优化:未分类连续维护与柏拉图预汇总,同步已发布修复

This commit is contained in:
2026-11-01 12:40:50 +08:00
parent 572873c6b8
commit 38290f5fec
135 changed files with 17613 additions and 663 deletions
+304
View File
@@ -0,0 +1,304 @@
"""Typed edits bind original spans to formal values, never to arbitrary prose."""
import copy
import io
import json
import sys
import threading
import unittest
from pathlib import Path
from unittest.mock import patch
sys.path[:0] = [str(Path(__file__).resolve().parents[1] / 'scripts'), str(Path(__file__).parent)]
from nao_evidence_finalizer import synthesize_native
from nao_native_answer import (_comparison_weight_changes, _attach_native_source_numbers,
_native_atomic_inventory, _native_number_match, _native_changes_weight_context, _metric_facts,
_native_black_category_share_context, _native_changes_decomposition)
from nao_native_answer import build_native_report
from test_agent_answer import evidence
class NativeAtomicBindingTests(unittest.TestCase):
def report(self):
return {'facts': [{'id':'F'+str(i),'kind':'product','label':label,
'text':label+'合格率95.00%,报废重量10.25 kg。','evidenceRef':'Q00000001',
'period':{'year':2026,'month':9},'scope':{'period':{'year':2026,'month':9},'dimension':'product','filters':{'product':label}}}
for i,label in [(1,'规格甲'),(2,'规格乙')]],'requiredFactIds':[],'priorityFactIds':[],
'missingEvidence':[],'rejectedEvidenceRefs':[],'sections':[],'notices':[]}
def invoke(self,draft,selector=None,report=None):
calls=[]
def provider(request,timeout):
payload=json.loads(json.loads(request.data)['messages'][1]['content']);calls.append(payload)
rows=[]
for block in payload['draftBlocks']:
row={'sourceId':block['sourceId'],'role':'core','factIds':['F1'],'supported':True,'statisticalClaimsSupported':True}
if selector: row.update(selector(payload,block))
rows.append(row)
raw={'reviews':rows,'missingGoalDescriptions':[]}
return io.BytesIO(json.dumps({'choices':[{'finish_reason':'stop','message':{'content':json.dumps(raw,ensure_ascii=False)}}]}).encode())
with patch('nao_native_answer.build_native_report',return_value=report or self.report()):
result=synthesize_native(base_url='https://offline.invalid',api_key='fixture',model='fixture',question='分析质量',
evidence=evidence(),context={'_agentTask':{'executionMode':'native_agent'}},native_answer=draft,
cancel_event=threading.Event(),opener=provider)
return result,calls
def binding(self,payload,span,owner='F1'):
key=next(key for key,value in payload['bindingCatalog'].items() if key.startswith('N') and value['factId']==owner and value.get('unit')=='kg')
return {'bindings':[{'span':span,'tokenId':key}]}
def test_bare_weight_and_bold_table_are_typed_once(self):
draft='规格甲报废为**10.25**。\n\n| 产品 | 9月报废 |\n|---|---|\n| 规格甲 | **10.25** |'
result,calls=self.invoke(draft)
self.assertTrue(result['evidenceSufficient'],result['compositionDiagnostics']['issues'])
self.assertEqual(len(calls),1)
self.assertIn('10.25 kg',result['answer'])
def test_atomic_span_uses_exact_source_and_qualified_token(self):
result,calls=self.invoke('规格甲报废10.25。',lambda p,b:self.binding(p,'10.25'))
self.assertTrue(result['evidenceSufficient'])
self.assertEqual(len(calls),1)
for draft,selector in [('规格甲报废10.25。',lambda p,b:self.binding(p,'10.25','F2')),
('规格甲报废10.25吨。',lambda p,b:self.binding(p,'10.25')),
('规格甲数量10.25件。',lambda p,b:self.binding(p,'10.25')),
('规格甲报废10.26。',lambda p,b:self.binding(p,'10.26')),
('规格甲A10.25质量。',lambda p,b:self.binding(p,'10.25')),
('| 序号 | 产品 |\n|---|---|\n| 10.25 | 规格甲 |',lambda p,b:self.binding(p,'10.25')),
