muse-agent-example/tests/契约/test_语义检测合同.py

"""B06核原字和完整声明；纯合同验证不认证模型执行或语义正确性。"""

import copy

import pytest

from muse.审校修订.接口 import 审校错误, 核对语义检测, 组装语义检测材料


def 材料():
    return 组装语义检测材料(
        "旅者等在城门。",
        {
            "sources": [{"source_id": "outline", "kind": "fine_outline", "text": "旅者守门。"}],
            "assertions": [{"statement_id": "guard", "text": "旅者是守卫。"}],
            "constraints": [{"statement_id": "stay", "text": "不得离开城门。"}],
        },
    )


def 输出(material, script="ok"):
    quote = material["candidate"]
    result = {
        "claims": [
            {
                "claim_id": "claim-1",
                "text": "候选中人物采取行动。",
                "candidate_quote": quote,
                "state": "supported",
                "evidence_refs": ["source:outline"],
                "reason": "参照细纲核对。",
            }
        ],
        "findings": [],
        "assertion_verdicts": [
            {
                "statement_id": id_,
                "verdict": "pass",
                "candidate_quote": quote,
                "evidence_refs": ["assertion:" + id_],
                "reason": "与给定命题相符。",
            }
            for id_ in material["required_assertions"]
        ],
        "constraint_verdicts": [
            {
                "statement_id": id_,
                "verdict": "pass",
                "candidate_quote": quote,
                "evidence_refs": ["constraint:" + id_],
                "reason": "符合给定约束。",
            }
            for id_ in material["required_constraints"]
        ],
        "new_setting_candidates": [],
    }
    if script == "high":
        result["findings"] = [
            {
                "finding_id": "high-1",
                "category": "fact",
                "severity": "high",
                "candidate_quote": quote,
                "evidence_refs": ["source:outline"],
                "message": "合成事实冲突。",
            }
        ]
    elif script == "unknown":
        result["claims"][0].update(state="unknown", evidence_refs=[], reason="给定来源不足以认定。")
        result["constraint_verdicts"][0].update(verdict="unknown", reason="约束无法确定。")
        result["new_setting_candidates"] = [
            {"claim_id": "claim-1", "proposal": "待作者核对的新设定。"}
        ]
    elif script == "missing":
        result["assertion_verdicts"] = []
    return result


def test_原字定位与命题覆盖来自实际材料__25f101():
    material = 材料()
    result = 核对语义检测(material, 输出(material))
    assert result["claims"][0]["start"] == 0
    assert result["claims"][0]["end"] == len(material["candidate"])
    assert result["constraint_counts"] == {"pass": 1, "fail": 0, "unknown": 0}
    assert result["status"] == "passed"


@pytest.mark.parametrize(
    "variant", ["passed", "quote", "missing", "reference", "duplicate", "position", "external"]
)
def test_缺项陌生依据及自报通过不能伪装检测__25f102(variant):
    material = 材料()
    output = 输出(material)
    if variant == "passed":
        output["passed"] = True
    elif variant == "quote":
        output["claims"][0]["candidate_quote"] = "原文没有这句话"
    elif variant == "missing":
        output["assertion_verdicts"] = []
    elif variant == "reference":
        output["claims"][0]["evidence_refs"] = ["source:stranger"]
    elif variant == "duplicate":
        output["claims"].append(copy.deepcopy(output["claims"][0]))
    elif variant == "position":
        output["claims"][0]["start"] = 0
    else:
        output = 输出(material, "high")
        output["findings"][0]["evidence_refs"] = ["candidate"]
    with pytest.raises(审校错误):
        核对语义检测(material, output)


def test_高严重度与未知不得被零发现覆盖__25f103():
    material = 材料()
    high = 核对语义检测(material, 输出(material, "high"))
    unknown = 核对语义检测(material, 输出(material, "unknown"))
    assert high["status"] == "failed" and high["high_severity_count"] == 1
    assert unknown["status"] == "inconclusive" and unknown["unknown_count"] == 2
    assert unknown["new_setting_candidates"][0]["claim_id"] == "claim-1"


def test_声明新事实须有细纲且建议不能借用已确认陈述__25f104():
    material = 材料()
    output = 输出(material)
    output["claims"][0].update(state="declared_new", evidence_refs=["assertion:guard"])
    with pytest.raises(审校错误, match="细纲"):
        核对语义检测(material, output)
    output["claims"][0]["evidence_refs"] = ["source:outline"]
    output["new_setting_candidates"] = [{"claim_id": "claim-1", "proposal": "新设定待确认。"}]
    assert 核对语义检测(material, output)["new_setting_candidates"]
    output["claims"][0]["state"] = "supported"
    with pytest.raises(审校错误, match="未决"):
        核对语义检测(material, output)