Files
breakpilot-compliance/backend-compliance/compliance/services/retention_comparator.py
T
Benjamin Admin d0e3621192 feat(audit): V2 mail render + 5 new findings (B4/B5/B6/B7/B8) + LLM-Plausibility-Phase
Mail Render V2 (compliance/services/mail_render_v2/) — 11-Modul-Subpackage
das einen einheitlichen Audit-Mail-Output erzeugt mit:
  - Header + KPI-Kacheln (Score / Findings / Docs / Vendors)
  - TOC + Sprung-Links
  - 3-Bucket-Trennung: Kritische Befunde / Manuelle Prüfung / Interne Reminder
  - Cookie-Inventar (Name·Vendor·Kategorie·Speicherdauer·Löschfrist·Sitzland·Quelle·Status)
  - Sofortmaßnahmen-Aggregator ("Sitzland ergänzen für 11 Cookies")
  - 24 Legacy-Wrappers — alle alten build_*_html in V2-Sections
  - Scope-Filter: FIN/GOV/MED/INS/EDU/LEG aus Berichten wenn nicht relevant
  - Hint/Action-Dedup: keine doppelten Sätze pro Card mehr
Aktiviert via env MAIL_RENDER_V2=true (Default: legacy renderer).

5 neue deterministische Findings als Phase D-2b/B4/B5/B6/B7/B8:

  B4 vendor_consistency_check — Cross-Doc-Provider-Widerspruch
     (Elli: DSE nennt Vertex AI für Chatbot, /de/cookies nennt Iadvize → HIGH).
     6 Service-Types: chatbot/analytics/tag_manager/pixel/cdn/cmp.

  B5 ai_act_transparency_check — AI Act Art. 50 Transparenzpflicht
     (Elli: Vertex AI vorhanden ohne Pre-Chat-Disclosure → HIGH).
     Plus B5-Erweiterung: Rechtsgrundlage Art-6-Abs-1-lit-f bei AI → MED
     (Einwilligung empfehlen).

  B6 cross_doc_dpo_check — DPO in DSE genannt, nicht im Impressum (LOW).

  B7 doc_staleness_check — Datum-Extraktion aus DSE/AGB/Nutzungsbedingungen.
     Cap: AGB/NB 3y, DSE 2y. Älter → MEDIUM (Elli NB Stand 2018 → HIGH).

  B8 cmp_fingerprint_check — Banner detected, aber CMP-Provider generic
     (kein Usercentrics/OneTrust/Cookiebot/etc → MED).

  B3-Erweiterung detect_intra_doc_contradictions — Widersprüchliche
     Speicherdauer im SELBEN Doc (Elli: Logfile 7d vs 30d → HIGH).

LLM-Plausibility-Phase (Phase D-2b, finding_plausibility_check.py):
  - Läuft AFTER MC pipeline, BEFORE D3 render
  - Prompt mit Beispiel-IDs + 3-Phase-Mapping: exact-ID / position-fallback /
    fuzzy-tail-match
  - Stempelt llm_title / llm_severity / llm_recommendation / llm_drop auf
    jeden FAIL CheckItem
  - V2-Render zeigt "🤖 LLM-Plausibility:" Box pro Finding wenn gestempelt
  - KNOWN ISSUE: qwen3:30b-a3b liefert oft empty content auf format='json' +
    8000-char-excerpt prompts. Pipeline läuft mit stamped=0 weiter. Task #16.

Coverage gegen Elli Ground Truth (zeroclaw/docs/ground-truth/elli_eco_2026-06-06.json,
13 expected findings via WebFetch-Agent-Crawl):
  - 4/4 HIGH-Findings ✓ (COOKIE-CONSENT-UX-001 + WIDERRUFSBELEHRUNG-001 +
    VENDOR-CONSISTENCY-001 + AI-ACT-TRANSPARENCY-001)
  - 4/6 MEDIUM ✓
  - 2/3 LOW ✓
  - Total: 10/13 = 77% (Sprung von 4/13 = 31%)

Restliche 3 Gaps als Task #17: IMPRESSUM-001 (multi-entity USt-IdNr),
TRANSFER-001 (Vendor-Mechanismus DPF/SCC), TH-RETENTION-002 (AI-Retention
pro Datenkategorie).

