343 lines
16 KiB
Python
343 lines
16 KiB
Python
import copy
|
|
import json
|
|
from pathlib import Path
|
|
|
|
import pytest
|
|
from pydantic import ValidationError
|
|
from test_support.external_qa import require_external_qa
|
|
|
|
from yuqei_sdk.contract_drafting import (
|
|
ContractDraftOutput, ContractDraftRequest, draft_digest, draft_output_schema,
|
|
validate_draft_output,
|
|
)
|
|
from yuqei_sdk.draft_change_evidence import (
|
|
CHANGE_EVIDENCE_POLICY, DraftChangeDeclarations, assess_draft_changes,
|
|
build_change_evidence_context,
|
|
)
|
|
|
|
|
|
BASE = "第一条：价款为100元。\n\n第二条：通知应书面送达。"
|
|
|
|
|
|
def request(*, enabled=True, base=BASE):
|
|
context = build_change_evidence_context(base)
|
|
snapshot = {"id": "tpl-1", "clause_assembly": {"items": [{"clause_id": "price"}]}}
|
|
if enabled:
|
|
snapshot.update(change_evidence_policy=CHANGE_EVIDENCE_POLICY,
|
|
change_evidence_context=context)
|
|
return ContractDraftRequest(contract_title="测试", contract_type="service", base_document=base,
|
|
facts=[{"name": "价款", "value": "100元"}], template_snapshot=snapshot,
|
|
playbook_snapshot={"rules": [{"id": "rule-1"}]},
|
|
retrieved_sources=[{"id": "source-1", "title": "来源", "content": "来源正文"}])
|
|
|
|
|
|
def output(body, declarations=None):
|
|
return ContractDraftOutput(schema_version="contract-draft.v1", status="candidate_ready",
|
|
draft_document=body, missing_fields=[], assumptions=[], risk_notes=[],
|
|
fact_bindings=[{"field": "价款", "value": "100元", "document_excerpt": "100元"}],
|
|
source_bindings=[], change_explanations=["说明"], quality_findings=[],
|
|
change_declarations=declarations)
|
|
|
|
|
|
def declaration(**updates):
|
|
payload = {"schema_version": "draft-change-declarations.v1", "items": [{
|
|
"id": "change-1", "target_id": "base-0002", "operation": "replace",
|
|
"before_excerpt": "书面送达", "after_excerpt": "电子邮件送达",
|
|
"basis": [{"type": "ai_suggestion"}], "summary": "调整通知方式",
|
|
"pending_confirmation": True,
|
|
}]}
|
|
payload["items"][0].update(updates)
|
|
return payload
|
|
|
|
|
|
def test_context_covers_every_codepoint_and_is_deterministic():
|
|
context = build_change_evidence_context(BASE)
|
|
assert "".join(item["text"] for item in context["targets"]) == BASE
|
|
assert [(item["start"], item["end"]) for item in context["targets"]] == [(0, 14), (14, len(BASE))]
|
|
assert context == build_change_evidence_context(BASE)
|
|
assert context["range_unit"] == "unicode_codepoint_half_open"
|
|
|
|
|
|
def test_verified_replace_produces_server_hashes_ranges_and_false_disclaimers():
|
|
candidate = BASE.replace("书面送达", "电子邮件送达")
|
|
assessment = assess_draft_changes(request(), output(candidate, declaration()))
|
|
assert assessment.status == "complete"
|
|
assert assessment.items[0].status == "verified"
|
|
assert assessment.items[0].before_range is not None and assessment.items[0].after_range is not None
|
|
assert assessment.uncovered_diff_ranges == []
|
|
assert assessment.semantic_equivalence_verified is False
|
|
assert assessment.legal_correctness_verified is False
|
|
assert assessment.authorizes_adoption is False
|
|
|
|
|
|
@pytest.mark.parametrize("candidate", [
|
|
BASE.replace("通知应书面送达", "通知应书面或电子邮件送达"),
|
|
BASE.replace("通知应书面送达", "通知应送达"),
|
|
])
|
|
def test_whole_paragraph_replace_accepts_internal_pure_insert_or_delete(candidate):
|
|
evidence = declaration(
|
|
before_excerpt="第二条：通知应书面送达。",
|
|
after_excerpt=candidate.split("\n\n", 1)[1],
|
|
basis=[{"type": "template", "ref": "tpl-1"}],
|
|
pending_confirmation=False,
|
|
)
|
|
assessment = assess_draft_changes(request(), output(candidate, evidence))
|
|
assert assessment.status == "complete"
|
|
assert assessment.items[0].status == "verified"
|
|
assert assessment.uncovered_diff_ranges == []
|
|
|
|
|
|
def test_saved_real_whole_paragraph_replacements_bind_to_their_internal_edits():
|
|
require_external_qa(
|
|
"multiscene-closeout-20260910/draft-timeout-diagnostic-real-v1/01-draft-before.json",
