File size: 20,500 Bytes
56257d2
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
e4347be
 
 
 
 
 
 
 
56257d2
 
 
 
 
 
 
 
 
e4347be
 
 
 
56257d2
e4347be
56257d2
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
dd054aa
56257d2
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
dd054aa
 
 
 
9399089
56257d2
 
 
 
9399089
56257d2
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
9399089
 
 
56257d2
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import re
import random
from typing import Any, Dict, List, Optional, Tuple
from uuid import uuid4
from dataclasses import dataclass, field
from models import (
    Observation as ObsModel,
    Action as ActModel,
    Reward as RewModel,
    Document,
    DataPractice,
)


@dataclass
class TaskConfig:
    task_id: str
    name: str
    difficulty: str
    description: str
    privacy_policy: str
    data_practices: List[Dict]
    compliance_requirements: List[str]
    hidden_issues: List[Dict]


TASKS = {
    "easy": TaskConfig(
        task_id="easy_clause_existence",
        name="Clause Existence Check",
        difficulty="easy",
        description="Verify if mandatory GDPR clauses are present in the privacy policy",
        privacy_policy="""Privacy Policy - TechCorp Inc.

We collect the following information:
- Email address for account creation
- Payment information for transactions
- Device information for analytics

We use your data to:
- Provide our services
- Improve user experience
- Communicate with you

Data Retention: We retain data for 3 years after account deletion.

Contact: privacy@techcorp.example.com""",
        data_practices=[
            {"id": "dp1", "category": "Account", "purpose": "Service delivery", "data_type": "Email", "shared_with_third_parties": False},
            {"id": "dp2", "category": "Payment", "purpose": "Transaction processing", "data_type": "Financial", "shared_with_third_parties": True},
            {"id": "dp3", "category": "Analytics", "purpose": "Improve services", "data_type": "Device", "shared_with_third_parties": True},
        ],
        compliance_requirements=[
            "Right to be Forgotten",
            "Data Portability",
            "Contact Information",
        ],
        hidden_issues=[
            {"type": "missing_clause", "expected": "Right to be Forgotten", "severity": "high"},
            {"type": "missing_clause", "expected": "Data Portability", "severity": "medium"},
        ],
    ),
    "medium": TaskConfig(
        task_id="medium_purpose_mapping",
        name="Purpose Mapping",
        difficulty="medium",
        description="Match data collection points to their stated purposes and identify mismatches",
        privacy_policy="""Privacy Policy - DataFlow Analytics Inc.

DATA COLLECTION:
1. Geolocation Data - Purpose: App functionality
2. Browsing History - Purpose: Improving user experience
3. Social Media Handles - Purpose: Account linking
4. Health Metrics - Purpose: Personalization
5. Device Identifiers - Purpose: Advertising

THIRD-PARTY SHARING:
We share data with:
- Analytics partners (Google Analytics)
- Advertising networks
- Social media platforms

We do NOT share health data with third parties.

User rights include access, correction, and deletion.""",
        data_practices=[
            {"id": "dp1", "category": "Geolocation", "purpose": "App functionality", "data_type": "Location", "shared_with_third_parties": True},
            {"id": "dp2", "category": "Browsing History", "purpose": "Advertising", "data_type": "Behavioral", "shared_with_third_parties": True},
            {"id": "dp3", "category": "Health Metrics", "purpose": "Advertising", "data_type": "Health", "shared_with_third_parties": True},
            {"id": "dp4", "category": "Device Identifiers", "purpose": "Advertising", "data_type": "Device", "shared_with_third_parties": True},
        ],
        compliance_requirements=[
            "Purpose limitation",
            "Data minimization",
            "Third-party disclosure",
        ],
        hidden_issues=[
            {"type": "purpose_mismatch", "issue": "Health Metrics collected for 'Personalization' but shared with advertisers", "severity": "high"},
            {"type": "misleading_statement", "issue": "Says 'We do NOT share health data' but Health Metrics shared with advertising networks", "severity": "critical"},
            {"type": "purpose_limitation", "issue": "Browsing History collected for 'Improving experience' but actually used for Advertising", "severity": "high"},
        ],
    ),
    "hard": TaskConfig(
        task_id="hard_dark_patterns",
        name="Dark Pattern Detection",
        difficulty="hard",
        description="Identify contradictions and dark patterns between stated policies and actual practices",
        privacy_policy="""Privacy Policy - GlobalTech Services

SECTION 1: DATA HANDLING
We value your privacy. Your data is NEVER shared with third parties under any circumstances.
We comply with GDPR, CCPA, and all applicable privacy laws.

SECTION 2: COOKIE POLICY
Our cookies are used ONLY for essential functionality.
We do not use tracking cookies or advertising cookies.

