resp_p.get('corrected', ''))
r.punctuation_raw_ms = resp_p.get('_ms', 0)
punct_changed = (r.punctuation_raw_output != v.input_text)
print(f" [B] Punct raw: {'CHANGED' if punct_changed else 'unchanged'} ({r.punctuation_raw_ms}ms)")
if punct_changed:
print(f" → {r.punctuation_raw_output[:80]}")
except Exception as e:
print(f" [B] Punct raw: ERROR ({e})")
# ── C) Full Pipeline ──
try:
resp = api.analyze(v.input_text)
r.pipeline_ms = resp.get('_ms', 0)
r.pipeline_timing = resp.get('timing_ms', {})
r.pipeline_corrected = resp.get('corrected', '')
r.pipeline_output = r.pipeline_corrected
r.pipeline_suggestions = resp.get('suggestions', [])
r.pipeline_changed = (r.pipeline_corrected != resp.get('original', v.input_text))
print(f" [C] Pipeline: {'CHANGED' if r.pipeline_changed else 'unchanged'} ({r.pipeline_ms}ms) [{len(r.pipeline_suggestions)} suggestions]")
if r.pipeline_changed:
print(f" → {r.pipeline_corrected[:80]}")
# ── Span validation ──
original = resp.get('original', v.input_text)
r.span_errors = validate_spans(original, r.pipeline_suggestions)
r.offset_drift = compute_offset_drift(original, r.pipeline_suggestions)
if r.span_errors:
print(f" ⚠️ SPAN ERRORS: {r.span_errors}")
if r.offset_drift > 0:
print(f" ⚠️ OFFSET DRIFT: {r.offset_drift} chars")
# ── Raw vs Pipeline comparison ──
norm_grammar = normalize(r.grammar_raw_output)
norm_pipeline = normalize(r.pipeline_corrected)
r.raw_vs_pipeline_match = (norm_grammar == norm_pipeline)
except Exception as e:
r.verdict = "ERROR"
r.failure_detail = str(e)
print(f" [C] Pipeline: ERROR ({e})")
# ── Verdict determination ──
_classify_result(v, r)
icon = {"PASS": "✅", "FAIL_MODEL": "🔴", "FAIL_PIPELINE": "🟡",
"FAIL_FILTER": "🟠", "FAIL_OFFSET": "🔵", "ERROR": "💥"}.get(r.verdict, "❓")
print(f" {icon} {r.verdict}: {r.root_cause}")
results.append(r)
return results
def _classify_result(v: TestVector, r: IsolationResult):
"""Classify the test result into a verdict + root cause."""
if r.verdict == "ERROR":
return
cat = v.category
# ── Category 1: Offset shift ──
if cat == "offset_shift":
if r.span_errors:
r.verdict = "FAIL_OFFSET"
r.root_cause = f"OffsetMapper coordinate drift ({r.offset_drift} chars): {r.span_errors[0]}"
r.failure_detail = "; ".join(r.span_errors)
elif r.offset_drift > 0:
r.verdict = "FAIL_OFFSET"
r.root_cause = f"Offset drift detected: {r.offset_drift} chars"
else:
r.verdict = "PASS"
r.root_cause = "Spans aligned correctly"
return
# ── Category 2: Prefix handling ──
if cat == "prefix_handling":
output = r.pipeline_corrected
inp = v.input_text
# Check for catastrophic prefix destruction
if _check_prefix_destruction(inp, output):
r.verdict = "FAIL_MODEL"
r.root_cause = "Prefix word was split, hallucinated, or destroyed"
r.failure_detail = f"Input had prefixed words that were corrupted in output"
elif r.span_errors:
r.verdict = "FAIL_OFFSET"
r.root_cause = f"Span errors in prefix test: {r.span_errors[0]}"
else:
r.verdict = "PASS"
r.root_cause = "Prefixed words handled correctly"
return
# ── Category 3: StageLocker ──
if cat == "stagelocker":
# Check if grammar fix was blocked by spelling lock
sugg_stages = [s.get('type', '') for s in r.pipeline_suggestions]
if v.subcategory == "spelling_lock_grammar_override":
has_grammar = 'grammar' in sugg_stages
has_spelling = 'spelling' in sugg_stages
if has_spelling and not has_grammar:
r.verdict = "FAIL_PIPELINE"
r.root_cause = "StageLocker STILL blocking grammar override of spelling (Phase 11 fix not deployed?)"