|
|
"multiscene-closeout-20260910/draft-timeout-diagnostic-real-v1/01-draft-after.json",
|
|
)
|
|
root = Path(__file__).resolve().parents[2]
|
|
evidence_dir = root / "docs/qa/multiscene-closeout-20260910/draft-timeout-diagnostic-real-v1"
|
|
before = json.loads((evidence_dir / "01-draft-before.json").read_text(encoding="utf-8"))
|
|
after = json.loads((evidence_dir / "01-draft-after.json").read_text(encoding="utf-8"))
|
|
req = ContractDraftRequest.model_validate(before["request"])
|
|
out = ContractDraftOutput.model_validate(after["job"]["result"]["output"])
|
|
|
|
assessment = assess_draft_changes(req, out)
|
|
affected = {item.declaration_id: item for item in assessment.items
|
|
if item.declaration_id in {"chg-0003", "chg-0005"}}
|
|
assert set(affected) == {"chg-0003", "chg-0005"}
|
|
assert all(item.status == "verified" for item in affected.values())
|
|
assert all("replace_ranges_not_same_hunk" not in item.errors for item in affected.values())
|
|
|
|
|
|
def test_whole_paragraph_replace_cannot_bind_an_edit_from_an_unrelated_target():
|
|
candidate = BASE.replace("价款为100元", "价款为200元")
|
|
evidence = declaration(
|
|
before_excerpt="第二条：通知应书面送达。",
|
|
after_excerpt="第一条：价款为200元。",
|
|
basis=[{"type": "template", "ref": "tpl-1"}],
|
|
pending_confirmation=False,
|
|
)
|
|
assessment = assess_draft_changes(request(), output(candidate, evidence))
|
|
assert assessment.status == "invalid"
|
|
assert assessment.items[0].status == "invalid"
|
|
|
|
|
|
@pytest.mark.parametrize("operation,fields", [
|
|
("insert", {"before_excerpt": "旧", "after_excerpt": "新"}),
|
|
("delete", {"before_excerpt": "旧", "after_excerpt": "新"}),
|
|
("replace", {"before_excerpt": "相同", "after_excerpt": "相同"}),
|
|
])
|
|
def test_operation_shape_is_strict(operation, fields):
|
|
with pytest.raises(ValidationError):
|
|
DraftChangeDeclarations.model_validate(declaration(operation=operation, **fields))
|
|
|
|
|
|
def test_repeated_excerpt_requires_independent_occurrence_or_anchor():
|
|
base = "重复。重复。"
|
|
req = request(base=base)
|
|
ambiguous = output("改。重复。", declaration(target_id="base-0001",
|
|
before_excerpt="重复", after_excerpt="改", pending_confirmation=True))
|
|
assessment = assess_draft_changes(req, ambiguous)
|
|
assert assessment.status == "invalid"
|
|
assert assessment.items[0].status == "ambiguous"
|
|
resolved = ambiguous.model_copy(deep=True)
|
|
resolved.change_declarations.items[0].before_occurrence = 1
|
|
resolved.change_declarations.items[0].after_occurrence = 1
|
|
assert assess_draft_changes(req, resolved).status == "complete"
|
|
|
|
|
|
@pytest.mark.parametrize("basis", [
|
|
{"type": "fact", "ref": "missing"},
|
|
{"type": "template", "ref": "unknown"},
|
|
{"type": "playbook_rule", "ref": "unknown"},
|
|
{"type": "retrieved_source", "ref": "unknown"},
|
|
{"type": "ai_suggestion", "ref": "fake"},
|
|
])
|
|
def test_forged_basis_is_invalid_and_validate_returns_hard_error(basis):
|
|
candidate = BASE.replace("书面送达", "电子邮件送达")
|
|
out = output(candidate, declaration(basis=[basis]))
|
|
assessment = assess_draft_changes(request(), out)
|
|
assert assessment.status == "invalid"
|
|
assert "change_declaration_invalid" in assessment.errors
|
|
assert "draft_change_evidence_invalid" in [item.code for item in validate_draft_output(request(), out)]
|
|
|
|
|
|
def test_uncovered_diff_is_incomplete_warning_not_hard_error():
|
|
candidate = BASE.replace("书面送达", "电子邮件送达") + "\n新增句。"
|
|
out = output(candidate, declaration())
|
|
assessment = assess_draft_changes(request(), out)
|
|
assert assessment.status == "incomplete" and assessment.uncovered_diff_ranges
|
|
findings = validate_draft_output(request(), out)
|
|
warning = next(item for item in findings if item.code == "draft_change_evidence_incomplete")
|
|
assert warning.severity == "warning"
|
|
assert "不代表语义或法律正确性" in warning.message
|
|
|
|
|
|
def test_unknown_target_overlap_and_unrelated_insert_fail_closed():
|
|