('规格甲报废10.25。',lambda p,b:self.binding(p,'不存在的原文'))]:
with self.subTest(draft=draft):
bad,calls=self.invoke(draft,selector);self.assertFalse(bad['evidenceSufficient']);self.assertEqual(len(calls),2)
def test_repeated_span_requires_occurrence_and_wrong_month_is_rejected(self):
bad,_=self.invoke('规格甲报废10.25,复核10.25。',lambda p,b:self.binding(p,'10.25'))
self.assertFalse(bad['evidenceSufficient'])
report=self.report();report['facts'][0]['period']={'year':2026,'month':8}
bad,_=self.invoke('规格甲质量。\n\n| 产品 | 2026年9月报废 |\n|---|---|\n| 规格甲 | 10.25 |',report=report)
self.assertFalse(bad['evidenceSufficient'])
def test_formal_weight_differences_use_unrounded_values_and_missing_is_not_zero(self):
compare={'current':{'period':{'year':2026,'month':9},'black':{'scrapWeight':'38375.282'},
'coverage':{'sourceMonths':[{'year':2026,'month':9,'blackRows':2}]}},
'comparison':{'period':{'year':2026,'month':8},'black':{'scrapWeight':'32375.0762'},
'coverage':{'sourceMonths':[{'year':2026,'month':8,'blackRows':2}]}}}
changes=_comparison_weight_changes(compare,'black')
item=next(item for item in changes if item['metricCode']=='scrapWeightDelta')
self.assertEqual(item['rawValue'],'6000.2058');self.assertEqual(item['displayValue'],'+6,000.21')
compare['comparison']['black']['scrapWeight']=None
self.assertFalse(any(item['metricCode']=='scrapWeightDelta' for item in _comparison_weight_changes(compare,'black')))
compare['comparison']['black']['scrapWeight']='32375.0762'
compare['comparison']['coverage']['sourceMonths'][0]['blackRows']=0
self.assertEqual(_comparison_weight_changes(compare,'black'),[])
def precision_report(self, *, rate=False):
before={'period':{'year':2026,'month':8},'black':{'rate':'0.9758','defects':[{'name':'外废','weight':'684.474'}]},
'coverage':{'sourceMonths':[{'year':2026,'month':8,'blackRows':2}]}}
now={'period':{'year':2026,'month':9},'black':{'rate':'0.9727','defects':[{'name':'外废','weight':'615.249'}]},
'coverage':{'sourceMonths':[{'year':2026,'month':9,'blackRows':2}]}}
result={'kind':'quality_comparison','comparison':before,'current':now,'stageRateChangesPP':{'black':'-0.30416066564718'}}
fact={'id':'F1','kind':'stage' if rate else 'category','label':'黑片合格率' if rate else '黑片·外废',
'text':'黑片合格率:97.58%→97.27%,下降0.30个百分点。' if rate else '黑片外废:684.47→615.25 kg,增减-69.23 kg。',
'evidenceRef':'Q00000001','scope':{'period':now['period'],'comparisonPeriod':before['period'],'filters':{}}}
_attach_native_source_numbers(fact,result)
return {'facts':[fact],'requiredFactIds':[],'priorityFactIds':[],'missingEvidence':[],
'rejectedEvidenceRefs':[],'sections':[],'notices':[]}
def test_native_raw_precision_does_not_double_round_or_accept_neighbour(self):
report=self.precision_report()
tokens=_native_atomic_inventory(report['facts'])
token=next(value for value in tokens.values() if value.get('rawValue')=='615.249')
self.assertTrue(_native_number_match('615.2',token))
self.assertFalse(_native_number_match('615.3',token))
self.assertFalse(_native_number_match('615.24',token))
good,calls=self.invoke('黑片外废报废615.2 kg。',report=report)
self.assertTrue(good['evidenceSufficient'],good['compositionDiagnostics']['issues'])
self.assertEqual(len(calls),1);self.assertIn('615.25',good['answer'])
for draft in ('黑片外废报废615.3 kg。','黑片外废报废-615.2 kg。','黑片外废数量615.2件。',
'黑片外废报废615.2吨。','亮片外废报废615.2 kg。',
'2026年8月黑片外废报废615.2 kg。',
'| 类别 | 2026年8月报废(kg) |\n|---|---|\n| 黑片外废 | 615.2 |'):
with self.subTest(draft=draft):
