muse-agent-example/tests/单元/test_文学报告合同.py

"""文学片的引文/依据/未知判断合同；纯组装不能声称模型执行。"""

from copy import deepcopy

import pytest

from muse.审校修订.接口 import (
    审校错误,
    核验文学片,
    生成审校分片,
    组装文学报告,
)


def 结果(片, *, 有声音=True):
    return {
        "shard_id": 片.片ID,
        "dimensions": {
            dim: {"status": status, "summary": "本片判断", "findings": []}
            for dim, status in [
                ("continuity", "reviewed"),
                ("craft", "reviewed"),
                ("voice", "reviewed" if 有声音 else "uncalibrated"),
            ]
        },
    }


def 问题(*, start=0, end=1, quote="甲", judgment="ask", sources=()):
    return {
        "start": start,
        "end": end,
        "quote": quote,
        "judgment": judgment,
        "reason": "需要结合语境判断。",
        "basis_ids": list(sources),
    }


def test_零命中报告仍保全范围和未标定声音__22f001():
    文本 = "甲走出门。\r\n乙仍在屋内。"
    片 = 生成审校分片(文本, "a" * 64, 核心上限=8, 上下文字数=2)
    报告 = 组装文学报告(
        文本,
        "a" * 64,
        片,
        [结果(p, 有声音=False) for p in 片],
        允许依据=("fact:1",),
        已确认声音=False,
    )
    assert 报告["coverage"]["范围完整"]
    assert 报告["counts"] == {"repair": 0, "retain": 0, "ask": 0}
    assert 报告["dimension_status"]["voice"] == "uncalibrated"
    assert 报告["persisted"] is False and 报告["model_verified"] is False
    assert "pass" not in 报告 and 报告["unreviewed_shards"] == []


def test_未审尾部及缺依据不能变整章通过__22f002():
    文本 = "甲乙丙丁戊己庚辛"
    片 = 生成审校分片(文本, "a" * 64, 核心上限=4, 上下文字数=1)
    一 = 结果(片[0])
    一["dimensions"]["continuity"]["status"] = "insufficient_evidence"
    报告 = 组装文学报告(文本, "a" * 64, 片, [一], 允许依据=(), 已确认声音=True)
    assert 报告["coverage"]["未审区"] == [[4, 8]] and not 报告["coverage"]["范围完整"]
    assert 报告["unreviewed_shards"] == [片[1].片ID]
    assert 报告["dimension_status"]["continuity"] == "partial"
    assert 报告["shards"][0]["dimensions"]["continuity"]["status"] == "insufficient_evidence"


def test_重复文本按实际位置核引文且修留问分开__22f003():
    文本 = "甲说。\n甲说。"
    片 = 生成审校分片(文本, "a" * 64, 核心上限=4, 上下文字数=2)
    二 = 结果(片[1])
    二["dimensions"]["craft"]["findings"] = [问题(start=4, end=6, quote="甲说", judgment="retain")]
    二["dimensions"]["continuity"]["findings"] = [
        问题(start=4, end=5, quote="甲", judgment="repair", sources=("fact:1",))
    ]
    校验 = 核验文学片(片[1], 二, 允许依据=("fact:1",), 已确认声音=True)
    assert 校验["dimensions"]["craft"]["findings"][0]["start"] == 4
    报告 = 组装文学报告(
        文本, "a" * 64, 片, [结果(片[0]), 二], 允许依据=("fact:1",), 已确认声音=True
    )
    assert 报告["counts"] == {"repair": 1, "retain": 1, "ask": 0}
    assert len({x["finding_id"] for x in 报告["findings"]}) == 2