candidate = BASE + "新增甲。新增乙。"
|
|
bad = declaration(target_id="base-9999", operation="insert", before_excerpt=None,
|
|
after_excerpt="新增甲", pending_confirmation=True)
|
|
assert assess_draft_changes(request(), output(candidate, bad)).status == "invalid"
|
|
unrelated = declaration(target_id="base-0001", operation="insert", before_excerpt=None,
|
|
after_excerpt="新增甲", pending_confirmation=True)
|
|
assert assess_draft_changes(request(), output(candidate, unrelated)).status == "invalid"
|
|
|
|
|
|
def test_context_tamper_and_missing_declarations_are_invalid():
|
|
req = request()
|
|
req.template_snapshot["change_evidence_context"]["targets"][0]["text"] = "tampered"
|
|
assessment = assess_draft_changes(req, output(BASE))
|
|
assert assessment.status == "invalid"
|
|
assert "change_evidence_context_invalid" in assessment.errors
|
|
assert "change_declarations_missing" in assessment.errors
|
|
|
|
|
|
def test_legacy_output_json_and_provider_schema_omit_new_contract():
|
|
legacy = request(enabled=False)
|
|
out = output(BASE)
|
|
dumped = out.model_dump(mode="json")
|
|
assert "change_declarations" not in dumped
|
|
assert "change_declarations" not in json.loads(out.model_dump_json())
|
|
schema = draft_output_schema()
|
|
assert "change_declarations" not in schema["properties"]
|
|
assert not any(name.startswith("DraftChange") for name in schema["$defs"])
|
|
assert draft_output_schema(legacy) == schema
|
|
|
|
|
|
def test_policy_provider_schema_requires_declarations_without_changing_legacy_schema():
|
|
legacy = copy.deepcopy(draft_output_schema())
|
|
policy = draft_output_schema(request())
|
|
assert "change_declarations" in policy["properties"]
|
|
assert "change_declarations" in policy["required"]
|
|
assert policy["properties"]["change_declarations"].get("$ref")
|
|
assert draft_output_schema() == legacy
|
|
assert draft_digest(draft_output_schema()) == draft_digest(legacy)
|
|
for definition in policy["$defs"].values():
|
|
if "properties" in definition:
|
|
assert set(definition["required"]) == set(definition["properties"])
|
|
|
|
|
|
@pytest.mark.parametrize("base,candidate,expected", [
|
|
("原文", "新增原文", [("after", 0, 2)]),
|
|
("原文", "原文新增", [("after", 2, 4)]),
|
|
("新增原文", "原文", [("before", 0, 2)]),
|
|
("原文新增", "原文", [("before", 2, 4)]),
|
|
])
|
|
def test_head_and_tail_insert_delete_ranges(base, candidate, expected):
|
|
from yuqei_sdk.draft_change_evidence import _diff_ranges
|
|
assert _diff_ranges(base, candidate) == expected
|
|
|
|
|
|
def test_malformed_context_is_invalid_but_never_traversed():
|
|
req = request()
|
|
req.template_snapshot["change_evidence_context"] = {"targets": [None, "secret"]}
|
|
candidate = BASE.replace("书面送达", "电子邮件送达")
|
|
assessment = assess_draft_changes(req, output(candidate, declaration()))
|
|
assert assessment.status == "invalid"
|
|
assert "change_evidence_context_invalid" in assessment.errors
|
|
assert assessment.items[0].status == "verified"
|
|
|
|
|
|
def test_partial_excerpt_does_not_cover_the_rest_of_one_diff_hunk():
|
|
candidate = BASE.replace("书面送达", "电子邮件及短信送达")
|
|
evidence = declaration(after_excerpt="电子邮件")
|
|
assessment = assess_draft_changes(request(), output(candidate, evidence))
|
|
assert assessment.status == "incomplete"
|
|
assert any(item.side == "after" and item.end > item.start
|
|
for item in assessment.uncovered_diff_ranges)
|
|
|
|
|
|
def test_cross_paragraph_and_oversized_excerpts_are_rejected():
|
|
with pytest.raises(ValidationError):
|
|
DraftChangeDeclarations.model_validate(declaration(before_excerpt="甲\n\n乙"))
|
|
with pytest.raises(ValidationError):
|
|
DraftChangeDeclarations.model_validate(declaration(before_excerpt="甲" * 2001))
|
|
|
|
|
|
@pytest.mark.parametrize("snapshot_patch,basis", [
|
|
({"clause_assembly": "not-an-object"}, {"type": "template", "ref": "None"}),
|
|
({"id": "", "clause_assembly": []}, {"type": "template", "ref": "tpl-1"}),
|
|
])
|
|