SECTION 3: THIRD-PARTY PARTNERS
We have carefully selected these partners who comply with privacy standards:
- AdTech Solutions (advertising)
- DataBroker Inc (data analytics)
- MarketingPro (campaign tracking)
- UserInsight LLC (behavioral analysis)
- ProfileEnhancer (data enhancement)

SECTION 4: USER RIGHTS
You have the right to:
- Access your data
- Request deletion (note: processing may continue for legal obligations)

SECTION 5: CHILDREN
We do not knowingly collect data from children under 13.

Last updated: January 2024""",
        data_practices=[
            {"id": "dp1", "category": "Advertising", "purpose": "Targeted ads", "data_type": "Behavioral", "shared_with_third_parties": True},
            {"id": "dp2", "category": "Analytics", "purpose": "Usage analysis", "data_type": "Behavioral", "shared_with_third_parties": True},
            {"id": "dp3", "category": "Marketing", "purpose": "Campaign tracking", "data_type": "Contact", "shared_with_third_parties": True},
            {"id": "dp4", "category": "Behavioral", "purpose": "User profiling", "data_type": "Behavioral", "shared_with_third_parties": True},
            {"id": "dp5", "category": "Data Enhancement", "purpose": "Profile enrichment", "data_type": "Demographic", "shared_with_third_parties": True},
        ],
        compliance_requirements=[
            "No contradiction in policy",
            "Accurate third-party disclosure",
            "Genuine consent mechanisms",
        ],
        hidden_issues=[
            {"type": "contradiction", "issue": "Section 1 says 'NEVER shared with third parties' but Section 3 lists 5 third-party partners", "severity": "critical"},
            {"type": "dark_pattern", "issue": "Section 2 says 'ONLY essential cookies' but lists advertising/tracking partners", "severity": "critical"},
            {"type": "false_statement", "issue": "Section 1 says 'We do not use tracking cookies' but partners include advertising networks", "severity": "high"},
            {"type": "deceptive_rights", "issue": "Section 4 says deletion requested but 'processing may continue' - undermines right to deletion", "severity": "high"},
            {"type": "missing_disclosure", "issue": "5 data practices shared with third parties but not clearly disclosed in policy", "severity": "medium"},
        ],
    ),
    "elite": TaskConfig(
        task_id="elite_multi_doc_reasoning",
        name="Multi-Document Reasoning",
        difficulty="elite",
        description="Find contradictions requiring cross-document reasoning across 3 different documents",
        privacy_policy="""Privacy Policy - MegaCorp International

DOCUMENT A - MAIN POLICY (https://megacorp.example.com/privacy)
==============================================================
Section 1.1: Data Collection
We collect minimal personal data necessary for service delivery.
We NEVER sell your personal data to any third party.

Section 1.2: Data Usage  
Your data is used ONLY for providing the services you requested.
We do not use your data for advertising, marketing, or profiling.

DOCUMENT B - PARTNERS PAGE (https://megacorp.example.com/partners)
================================================================
Our Trusted Partners:
- AdTech Global (advertising)
- DataMine Analytics (behavioral analysis)  
- ProfileBuilders (user profiling)
- MarketingForce (campaign management)

DOCUMENT C - COOKIE POLICY (https://megacorp.example.com/cookies)
==================================================================
We use the following cookies:
- Essential cookies (login, cart)
- Advertising cookies (targeted ads based on browsing history)
- Analytics cookies (usage patterns)
- Third-party tracking cookies

Your consent is required for non-essential cookies.""",
        data_practices=[
            {"id": "dp1", "category": "Advertising", "purpose": "Targeted ads", "data_type": "Behavioral", "shared_with_third_parties": True},
            {"id": "dp2", "category": "Analytics", "purpose": "Behavior analysis", "data_type": "Behavioral", "shared_with_third_parties": True},
            {"id": "dp3", "category": "Profiling", "purpose": "User profiling", "data_type": "Demographic", "shared_with_third_parties": True},
            {"id": "dp4", "category": "Marketing", "purpose": "Campaign mgmt", "data_type": "Contact", "shared_with_third_parties": True},
            {"id": "dp5", "category": "Tracking", "purpose": "Cross-site tracking", "data_type": "Behavioral", "shared_with_third_parties": True},
        ],
        compliance_requirements=[
            "Consistency across all documents",
            "No selling of personal data",
            "Clear purpose limitation",
            "Transparent third-party usage",
        ],
        hidden_issues=[
            {"type": "contradiction_ab", "issue": "Doc A says 'NEVER sell data' but Doc B lists 4 advertising/analytics partners", "severity": "critical", "requires": ["A", "B"]},
            {"type": "contradiction_ac", "issue": "Doc A says 'used ONLY for service delivery' but Doc C lists advertising cookies", "severity": "critical", "requires": ["A", "C"]},
            {"type": "contradiction_abc", "issue": "Complete contradiction: A says no advertising, B has advertising partners, C has advertising cookies", "severity": "critical", "requires": ["A", "B", "C"]},
            {"type": "false_statement", "issue": "Doc A says 'minimal data necessary' but Doc B/C show extensive data sharing", "severity": "high", "requires": ["A", "B"]},
            {"type": "misleading_consent", "issue": "Doc C says 'consent required' but Doc A implies data used for requested services only", "severity": "high", "requires": ["A", "C"]},
            {"type": "hidden_practice", "issue": "5 data practices with third parties completely contradicts 'never sell' statement", "severity": "critical", "requires": ["A", "B", "C"]},
        ],
    ),
}


