o2o-plagiarism-ai/tests/test_legal_risk.py

import json
from types import SimpleNamespace

from app.engine.legal_risk import (
    LegalJudgeDecision,
    LegalRiskEngine,
    Precedent,
    load_precedents,
)
from app.engine.openai_legal_judge import OpenAILegalJudge


def _case() -> Precedent:
    return Precedent(
        case_id="case-1",
        title="테스트 판례",
        source_url="https://example.test/case-1",
        work_types=("literary",),
        legal_tags=("reproduction",),
        criteria=("실질적 유사성",),
        holding_summary="테스트 요약",
    )


def test_empty_precedents_never_claim_legal_conclusion():
    result = LegalRiskEngine([]).assess(
        max_similarity=0.95, coverage=0.8, longest_span=300,
        legal_tags=["reproduction"],
    )
    assert result.status == "insufficient_precedent_data"
    assert result.risk_level is None


def test_risk_requires_missing_facts_and_only_registered_case_ids():
    result = LegalRiskEngine([_case()]).assess(
        max_similarity=0.9, coverage=0.5, longest_span=200,
        legal_tags=["reproduction"],
    )
    assert result.status == "review_required"
    assert result.risk_level == "high"
    assert result.precedent_ids == ("case-1",)
    assert len(result.missing_factors) == 3


def test_repository_seed_is_valid():
    cases = load_precedents("data/precedents/precedents.jsonl")
    assert cases
    assert any(case.case_id == "2012다73493" for case in cases)
    assert len(cases) == len({case.case_id for case in cases})


class _FakeJudge:
    def __init__(self, case_ids=("case-1",), review_required=True):
        self.case_ids = case_ids
        self.review_required = review_required
        self.request = None

    def judge(self, request):
        self.request = request
        return LegalJudgeDecision(
            verdict="likely",
            confidence=0.82,
            matched_precedent_ids=self.case_ids,
            supporting_reasons=("표현 일치 구간이 길다",),
            counter_reasons=("접근 가능성은 확인되지 않았다",),
            missing_factors=("시장 영향",),
            review_required=self.review_required,
            model="test-gpt",
            prompt_version="test-v1",
        )


def test_llm_judge_augments_rule_result_and_keeps_deterministic_missing_factors():
    judge = _FakeJudge()
    result = LegalRiskEngine([_case()], judge=judge).assess(
        max_similarity=0.9,
        coverage=0.5,
        longest_span=200,
        legal_tags=["reproduction"],
        evidence=[{"source_document_id": "doc-1", "excerpts": ["일치 문장"]}],
    )
    assert result.judgment_method == "llm"
    assert result.llm_verdict == "likely"
    assert result.llm_confidence == 0.82
    assert result.llm_matched_precedent_ids == ("case-1",)
    assert "case-1 판례" in result.judgment_summary
    assert "저작권 침해가 의심" in result.judgment_summary
    assert "GPT" not in (result.judge_note or "")
    assert "보호되는 창작적 표현인지에 대한 사람 검토" in result.missing_factors
    assert "시장 영향" in result.missing_factors
    assert judge.request.evidence[0]["source_document_id"] == "doc-1"


def test_llm_hallucinated_precedent_id_falls_back_to_rules():
    result = LegalRiskEngine([_case()], judge=_FakeJudge(("invented-case",))).assess(
        max_similarity=0.9,
        coverage=0.5,
        longest_span=200,
        legal_tags=["reproduction"],
    )
    assert result.judgment_method == "rule_fallback"
    assert result.llm_verdict is None
    assert result.precedent_ids == ("case-1",)
    assert "GPT" not in (result.judge_note or "")


def test_server_keeps_human_review_even_if_judge_returns_false():
    result = LegalRiskEngine([_case()], judge=_FakeJudge(review_required=False)).assess(
        max_similarity=0.9,
        coverage=0.5,
        longest_span=200,
        legal_tags=["reproduction"],
    )
    assert result.judgment_method == "llm"
    assert result.llm_review_required is True


def test_likely_judgment_without_precedent_id_falls_back():
    result = LegalRiskEngine([_case()], judge=_FakeJudge(())).assess(
        max_similarity=0.9,
        coverage=0.5,
        longest_span=200,
        legal_tags=["reproduction"],
    )
    assert result.judgment_method == "rule_fallback"


def test_openai_judge_uses_strict_schema_and_disables_response_storage():
    captured = {}

    class _Responses:
        def create(self, **kwargs):
            captured.update(kwargs)
            return SimpleNamespace(output_text=json.dumps({
                "verdict": "insufficient_evidence",
                "confidence": 0.35,
                "matched_precedent_ids": ["case-1"],
                "supporting_reasons": [],
                "counter_reasons": ["법적 맥락 부족"],
                "missing_factors": ["의거관계"],
                "review_required": True,
            }, ensure_ascii=False))

    client = SimpleNamespace(responses=_Responses())
    fake = _FakeJudge()
    engine = LegalRiskEngine([_case()], judge=fake)
    engine.assess(
        max_similarity=0.9, coverage=0.5, longest_span=200,
        legal_tags=["reproduction"],
    )
    judge = OpenAILegalJudge(api_key="unused", model="test-gpt", client=client)
    decision = judge.judge(fake.request)

    assert decision.verdict == "insufficient_evidence"
    assert captured["store"] is False
    assert captured["text"]["format"]["type"] == "json_schema"
    assert captured["text"]["format"]["strict"] is True