def test_arbitrary_template_snapshot_shapes_and_empty_ids_fail_closed(snapshot_patch, basis):
|
|
req = request()
|
|
req.template_snapshot.update(snapshot_patch)
|
|
candidate = BASE.replace("书面送达", "电子邮件送达")
|
|
assessment = assess_draft_changes(req, output(candidate, declaration(basis=[basis])))
|
|
assert assessment.status == "invalid"
|
|
assert "basis_invalid:template" in assessment.items[0].errors
|
|
|
|
|
|
def test_missing_and_disabled_rule_ids_are_not_valid_bases():
|
|
req = request()
|
|
req.playbook_snapshot = {"rules": [{"id": None}, {"id": "off", "enabled": False}]}
|
|
candidate = BASE.replace("书面送达", "电子邮件送达")
|
|
for ref in ("None", "off"):
|
|
assessment = assess_draft_changes(req, output(candidate,
|
|
declaration(basis=[{"type": "playbook_rule", "ref": ref}])))
|
|
assert assessment.status == "invalid"
|
|
assert "basis_invalid:playbook_rule" in assessment.items[0].errors
|
|
|
|
|
|
def test_ai_suggestion_cannot_be_silently_confirmed_and_pending_is_correlated():
|
|
candidate = BASE.replace("书面送达", "电子邮件送达")
|
|
silent = assess_draft_changes(request(), output(candidate,
|
|
declaration(pending_confirmation=False)))
|
|
assert "ai_suggestion_requires_confirmation" in silent.items[0].errors
|
|
unrelated = assess_draft_changes(request(), output(candidate,
|
|
declaration(basis=[{"type": "template", "ref": "tpl-1"}], pending_confirmation=True)))
|
|
assert "pending_confirmation_unjustified" in unrelated.items[0].errors
|
|
|
|
|
|
def test_large_distant_edits_are_refined_into_their_own_targets():
|
|
base = "开头旧甲" + ("中" * 5000) + "结尾旧乙"
|
|
candidate = "开头新甲" + ("中" * 5000) + "结尾新乙"
|
|
req = request(base=base)
|
|
changes = DraftChangeDeclarations.model_validate({
|
|
"schema_version": "draft-change-declarations.v1",
|
|
"items": [
|
|
{"id": "change-a", "target_id": "base-0001", "operation": "replace",
|
|
"before_excerpt": "旧甲", "after_excerpt": "新甲",
|
|
"basis": [{"type": "ai_suggestion"}], "summary": "调整甲", "pending_confirmation": True},
|
|
{"id": "change-b", "target_id": "base-0003", "operation": "replace",
|
|
"before_excerpt": "旧乙", "after_excerpt": "新乙",
|
|
"basis": [{"type": "ai_suggestion"}], "summary": "调整乙", "pending_confirmation": True},
|
|
],
|
|
})
|
|
assessment = assess_draft_changes(req, output(candidate, changes))
|
|
assert assessment.status == "complete"
|
|
assert [item.status for item in assessment.items] == ["verified", "verified"]
|
|
|
|
|
|
def test_user_instruction_basis_requires_an_actual_request_instruction():
|
|
candidate = BASE.replace("书面送达", "电子邮件送达")
|
|
evidence = declaration(basis=[{"type": "user_instruction"}])
|
|
absent = assess_draft_changes(request(), output(candidate, evidence))
|
|
assert "basis_invalid:user_instruction" in absent.items[0].errors
|
|
req = request()
|
|
req.drafting_instructions = "请调整通知方式。"
|
|
assert assess_draft_changes(req, output(candidate, evidence)).status == "complete"
|
|
|
|
|
|
def test_maximum_fragmented_context_ids_remain_valid_declaration_ids():
|
|
base = "x\n\n" * 33_333
|
|
context = build_change_evidence_context(base)
|
|
assert len(base) == 99_999
|
|
assert len(context["targets"]) == 33_333
|
|
assert context["targets"][-1]["target_id"] == "base-33333"
|
|
DraftChangeDeclarations.model_validate(declaration(
|
|
target_id=context["targets"][-1]["target_id"]))
|
|
|
|
|
|
def test_many_repeated_short_lines_use_bounded_token_matching():
|
|
from yuqei_sdk.draft_change_evidence import _diff_hunks
|
|
base = "相同\n" * 25_000
|
|
candidate = base[:50_000] + "改" + base[50_001:]
|
|
hunks = _diff_hunks(base, candidate)
|
|
assert hunks
|
|
assert all(0 <= left[0] <= left[1] <= len(base) for left, _ in hunks)
|
|
|
|
|
|
def test_policy_disabled_returns_none_and_does_not_validate_declarations():
|
|
legacy = request(enabled=False)
|
|
assert assess_draft_changes(legacy, output(BASE)) is None
|
|
assert not any(item.code.startswith("draft_change_evidence")
|
|
for item in validate_draft_output(legacy, output(BASE)))
|