bad,_=self.invoke(draft,report=report);self.assertFalse(bad['evidenceSufficient'])
def test_native_three_decimal_pp_uses_formal_raw_signed_value(self):
report=self.precision_report(rate=True)
token=next(value for value in _native_atomic_inventory(report['facts']).values() if value.get('nativeSignedChange'))
self.assertEqual(token['rawValue'],'-0.30416066564718')
self.assertTrue(_native_number_match('-0.304',token));self.assertFalse(_native_number_match('-0.305',token))
self.assertFalse(_native_number_match('+0.304',token))
good,calls=self.invoke('黑片合格率变化-0.304个百分点。',report=report)
self.assertTrue(good['evidenceSufficient'],good['compositionDiagnostics']['issues']);self.assertEqual(len(calls),1)
for draft in ('黑片合格率变化-0.305个百分点。','黑片合格率变化+0.304个百分点。','黑片合格率变化-0.304%。'):
with self.subTest(draft=draft):
bad,_=self.invoke(draft,report=report);self.assertFalse(bad['evidenceSufficient'])
def test_changes_stage_weight_delta_stays_same_row_and_never_fills_missing(self):
result={'kind':'quality_changes','currentPeriod':{'year':2026,'month':9},'comparisonPeriod':{'year':2026,'month':8},
'rows':[{'label':'产业部甲','present':{'current':True,'comparison':True},
'current':{'black':{'scrapWeight':'26650.2387'},'bright':{'scrapWeight':'15045.38117'}},
'comparison':{'black':{'scrapWeight':'20372.4643'},'bright':{'scrapWeight':'15011.77401'}}}]}
fact={'id':'F1','kind':'supporting','label':'产业部甲规模依据','text':'产业部甲规模依据。','evidenceRef':'Q00000001'}
projected=_native_changes_weight_context(fact,result)
self.assertEqual([x['rawValue'] for x in projected['verifiedWeightChanges']],['6277.7744','33.60716'])
self.assertEqual(fact['text'],'产业部甲规模依据。')
result['rows'][0]['comparison']['bright']['scrapWeight']=None
self.assertEqual(len(_native_changes_weight_context(fact,result)['verifiedWeightChanges']),1)
result['rows'][0]['present']['comparison']=False
self.assertNotIn('verifiedWeightChanges',_native_changes_weight_context(fact,result))
def test_narrow_metric_row_keeps_its_own_raw_value_not_same_rounded_parent(self):
row={'period':{'year':2026,'month':9},'black':{'rate':'0.95129','totalWeight':100},
'coverage':{'sourceMonths':[{'year':2026,'month':9,'blackRows':2}]}}
parent={'kind':'quality_metrics','period':row['period'],'summary':{'black':{'rate':'0.95126'}},'coverage':row['coverage']}
fact=next(f for f in _metric_facts(row,'2026年9月;规格甲','Q00000001') if f['label']=='黑片正式指标')
fact.update(id='F1');_attach_native_source_numbers(fact,parent)
token=next(v for v in _native_atomic_inventory([fact]).values() if v.get('unit')=='%')
self.assertEqual(token['rawValue'],'95.12900')
self.assertTrue(_native_number_match('95.129',token));self.assertFalse(_native_number_match('95.126',token))
def test_nested_scrap_rate_is_measured_formal_value_with_own_period_and_unit(self):
row={'label':'规格甲','present':{'current':True,'comparison':True},'comparison':{},
'current':{'black':{'scrapRate':'0.13192884159534718','totalWeight':'9384.42','scrapWeight':'1238.08'},
'rateSupport':{'blackMeasured':True}}}
result={'kind':'quality_changes','dimension':'product','currentPeriod':{'year':2026,'month':9},
'comparisonPeriod':{'year':2026,'month':8},'rows':[row]}
base={'id':'F1','kind':'supporting','label':'规格甲规模依据','text':'规格甲规模依据。','evidenceRef':'Q00000001',
'scope':{'period':result['currentPeriod'],'dimension':'product','filters':{'product':'规格甲'}}}
fact=_attach_native_source_numbers(_native_changes_weight_context(base,result),result)
token=next(value for value in _native_atomic_inventory([fact]).values() if value.get('metricField')=='scrapRate' and value.get('unit')=='%')