@dataclass
class EpisodeState:
    task_config: TaskConfig
    documents: List[Document]
    data_practices: List[DataPractice]
    flagged_issues: List[str]
    found_issues: List[str]
    steps: int = 0
    episode_id: str = field(default_factory=lambda: str(uuid4()))


class GDPRAuditorEnvironment:
    """GDPR Compliance Auditor Environment.
    
    The agent acts as a Data Protection Officer auditing privacy policies.
    """
    
    def __init__(self, max_steps: int = 8):
        self._max_steps = max_steps
        self._ep: Optional[EpisodeState] = None
        
    # Map full task IDs (from openenv.yaml) → short keys used in TASKS dict
    _TASK_ID_MAP = {
        "easy_clause_existence":    "easy",
        "medium_purpose_mapping":   "medium",
        "hard_dark_patterns":       "hard",
        "elite_multi_doc_reasoning": "elite",
    }

    def reset(
        self,
        seed: Optional[int] = None,
        episode_id: Optional[str] = None,
        task_id: Optional[str] = None,
        **kwargs: Any,
    ) -> ObsModel:
        if seed is not None:
            random.seed(seed)

        # Accept both short keys ("easy") and full IDs ("easy_clause_existence")
        raw_key = task_id or random.choice(["easy", "medium", "hard"])
        task_key = self._TASK_ID_MAP.get(raw_key, raw_key)  # resolve full→short
        if task_key not in TASKS:
            task_key = "easy"  # safe fallback
            
        task = TASKS[task_key]
        
        documents = [
            Document(
                id="privacy_policy",
                title="Privacy Policy",
                content=task.privacy_policy,
                doc_type="policy"
            )
        ]
        
        data_practices = [
            DataPractice(**dp) for dp in task.data_practices
        ]
        
        self._ep = EpisodeState(
            task_config=task,
            documents=documents,
            data_practices=data_practices,
            flagged_issues=[],
            found_issues=[],
            steps=0,
            episode_id=episode_id or str(uuid4()),
        )
        
        return self._build_observation("Review the privacy policy and data practices. Identify compliance issues.")
    
    def step(self, action: ActModel, **kwargs: Any) -> Tuple[ObsModel, RewModel, bool, Dict]:
        if self._ep is None:
            return (
                self._error_obs("Environment not reset"),
                RewModel(value=0.001, reason="Environment not initialized", issues_found=0, total_issues=0),
                True,
                {"error": "Environment not reset. Call /reset first."},
            )
            
        self._ep.steps += 1
        msg = action.message.lower()
        
        found_issue = self._parse_and_record_finding(msg)
        
        if found_issue:
            if found_issue not in self._ep.found_issues:
                self._ep.found_issues.append(found_issue)
        
        reward = self._calculate_reward()
        
        done = (
            self._ep.steps >= self._max_steps or
            reward.value >= 0.95
        )
        
        obs = self._build_observation(f"Issue recorded: {found_issue or 'No valid finding'}")
        return obs, reward, done, {"found_issues": len(self._ep.found_issues)}
    
    def _parse_and_record_finding(self, msg: str) -> Optional[str]:
        task = self._ep.task_config
        issues = task.hidden_issues
        
        for issue in issues:
            issue_text = issue.get("issue", "").lower()
            issue_type = issue.get("type", "")
            
            # --- Easy task: missing clause detection ---
            if issue_type == "missing_clause":
                expected = issue.get("expected", "").lower()
                if expected in msg:
                    return f"MISSING_CLAUSE: {issue.get('expected')}"
            
            # --- Medium task: purpose mismatch / misleading / purpose limitation ---
            elif issue_type in ["purpose_mismatch", "misleading_statement", "purpose_limitation"]:
                keywords = [w.lower() for w in issue_text.split() if len(w) > 4]
                matches = sum(1 for kw in keywords if kw in msg)
                if matches >= 2:
                    return f"POLICY_VIOLATION: {issue.get('severity')}"
            
            # --- Elite task: multi-doc contradiction types (must come BEFORE generic) ---
            elif issue_type in ["contradiction_ab", "contradiction_ac", "contradiction_abc"]:
                if any(word in msg for word in ["never", "sell", "advertising", "partner", "contradict", "doc a", "doc b", "doc c", "document"]):
                    return f"MULTI_DOC_CONTRADICTION: {issue.get('severity')}"
            