elif has_spelling and has_grammar:
r.verdict = "PASS"
r.root_cause = "Grammar successfully overrode spelling lock"
else:
r.verdict = "PASS"
r.root_cause = "Pipeline handled correctly"
elif v.subcategory == "backward_overwrite":
# Check that spelling fixes were preserved
if 'ه' in r.pipeline_corrected and 'المدين' in r.pipeline_corrected:
# Check if any ه→ة fix was reverted
for sg in r.pipeline_suggestions:
if sg.get('type') == 'grammar':
orig_sg = sg.get('original', '')
corr_sg = sg.get('correction', '')
if orig_sg.endswith('ة') and corr_sg.endswith('ه'):
r.verdict = "FAIL_PIPELINE"
r.root_cause = f"Grammar reverted spelling fix: '{orig_sg}'→'{corr_sg}'"
return
r.verdict = "PASS"
r.root_cause = "No backward overwrite detected"
else:
r.verdict = "PASS"
r.root_cause = "StageLocker operating correctly"
return
# ── Category 4: Overcorrection ──
if cat == "overcorrection":
if v.subcategory in ("correct_text_unchanged", "correct_grammar_unchanged",
"correct_feminine_form", "tanween_preservation"):
if r.pipeline_changed:
changes = [f"{s.get('original','')}→{s.get('correction','')}"
for s in r.pipeline_suggestions]
r.verdict = "FAIL_FILTER"
r.root_cause = f"Correct text was modified: {changes[:3]}"
r.failure_detail = f"All changes: {changes}"
else:
r.verdict = "PASS"
r.root_cause = "Correct text preserved"
elif v.subcategory == "meaning_change_block":
# Check if وكان was changed to وكأن
if 'وكأن' in r.pipeline_corrected:
r.verdict = "FAIL_FILTER"
r.root_cause = "Directional block failed: وكان→وكأن leaked through"
else:
r.verdict = "PASS"
r.root_cause = "Meaning preserved"
elif v.subcategory == "digits_protection":
if not re.search(r'2020', r.pipeline_corrected):
r.verdict = "FAIL_FILTER"
r.root_cause = "Digits were corrupted"
elif not re.search(r'150', r.pipeline_corrected):
r.verdict = "FAIL_FILTER"
r.root_cause = "Digits were corrupted"
else:
r.verdict = "PASS"
r.root_cause = "Digits preserved"
elif v.subcategory == "brackets_protection":
if '(' in v.input_text and '(' not in r.pipeline_corrected:
r.verdict = "FAIL_FILTER"
r.root_cause = "Brackets were removed"
else:
r.verdict = "PASS"
r.root_cause = "Brackets preserved"
else:
if r.pipeline_changed and not r.pipeline_suggestions:
r.verdict = "FAIL_FILTER"
r.root_cause = "Text changed without suggestions"
else:
r.verdict = "PASS"
r.root_cause = "No overcorrection detected"
return
# ── Category 5: Model mismatch ──
if cat == "model_mismatch":
if v.subcategory == "grammar_undoes_spelling":
# Check if spelling corrections were preserved
if 'الي' in r.pipeline_corrected or 'المكتبه' in r.pipeline_corrected:
r.verdict = "FAIL_PIPELINE"
r.root_cause = "Grammar undid spelling corrections"
else:
r.verdict = "PASS"
r.root_cause = "Spelling corrections preserved through grammar"
elif v.subcategory == "spelling_output_breaks_grammar":
if r.verdict != "ERROR" and r.pipeline_corrected:
r.verdict = "PASS"
r.root_cause = "Grammar handled spelling-corrected input"
else:
r.verdict = "FAIL_PIPELINE"
r.root_cause = "Grammar broke on spelling output"
else:
r.verdict = "PASS"
r.root_cause = "No model mismatch detected"
return
r.verdict = "PASS"
r.root_cause = "No issues detected"
def _check_prefix_destruction(input_text: str, output: str) -> bool:
"""Check if prefixed words were catastrophically destroyed."""
PREFIXED_PATTERNS = [
(r'\bوبال\w+', 'وبال'),
(r'\bبال\w+', 'بال'),
(r'\bكال\w+', 'كال'),
(r'\bفبال\w+', 'فبال'),
(r'\bلل\w+', 'لل'),
]
input_words = set(input_text.split())
output_words = set(output.split())
for word in input_words:
for pattern, prefix in PREFIXED_PATTERNS:
if re.match(pattern, word):
# This word has a prefix — check if it survived
# Allow ه→ة change but not complete destruction
stripped_orig = re.sub(r'[ةه]$', '', word)
found = False
for ow in output_words:
stripped_out = re.sub(r'[ةه]$', '', ow)
if stripped_orig == stripped_out or ow.startswith(prefix):
found = True
break
if not found:
return True # Prefixed word was destroyed
return False
# ═══════════════════════════════════════════════════════════════════════════════
# STEP 3: ANALYSIS & REPORTING
# ═══════════════════════════════════════════════════════════════════════════════
def generate_report(results: List[IsolationResult], url: str) -> dict:
"""Generate comprehensive analysis report."""