self.assertEqual(token['rawValue'],'13.19288415953471800');self.assertEqual(token['unit'],'%')
report={**self.report(),'facts':[fact]}
good,calls=self.invoke('规格甲9月黑片正式报废率13.19%。',report=report)
self.assertTrue(good['evidenceSufficient'],good['compositionDiagnostics']['issues']);self.assertEqual(len(calls),1)
for text in ('规格甲8月黑片正式报废率13.19%。','规格甲9月亮片正式报废率13.19%。','规格甲9月黑片正式报废率13.19个百分点。'):
bad,_=self.invoke(text,report=report);self.assertFalse(bad['evidenceSufficient'])
row['current']['rateSupport']['blackMeasured']=False
self.assertNotIn('verifiedStageScrapRates',_native_changes_weight_context(base,result))
def test_black_share_requires_formal_denominator_and_never_borrows_bright_basis(self):
side={'period':{'year':2026,'month':9},'black':{'scrapWeight':100,'defects':[
{'name':'责废','weight':60,'rate':'0.6'},{'name':'料废','weight':40,'rate':'0.4'}]},
'coverage':{'sourceMonths':[{'year':2026,'month':9,'blackRows':2}]}}
result={'kind':'quality_comparison','current':side,'comparison':{}}
base={'id':'F1','kind':'category','label':'黑片·责废','text':'黑片责废:60.00 kg。','evidenceRef':'Q00000001',
'scope':{'period':side['period'],'filters':{}}}
fact=_attach_native_source_numbers(_native_black_category_share_context(base,result),result)
report={**self.report(),'facts':[fact]}
good,calls=self.invoke('9月黑片责废占正式报废60.0%。',report=report)
self.assertTrue(good['evidenceSufficient'],good['compositionDiagnostics']['issues']);self.assertEqual(len(calls),1)
bad,_=self.invoke('9月黑片料废占正式报废60.0%。',report=report);self.assertFalse(bad['evidenceSufficient'])
side['black']['scrapWeight']=101
self.assertNotIn('verifiedCategoryShares',_native_black_category_share_context(base,result))
side['black']['scrapWeight']=100
self.assertNotIn('verifiedCategoryShares',_native_black_category_share_context({**base,'label':'亮片·加工废'},result))
def test_changes_decomposition_owns_query_aggregate_scope_not_ranked_row(self):
period={'year':2026,'month':9};previous={'year':2026,'month':8}
def snapshot(selected):return {'period':selected,'black':{'rate':'.97','totalWeight':100},
'bright':{'rate':'.98','totalWeight':100},'coverage':{'sourceMonths':[{'year':2026,'month':selected['month'],'blackRows':2,'brightRows':2}]}}
result={'kind':'quality_changes','current':snapshot(period),'comparison':snapshot(previous),
'stageDecomposition':{'status':'available','blackContributionToCompositePP':'-0.36330259939709003',
'brightContributionToCompositePP':'-0.2318568021276503'}}
scope={'period':period,'comparisonPeriod':previous,'filters':{'department':'新能源汽车磁钢'},'dimension':'group'}
fact=_native_changes_decomposition(result,'Q00000001',scope,'2026年9月;新能源汽车磁钢')
fact['id']='F1';_attach_native_source_numbers(fact,result)
self.assertEqual(fact['scope']['filters'],{'department':'新能源汽车磁钢'});self.assertIsNone(fact['scope']['dimension'])
self.assertEqual([v['rawValue'] for v in fact['verifiedNativeNumericValues']],['-0.36330259939709003','-0.2318568021276503'])
good,calls=self.invoke('新能源汽车磁钢黑片贡献-0.363个百分点,亮片贡献-0.232个百分点。',report={**self.report(),'facts':[fact]})
self.assertTrue(good['evidenceSufficient'],good['compositionDiagnostics']['issues']);self.assertEqual(len(calls),1)
result['current']['coverage']['sourceMonths'][0]['blackRows']=0
self.assertIsNone(_native_changes_decomposition(result,'Q00000001',scope,'范围'))
class NativeAtomicFullPipelineGuards(unittest.TestCase):
"""Build real formal facts; mock only the independent review transport."""