V2-Mail-Preview in Mailpit: 'v2all@local.test' Subject '[V2 ALL] ELLI'.
Backend healthy, B1+B3+B4+B5+B6+B7+B8 alle live im Orchestrator.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
2026-06-06 21:19:49 +02:00

432 lines
15 KiB
Python

"""
B3 — Cross-Doc Retention Consistency Comparator.
Compares three sources of truth for cookie storage duration:
1. DSI claim — sentence(s) in the privacy policy mentioning retention
("Die Speicherdauer beträgt 14 Monate", "_ga: 14 Monate", ...).
2. Cookie-table — the `duration` field parsed from the cookie policy
table (parse_flat_cookie_text / OCR / vendor-extract).
3. Actual cookie — `Max-Age` / `Expires` from the real Set-Cookie
header captured by the consent-tester.
Output is a per-cookie finding usable by the audit report:
- matches=True → all three sources agree (within tolerance)
- matches=False → mismatch with explicit type + severity_reason
Severity hierarchy (see project_audit_report_architecture.md):
HIGH/factually_wrong : DSI claim is shorter than reality
→ user is told "X" but tracked for longer
HIGH/factually_wrong : table duration is shorter than reality
→ cookie table understates what is set
MEDIUM/misclassified : DSI is shorter than table (internal docs disagree)
LOW/incomplete : only one source has data
The module is pure (no DB, no network) and meant to be called from the
report pipeline after cookies+DSI+HAR have already been collected.
"""
from __future__ import annotations
import logging
import re
from dataclasses import dataclass
logger = logging.getLogger(__name__)
# 5% tolerance — Safari ITP, leap years, server clocks etc.
_MATCH_TOLERANCE_PCT = 5
# Multipliers in DAYS for the German + English unit vocabulary used in
# our cookie tables and policies.
_UNIT_DAYS: dict[str, float] = {
"sekunden": 1 / 86400, "sekunde": 1 / 86400, "sec": 1 / 86400, "s": 1 / 86400,
"minuten": 1 / 1440, "minute": 1 / 1440, "min": 1 / 1440,
"stunden": 1 / 24, "stunde": 1 / 24, "h": 1 / 24,
"tage": 1, "tag": 1, "d": 1, "day": 1, "days": 1,
"wochen": 7, "woche": 7, "week": 7, "weeks": 7,
"monate": 30, "monat": 30, "month": 30, "months": 30,
"jahre": 365, "jahr": 365, "year": 365, "years": 365,
}
# Phrases that mean "session" — cookie deleted when browser closes.
_SESSION_TOKENS = {
"session", "sitzung", "sitzungsdauer", "browsersitzung",
"browser session", "browsing session", "tab",
}
# Phrases that mean "persistent without explicit cap".
_NO_EXPIRY_TOKENS = {
"unbegrenzt", "unbestimmt", "kein ablaufdatum",
"no expiry", "persistent", "permanent",
}
@dataclass
class RetentionClaim:
"""One retention statement found in the DSI text."""
sentence: str
days: float | None # None for session/unknown
is_session: bool
is_persistent: bool
context_terms: list[str] # cookie names / provider names mentioned nearby
def parse_duration_to_days(text: str) -> tuple[float | None, str]:
"""Convert a duration phrase to days.
Returns (days, kind) where kind ∈
{"days", "session", "persistent", "unknown"}.
For "session" / "persistent" days is None — comparisons must
handle these as special cases, not as 0 or infinity.
"""
if text is None:
return None, "unknown"
s = text.strip().lower()
if not s:
return None, "unknown"
for tok in _SESSION_TOKENS:
if tok in s:
return None, "session"
for tok in _NO_EXPIRY_TOKENS:
if tok in s:
return None, "persistent"
# "14 Monate", "1 Jahr", "24h", "30 Tage", "365 Tage", "30d"
m = re.search(
r"(?P<num>\d+(?:[.,]\d+)?)\s*(?P<unit>"
r"sekunden?|sec|s|minuten?|min|stunden?|h|"
r"tage?|d(?:ays?)?|wochen?|weeks?|"
r"monate?|months?|jahre?|years?)\b",
s,
)
if not m:
return None, "unknown"
num = float(m.group("num").replace(",", "."))
unit = m.group("unit")
mult = _UNIT_DAYS.get(unit)
if mult is None:
return None, "unknown"
return num * mult, "days"
def max_age_to_days(max_age_seconds: int | float | None) -> float | None:
"""Convert a Set-Cookie Max-Age (in seconds) to days."""
if max_age_seconds is None:
return None
try:
return float(max_age_seconds) / 86400.0
except (TypeError, ValueError):
return None
# Sentence splitter that respects German legal text style (lots of
# semicolons + parentheses but few capitalised abbreviations).
_SENTENCE_SPLIT = re.compile(r"(?<=[.!?])\s+(?=[A-ZÄÖÜ])")
# Quick anchor terms for retention sentences.
_RETENTION_ANCHORS = (
"speicherdauer", "speicherfrist", "speicher",
"aufbewahrungsdauer", "aufbewahrungsfrist", "aufbewahr",