# Aggregate
total = len(results)
by_verdict = {}
by_category = {}
offset_issues = []
filter_issues = []
pipeline_issues = []
model_issues = []
for r in results:
by_verdict[r.verdict] = by_verdict.get(r.verdict, 0) + 1
cat = r.category
if cat not in by_category:
by_category[cat] = {"total": 0, "pass": 0, "fail": 0}
by_category[cat]["total"] += 1
if r.verdict == "PASS":
by_category[cat]["pass"] += 1
else:
by_category[cat]["fail"] += 1
if r.verdict == "FAIL_OFFSET":
offset_issues.append(r)
elif r.verdict == "FAIL_FILTER":
filter_issues.append(r)
elif r.verdict == "FAIL_PIPELINE":
pipeline_issues.append(r)
elif r.verdict == "FAIL_MODEL":
model_issues.append(r)
pass_count = by_verdict.get("PASS", 0)
pass_rate = pass_count / total if total > 0 else 0
report = {
"timestamp": time.strftime("%Y-%m-%dT%H:%M:%SZ"),
"target": url,
"total_vectors": total,
"pass_rate": round(pass_rate * 100, 1),
"by_verdict": by_verdict,
"by_category": by_category,
"offset_issues": [asdict(r) for r in offset_issues],
"filter_issues": [asdict(r) for r in filter_issues],
"pipeline_issues": [asdict(r) for r in pipeline_issues],
"model_issues": [asdict(r) for r in model_issues],
"results": [asdict(r) for r in results],
}
return report
def print_report(results: List[IsolationResult], report: dict):
"""Print detailed report to console."""
print(f"\n{'='*70}")
print("ADVERSARIAL DIAGNOSTIC REPORT")
print(f"{'='*70}")
print(f"\n## Overall: {report['pass_rate']}% pass rate ({report['total_vectors']} vectors)")
print(f"\n| Verdict | Count |")
print(f"|-----------------|-------|")
for verdict, count in sorted(report['by_verdict'].items()):
icon = {"PASS": "✅", "FAIL_MODEL": "🔴", "FAIL_PIPELINE": "🟡",
"FAIL_FILTER": "🟠", "FAIL_OFFSET": "🔵", "ERROR": "💥"}.get(verdict, "❓")
print(f"| {icon} {verdict:<14} | {count:>5} |")
print(f"\n## By Edge-Case Category")
print(f"| Category | Total | Pass | Fail | Rate |")
print(f"|----------------------|-------|------|------|------|")
for cat, data in sorted(report['by_category'].items()):
rate = data['pass'] / data['total'] * 100 if data['total'] > 0 else 0
print(f"| {cat:<20} | {data['total']:>5} | {data['pass']:>4} | {data['fail']:>4} | {rate:>3.0f}% |")
# ── Failure details ──
failures = [r for r in results if r.verdict != "PASS"]
if failures:
print(f"\n## Failure Details ({len(failures)} issues)")
print(f"{'─'*70}")
for r in failures:
icon = {"FAIL_MODEL": "🔴", "FAIL_PIPELINE": "🟡",
"FAIL_FILTER": "🟠", "FAIL_OFFSET": "🔵", "ERROR": "💥"}.get(r.verdict, "❓")
print(f"\n {icon} [{r.vector_id}] {r.category}/{r.subcategory}")
print(f" Input: {r.input_text[:70]}")
print(f" Pipeline: {r.pipeline_corrected[:70]}")
print(f" Cause: {r.root_cause}")
if r.span_errors:
for se in r.span_errors[:3]:
print(f" Span: {se}")
# ── OffsetMapper analysis ──
all_drifts = [r.offset_drift for r in results if r.offset_drift > 0]
if all_drifts:
print(f"\n## OffsetMapper Drift Analysis")
print(f" Vectors with drift: {len(all_drifts)}/{len(results)}")
print(f" Max drift: {max(all_drifts)} chars")
print(f" Affected vectors: {[r.vector_id for r in results if r.offset_drift > 0]}")
# ═══════════════════════════════════════════════════════════════════════════════
# MAIN
# ═══════════════════════════════════════════════════════════════════════════════
def main():
parser = argparse.ArgumentParser(description="Phase 11 Adversarial Diagnostic")
parser.add_argument("--url", default=DEFAULT_URL)
parser.add_argument("--category", default=None, help="Filter by category (offset_shift, prefix_handling, etc.)")
args = parser.parse_args()
api = API(args.url)
# Filter vectors if category specified
vectors = ADVERSARIAL_VECTORS
if args.category:
vectors = [v for v in vectors if v.category == args.category]
print(f"\n{'='*70}")
print(f"BAYAN ADVERSARIAL DIAGNOSTIC — Phase 11")
print(f"{'='*70}")
print(f" Target: {args.url}")
print(f" Vectors: {len(vectors)}")
print(f" Categories: {sorted(set(v.category for v in vectors))}")
print(f"{'='*70}")
results = run_isolation_matrix(api, vectors)
report = generate_report(results, args.url)
print_report(results, report)
# Save JSON report
REPORT_DIR.mkdir(parents=True, exist_ok=True)
out_path = REPORT_DIR / "adversarial_diagnostic.json"
with open(out_path, 'w', encoding='utf-8') as f:
json.dump(report, f, ensure_ascii=False, indent=2)
print(f"\n[DIAG] Full report → {out_path}")
if __name__ == "__main__":
main()
|