def documents(self, *, rising=False):
docs=evidence();result=docs[0]['result']
for side,weight,rate in [('comparison','684.474','.9727' if rising else '.9758'),
('current','615.249','.9758' if rising else '.9727')]:
result[side]['black']={'rate':rate,'scrapWeight':weight,'totalWeight':'10000','qualifiedWeight':'9000',
'defects':[{'name':'外废','weight':weight}]}
result['stageRateChangesPP']={'black':'.30416066564718' if rising else '-.30416066564718'}
return docs
def invoke(self,draft,*,rising=False,owner='黑片·外废'):
docs=self.documents(rising=rising);context={'_agentTask':{'executionMode':'native_agent'}}
report=build_native_report(docs,context);self.assertEqual(report['missingEvidence'],[])
fact=next(f for f in report['facts'] if f.get('label')==owner);calls=[]
def provider(request,timeout):
payload=json.loads(json.loads(request.data)['messages'][1]['content']);calls.append(payload)
value={'reviews':[{'sourceId':b['sourceId'],'role':'core','factIds':[fact['id']],
'supported':True,'statisticalClaimsSupported':True} for b in payload['draftBlocks']],
'missingGoalDescriptions':[]}
return io.BytesIO(json.dumps({'choices':[{'finish_reason':'stop','message':{'content':json.dumps(value)}}]}).encode())
result=synthesize_native(base_url='https://offline.invalid',api_key='fixture',model='fixture',question='分析质量',
evidence=docs,context=context,native_answer=draft,cancel_event=threading.Event(),opener=provider)
return result,calls
def test_derived_decrease_cannot_escape_wrong_stage_through_later_binder(self):
bad,_=self.invoke('亮片外废减少69.225 kg。')
self.assertFalse(bad['evidenceSufficient']);self.assertNotIn('亮片外废减少69.23',bad['answer'])
good,calls=self.invoke('黑片外废减少69.225 kg。')
self.assertTrue(good['evidenceSufficient'],good['compositionDiagnostics']['issues']);self.assertEqual(len(calls),1)
self.assertIn('黑片外废减少69.23 kg',good['answer'])
def test_table_recovery_cannot_escape_stage_or_period_authorization(self):
for label,header in [('亮片','变化(个百分点)'),('黑片','2026年7月变化(个百分点)'),('黑片','2026年8月变化(个百分点)')]:
draft='黑片质量对比。\n\n| 阶段 | '+header+' |\n|---|---|\n| '+label+' | +0.30 |'
with self.subTest(label=label,header=header):
bad,_=self.invoke(draft,rising=True,owner='黑片合格率');self.assertFalse(bad['evidenceSufficient'])
for header in ('变化(个百分点)','2026年9月变化(个百分点)'):
good,calls=self.invoke('黑片质量对比。\n\n| 阶段 | '+header+' |\n|---|---|\n| 黑片 | +0.30 |',
rising=True,owner='黑片合格率')
self.assertTrue(good['evidenceSufficient'],good['compositionDiagnostics']['issues']);self.assertEqual(len(calls),1)
def test_comparison_dates_keep_current_and_baseline_roles(self):
for owner,positive,negative in [
('黑片·外废','9月黑片外废较8月减少69.225 kg。','8月黑片外废较9月减少69.225 kg。'),
('黑片合格率','9月黑片合格率较8月下降0.304个百分点。','8月黑片合格率较9月下降0.304个百分点。')]:
good,calls=self.invoke(positive,owner=owner)
self.assertTrue(good['evidenceSufficient'],good['compositionDiagnostics']['issues']);self.assertEqual(len(calls),1)
bad,_=self.invoke(negative,owner=owner);self.assertFalse(bad['evidenceSufficient'])
good,calls=self.invoke('黑片外废与8月相比减少69.225 kg。')
self.assertTrue(good['evidenceSufficient'],good['compositionDiagnostics']['issues']);self.assertEqual(len(calls),1)