"löschfrist", "löschung", "gelöscht",
"gespeichert für", "wird gespeichert", "wird für", "werden für",
"in der regel", "bis zu",
"retention", "expires", "expiration", "lifetime",
"gültigkeit", "laufzeit",
)
def _looks_like_retention(sentence: str) -> bool:
s = sentence.lower()
if not any(a in s for a in _RETENTION_ANCHORS):
return False
# Need a unit token nearby — otherwise it's metadata not duration.
return bool(re.search(
r"\b\d[\d.,]*\s*("
r"sekunden?|minuten?|stunden?|tage?|wochen?|"
r"monate?|jahre?|sec|min|h|d|"
r"weeks?|months?|years?|days?)\b",
s,
))
def extract_retention_claims(
dsi_text: str,
cookie_names: list[str] | None = None,
vendor_names: list[str] | None = None,
) -> list[RetentionClaim]:
"""Find sentences in the DSI that state a retention period.
cookie_names / vendor_names attach themselves to a sentence when
they are mentioned in it; the comparator uses this to prefer the
most specific claim available for a given cookie.
"""
if not dsi_text:
return []
cookie_names = cookie_names or []
vendor_names = vendor_names or []
# Normalise — keep original case for the sentence so it can be
# cited verbatim in the audit report.
sentences = _SENTENCE_SPLIT.split(dsi_text)
claims: list[RetentionClaim] = []
for raw in sentences:
s = raw.strip()
if not s:
continue
if not _looks_like_retention(s):
continue
days, kind = parse_duration_to_days(s)
lower = s.lower()
contexts: list[str] = []
for n in cookie_names:
if n and n.lower() in lower:
contexts.append(n)
for v in vendor_names:
if v and v.lower() in lower:
contexts.append(v)
claims.append(RetentionClaim(
sentence=s[:400],
days=days,
is_session=(kind == "session"),
is_persistent=(kind == "persistent"),
context_terms=contexts,
))
return claims
def _best_dsi_claim(
claims: list[RetentionClaim],
cookie_name: str,
vendor_name: str | None,
) -> RetentionClaim | None:
"""Pick the most specific DSI claim for a given cookie.
Priority: claim that mentions the cookie name > claim that mentions
the vendor > generic (no context).
"""
if not claims:
return None
by_cookie = [c for c in claims if cookie_name and cookie_name in c.context_terms]
if by_cookie:
return by_cookie[0]
if vendor_name:
by_vendor = [c for c in claims if vendor_name in c.context_terms]
if by_vendor:
return by_vendor[0]
generic = [c for c in claims if not c.context_terms]
return generic[0] if generic else claims[0]
def _within_tolerance(a: float, b: float) -> bool:
if a == 0 and b == 0:
return True
base = max(abs(a), abs(b))
return abs(a - b) <= base * (_MATCH_TOLERANCE_PCT / 100.0)
def compare_retention(
cookie_name: str,
table_duration: str | None,
actual_max_age_seconds: int | float | None,
dsi_claims: list[RetentionClaim] | None = None,
vendor_name: str | None = None,
) -> dict:
"""Per-cookie three-way retention comparison.
Returns a finding dict suitable for the audit-report aggregator
(theme = TH-RETENTION). Output schema is stable — extending it must
be additive so existing tests stay green.
"""
table_days, table_kind = parse_duration_to_days(table_duration or "")
actual_days = max_age_to_days(actual_max_age_seconds)
dsi_claim = _best_dsi_claim(
dsi_claims or [], cookie_name, vendor_name,
)
dsi_days = dsi_claim.days if dsi_claim else None
out: dict = {
"cookie_name": cookie_name,
"vendor_name": vendor_name,
"table_duration_raw": table_duration,
"table_days": table_days,
"table_kind": table_kind,
"actual_max_age_seconds": actual_max_age_seconds,
"actual_days": actual_days,
"dsi_days": dsi_days,
"dsi_sentence": dsi_claim.sentence if dsi_claim else None,
"dsi_context_terms": dsi_claim.context_terms if dsi_claim else [],
"matches": True,
"mismatch_type": None,
"severity_reason": None,
"severity": None,
"diff_days": None,
"notes": [],
}
sources = [v for v in (table_days, actual_days, dsi_days) if v is not None]
if len(sources) <= 1:
out["severity_reason"] = "incomplete"
out["severity"] = "LOW"
out["notes"].append("only_one_source_has_data")
return out
# Highest-severity check first: DSI claim is shorter than the cookie
# actually lives — user was misled.
if dsi_days is not None and actual_days is not None:
if not _within_tolerance(dsi_days, actual_days):
if dsi_days < actual_days:
out["matches"] = False
out["mismatch_type"] = "dsi_under_actual"
out["severity_reason"] = "factually_wrong"
out["severity"] = "HIGH"
out["diff_days"] = actual_days - dsi_days