            # --- Elite task: misleading consent ---
            elif issue_type == "misleading_consent":
                if any(word in msg for word in ["consent", "cookie", "non-essential", "service", "implies", "mislead"]):
                    return f"MISLEADING_CONSENT: {issue.get('severity')}"
            
            # --- Elite task: hidden practice ---
            elif issue_type == "hidden_practice":
                if any(word in msg for word in ["sell", "share", "third", "advertising", "track", "never", "practice"]):
                    return f"HIDDEN_PRACTICE: {issue.get('severity')}"
            
            # --- Hard task: contradiction / dark pattern / false statement ---
            elif issue_type in ["contradiction", "dark_pattern"]:
                if "never" in msg or "not" in msg or "contradict" in msg or "section" in msg:
                    return f"CONTRADICTION: {issue.get('severity')}"
            
            elif issue_type == "false_statement":
                if any(word in msg for word in ["never", "not", "contradict", "false", "minimal", "necessary", "tracking"]):
                    return f"FALSE_STATEMENT: {issue.get('severity')}"
            
            # --- Shared: deceptive rights ---
            elif issue_type == "deceptive_rights":
                if "deletion" in msg or "delete" in msg or "right" in msg or "continue" in msg:
                    return f"DECEPTIVE_CLAUSE: {issue.get('severity')}"
            
            # --- Shared: missing disclosure ---
            elif issue_type == "missing_disclosure":
                if "third" in msg or "partner" in msg or "disclose" in msg or "shared" in msg:
                    return f"MISSING_DISCLOSURE: {issue.get('severity')}"
        
        # Fallback: generic finding if agent mentions relevant terms
        if any(word in msg for word in ["issue", "violation", "problem", "concern", "missing", "contradict"]):
            return "GENERAL_FINDING"
        
        return None
    
    def _calculate_reward(self) -> RewModel:
        # Scores must be strictly in (0, 1) and visible in :.4f format
        # 1e-6 formats as "0.0000" which the validator reads as 0.0 — use 0.001 minimum
        _MIN_SCORE = 0.001
        _MAX_SCORE = 0.999

        task = self._ep.task_config
        total_issues = len(task.hidden_issues)
        found_count = len(self._ep.found_issues)
        
        base_score = found_count / total_issues if total_issues > 0 else 0.0
        
        severity_bonus = 0.0
        critical_found = any("critical" in f.lower() for f in self._ep.found_issues)
        high_found = any("high" in f.lower() for f in self._ep.found_issues)
        
        if critical_found:
            severity_bonus += 0.25
        if high_found:
            severity_bonus += 0.15
        
        multi_doc_bonus = 0.0
        if task.difficulty == "elite":
            multi_doc_found = any("multi_doc" in f.lower() for f in self._ep.found_issues)
            if multi_doc_found:
                multi_doc_bonus += 0.2
        
        exploration_bonus = min(self._ep.steps * 0.02, 0.1)
        
        raw_reward = base_score + severity_bonus + multi_doc_bonus + exploration_bonus
        # Clamp to strictly open interval (0, 1)
        total_reward = max(_MIN_SCORE, min(_MAX_SCORE, raw_reward))
        
        reason = f"Found {found_count}/{total_issues} issues"
        
        return RewModel(
            value=total_reward,
            reason=reason,
            issues_found=found_count,
            total_issues=total_issues,
        )
    
    def _build_observation(self, message: str) -> ObsModel:
        if self._ep is None:
            return self._error_obs()
            
        return ObsModel(
            task_id=self._ep.task_config.task_id,
            task_name=self._ep.task_config.name,
            difficulty=self._ep.task_config.difficulty,
            step=self._ep.steps,
            documents=self._ep.documents,
            data_practices=self._ep.data_practices,
            compliance_requirements=self._ep.task_config.compliance_requirements,
            flagged_issues=self._ep.found_issues,
            echoed_message=message,
        )
    
    def _error_obs(self, message: str = "Error: Environment not initialized") -> ObsModel:
        return ObsModel(
            task_id="",
            task_name="",
            difficulty="",
            step=0,
            documents=[],
            data_practices=[],
            compliance_requirements=[],
            flagged_issues=[],
            echoed_message=message,
        )
    
    def state(self) -> Dict[str, Any]:
        if self._ep is None:
            return {}
        
        return {
            "episode_id": self._ep.episode_id,
            "task_id": self._ep.task_config.task_id,
            "task_name": self._ep.task_config.name,
            "difficulty": self._ep.task_config.difficulty,
            "steps": self._ep.steps,
            "found_issues": self._ep.found_issues,
            "total_issues": len(self._ep.task_config.hidden_issues),
        }


Environment = GDPRAuditorEnvironment