@pytest.mark.parametrize(
    "变动",
    [
        "错引文",
        "上下文冒核心",
        "越界",
        "错片",
        "重复",
        "陌生依据",
        "未知处置",
        "未知维度",
        "缺维度",
        "额外分数",
        "无依据修",
        "未标定修",
        "缺声音",
    ],
)
def test_文学片拒绝错误归属与强判__22f004(变动):
    文本 = "甲乙丙丁戊己庚辛"
    片 = 生成审校分片(文本, "a" * 64, 核心上限=4, 上下文字数=2)[1]
    r = 结果(片)
    f = 问题(start=4, end=5, quote="戊")
    r["dimensions"]["continuity"]["findings"] = [f]
    有声音 = True
    if 变动 == "错引文":
        f["quote"] = "甲"
    elif 变动 == "上下文冒核心":
        f.update(start=2, end=3, quote="丙")
    elif 变动 == "越界":
        f.update(end=10, quote="戊己庚辛")
    elif 变动 == "错片":
        r["shard_id"] = "b" * 64
    elif 变动 == "重复":
        r["dimensions"]["continuity"]["findings"].append(deepcopy(f))
    elif 变动 == "陌生依据":
        f["basis_ids"] = ["foreign"]
    elif 变动 == "未知处置":
        f["judgment"] = "must_rewrite"
    elif 变动 == "未知维度":
        r["dimensions"]["human_probability"] = {}
    elif 变动 == "缺维度":
        del r["dimensions"]["craft"]
    elif 变动 == "额外分数":
        r["score"] = 100
    elif 变动 == "无依据修":
        f["judgment"] = "repair"
    elif 变动 == "未标定修":
        f.update(judgment="repair", basis_ids=["fact:1"])
        r["dimensions"]["continuity"]["status"] = "uncalibrated"
    else:
        有声音 = False
    with pytest.raises(审校错误):
        核验文学片(片, r, 允许依据=("fact:1",), 已确认声音=有声音)


@pytest.mark.parametrize("变动", ["重复片", "陌生片", "旧正文", "旧依据"])
def test_报告不能复用旧片结论或伪造完成声明__22f005(变动):
    文本 = "甲乙丙丁"
    片 = 生成审校分片(文本, "a" * 64)
    rs = [结果(片[0])]
    依据 = "a" * 64
    if 变动 == "重复片":
        rs *= 2
    elif 变动 == "陌生片":
        rs[0]["shard_id"] = "b" * 64
    elif 变动 == "旧正文":
        文本 = "甲乙丁丙"
    else:
        依据 = "b" * 64
    with pytest.raises(审校错误):
        组装文学报告(文本, 依据, 片, rs, 允许依据=(), 已确认声音=True)


@pytest.mark.parametrize("变动", ["空结果", "重复依据", "空依据", "缺片ID"])
def test_未审报告也不能接受不明确的授权声明__22f006(变动):
    片 = 生成审校分片("甲乙", "a" * 64)
    参数 = {"允许依据": (), "已确认声音": True}
    结果表 = []
    if 变动 == "空结果":
        参数["已确认声音"] = 1
    elif 变动 == "重复依据":
        参数["允许依据"] = ("fact:1", "fact:1")
    elif 变动 == "空依据":
        参数["允许依据"] = ("",)
    else:
        结果表 = [{}]
    with pytest.raises(审校错误):
        组装文学报告("甲乙", "a" * 64, 片, 结果表, **参数)


@pytest.mark.parametrize("非标准值", ["NaN", "Infinity", "-Infinity", "1e999"])
def test_非标准JSON数值不产生可哈希的模型最终交付__22f007(非标准值):
    from muse.任务运行.接口 import 校验模型输出, 模型协议错误, 模型结果, 模型请求

    请求 = 模型请求(
        "call",
        "synthetic",
        "synthetic-model",
        "固定结构",
        "合成输入",
        {"type": "object", "properties": {"n": {"type": "number"}}, "required": ["n"]},
        50,
        10,
    )
    响应 = 模型结果("completed", '{"n":' + 非标准值 + "}", "synthetic-model", None)
    with pytest.raises(模型协议错误):
        校验模型输出(请求, 响应)