bad,_=self.invoke('9月黑片外废较7月减少69.225 kg。');self.assertFalse(bad['evidenceSufficient'])
def test_native_repair_never_issues_identifier_numeric_fragments(self):
name='R127.985X57.516X15.786X75.485X32.232°X5.5'
fact={'id':'F1','kind':'product','label':name,'text':name+'报废重量10.25 kg。','evidenceRef':'Q00000001',
'scope':{'period':{'year':2026,'month':9},'filters':{'product':name}}}
unused={**fact,'id':'F2','label':'其他规格甲','text':'其他规格甲报废重量9.99 kg。','scope':{'filters':{'product':'其他规格甲'}}}
report={'facts':[fact,unused],'requiredFactIds':[],'priorityFactIds':[],'missingEvidence':[],
'rejectedEvidenceRefs':[],'sections':[],'notices':[]};requests=[]
def provider(request,timeout):
body=json.loads(request.data);payload=json.loads(body['messages'][1]['content']);requests.append(body)
value={'reviews':[{'sourceId':b['sourceId'],'role':'core','factIds':['F1'],
'supported':True,'statisticalClaimsSupported':True} for b in payload['draftBlocks']],'missingGoalDescriptions':[]}
return io.BytesIO(json.dumps({'choices':[{'finish_reason':'stop','message':{'content':json.dumps(value)}}]}).encode())
with patch('nao_native_answer.build_native_report',return_value=report):
out=synthesize_native(base_url='https://offline.invalid',api_key='fixture',model='fixture',question='分析质量',
evidence=evidence(),context={'_agentTask':{'executionMode':'native_agent'}},native_answer='R127.985规格报废10.25 kg。',
cancel_event=threading.Event(),opener=provider)
self.assertFalse(out['evidenceSufficient']);self.assertEqual(len(requests),2)
catalog=json.loads(requests[0]['messages'][1]['content'])['bindingCatalog']
self.assertTrue(any(key.startswith('I') and value['value']=='其他规格甲' for key,value in catalog.items()))
self.assertFalse(any(key.startswith('N') and value['value']=='127.985' for key,value in catalog.items()))
self.assertNotIn('应替换为{{N',requests[1]['messages'][-1]['content'])
self.assertIn('标识',requests[1]['messages'][-1]['content'])
def test_strict_bound_requires_exact_raw_and_preserves_boundary(self):
for operator in ('不足','不到','低于','高于','少于','多于','小于','大于','超过','至少','至多','上限','下限','<','>','<=','>=','≤','≥'):
with self.subTest(operator=operator):
result,_=self.invoke('黑片外废报废'+operator+'615.2 kg。')
# The formal current value is 615.249. A true threshold need
# not itself be an observed point, and must stay 615.2.
if operator in ('高于','多于','大于','超过','至少','>','>=','≥'):
self.assertTrue(result['evidenceSufficient'],result['compositionDiagnostics']['issues'])
self.assertIn('615.2 kg',result['answer']);self.assertNotIn('615.25 kg',result['answer'])
else:self.assertFalse(result['evidenceSufficient'])
for text,number in [('黑片外废报废低于684.474 kg。','684.474'),('黑片外废报废至多615.249 kg。','615.249')]:
good,calls=self.invoke(text)
self.assertTrue(good['evidenceSufficient'],good['compositionDiagnostics']['issues']);self.assertEqual(len(calls),1)
self.assertIn(number+' kg',good['answer'])
good,calls=self.invoke('黑片外废报废615.2 kg。')
self.assertTrue(good['evidenceSufficient']);self.assertEqual(len(calls),1);self.assertIn('615.25 kg',good['answer'])
if __name__=='__main__':unittest.main()