# Cookie table understates reality — second highest.
if (out["matches"] and table_days is not None
and actual_days is not None):
if not _within_tolerance(table_days, actual_days):
if table_days < actual_days:
out["matches"] = False
out["mismatch_type"] = "table_under_actual"
out["severity_reason"] = "factually_wrong"
out["severity"] = "HIGH"
out["diff_days"] = actual_days - table_days
# Internal disagreement DSI vs. table (less severe — both are
# documentation, neither contradicts the live cookie).
if (out["matches"] and dsi_days is not None and table_days is not None):
if not _within_tolerance(dsi_days, table_days):
out["matches"] = False
out["mismatch_type"] = "dsi_vs_table"
out["severity_reason"] = "misclassified"
out["severity"] = "MEDIUM"
out["diff_days"] = abs(dsi_days - table_days)
# Catch over-declaration too — table says "2 years" but cookie
# lives 7 days (Safari ITP). Less severe but worth flagging.
if (out["matches"] and table_days is not None
and actual_days is not None):
if (not _within_tolerance(table_days, actual_days)
and table_days > actual_days):
out["matches"] = False
out["mismatch_type"] = "actual_under_table"
out["severity_reason"] = "incomplete"
out["severity"] = "LOW"
out["notes"].append("possible_safari_itp_cap")
out["diff_days"] = table_days - actual_days
return out
def detect_intra_doc_contradictions(
dsi_text: str,
) -> list[dict]:
"""Find sentences in the SAME doc that claim different retention
values for what looks like the same data category.
Catches the Elli pattern:
"Logfiles werden 7 Tage gespeichert" + "Logfiles werden 30 Tage
aufbewahrt" → contradiction in one DSE.
Heuristik: group retention-bearing sentences by a category-anchor
keyword (logfile / log / chatverlauf / cookies / nutzungsdaten /
server-log) and report when ≥2 different day-values exist for the
same group.
"""
if not dsi_text:
return []
claims = extract_retention_claims(dsi_text)
if len(claims) < 2:
return []
anchors = (
("logfile", ("logfile", "log-file", "log file", "server-log")),
("chat", ("chat", "chatverlauf", "konversation")),
("cookie", ("cookie",)),
("session", ("session", "sitzung")),
("nutzungsdaten", ("nutzungsdaten", "usage data")),
)
by_group: dict[str, list[RetentionClaim]] = {}
for cl in claims:
if cl.days is None:
continue
sentence_lc = cl.sentence.lower()
for group, kws in anchors:
if any(k in sentence_lc for k in kws):
by_group.setdefault(group, []).append(cl)
break
findings: list[dict] = []
for group, group_claims in by_group.items():
days_set = {round(c.days, 1) for c in group_claims if c.days}
if len(days_set) < 2:
continue
values = sorted(days_set)
delta = values[-1] - values[0]
sev = "HIGH" if delta > values[0] * 3 else "MEDIUM"
findings.append({
"check_id": "TH-RETENTION-INTRA-001",
"category": group,
"severity": sev,
"severity_reason": "factually_wrong",
"values_days": values,
"claims": [c.sentence[:200] for c in group_claims[:3]],
"title": (
f"Speicherdauer-Widerspruch in DSE für '{group}': "
f"{values} Tage"
),
"norm": "DSGVO Art. 5 Abs. 1 lit. a (Transparenz)",
"action": (
f"In der DSE einheitlichen Wert für '{group}' angeben. "
"Aktuell mindestens zwei verschiedene Werte genannt — "
"ein Mandant kann die Frist nicht eindeutig erkennen."
),
})
return findings
def build_retention_theme_summary(
findings: list[dict],
) -> dict:
"""Aggregate per-cookie findings into the per-theme block used by
the report (theme = TH-RETENTION).
"""
total = len(findings)
incomplete = sum(
1 for f in findings if f.get("severity_reason") == "incomplete"
)
# Incomplete findings keep matches=True (we did not observe a
# mismatch), but they don't count as a verified pass either.
passed = sum(
1 for f in findings
if f.get("matches") and f.get("severity_reason") != "incomplete"
)
failed = total - passed - incomplete
by_severity: dict[str, int] = {}
by_type: dict[str, int] = {}
for f in findings:
sev = f.get("severity")
if sev:
by_severity[sev] = by_severity.get(sev, 0) + 1
mt = f.get("mismatch_type")
if mt:
by_type[mt] = by_type.get(mt, 0) + 1
return {
"theme_id": "TH-RETENTION",
"total": total,
"passed": passed,
"failed": failed,
"incomplete": incomplete,
"pct": int(round(100 * passed / total)) if total else 0,
"by_severity": by_severity,
"by_mismatch_type": by_type,
"top_fails": sorted(
(f for f in findings
if not f.get("matches")
and f.get("severity_reason") == "factually_wrong"),
key=lambda f: -(f.get("diff_days") or 0),
)[:10],
}