クイック回答
AI performance review summaryのrelease前に7 check：pre-existing role standard/period、material sentence-source link、observation/interpretation/decision分離、missing/contrary/changing context保持、consistent application/proxy test、accommodation/leave/complaint/retaliation保護、employee response/correction/authorized release。reviewerがwhat/whyをreconstructできる場合のみshort summaryが有用。
compression前にsummary contractを定義
permitted evidence
approved system、artifact、metric、observation、feedback、self-report、checkpoint、date range、access、reliability、prohibited sourceを列挙しrestricted dataをneed-to-know。
evidence statement
criterion、exact fact、source、period、context、employee role、contrary evidence、uncertainty、reviewer interpretation、evidence stateを別fieldで保存。
decision boundary
summarize、verify、correct、adjudicate、rate、compensation/employment action、restricted access、employee notice、unsafe stopのauthorityを明示。
release record
input version、generated draft、citation、validation、human edit/reason、employee response、correction、approval、recipient、notice、retention、appeal、downstream actionを保持。
evidence-preserving review summaryの7 checks
各checkをseparate release gateとして使用。一項passでmissing source、inaccessible process、unresolved dispute、unauthorized decisionを相殺しない。
review contract/evaluation periodをlock
summary前にrole/level、essential function、approved standard/weight、period、reviewer authority、policy version、prior goal、permitted evidence、accommodation status、summary-supported decisionを特定。
- release evidence
- 全criterionをpre-existing job-related standardへtraceしperiod中のchangeを確認可能。
- failure to prevent
- result後のcriterion reverse-engineer、other-role expectation、AI-generated competencyのapproved扱いを禁止。
全material statementのsource provenanceを保持
dated artifact、system record、approved metric、stakeholder feedback、manager observation、employee self-report、prior checkpointへclaimをlinkしauthor、context、access、direct/hearsayを記録。
- release evidence
- criterion stateを変えるsentenceにopenable referenceとfaithful excerpt/structured fact。
- failure to prevent
- unverified recollection、anonymous sentiment、surveillance、message volume、presence、model inferenceをperformance factへ変換しない。
observation・interpretation・decisionを分離
what happened/sourceを先に、criterion interpretationを次に、rating、compensation、promotion、discipline、development、employment decisionはauthorized fieldへ。
- release evidence
- direct fact、employee claim、reviewer inference、disputed interpretation、final determinationを区別。
- failure to prevent
- defined observable behavior/sourceなしにleadership、attitude、commitment、ownership、executive presence、fitをevidence化しない。
missing・contrary・changing contextを可視化
criterionごとにsupport/contrary evidence、incomplete period、scope change、leave、access delay、resource unavailable、ownership transfer、metric change、unresolved dependencyを確認しexplicit stateを使う。
- release evidence
- coverage date、gap、contrary example、context change、organization blocker、uncertainty resolutionを記載。
- failure to prevent
- missing monthをaverageで埋めずsilence=success、learning-producing failure削除、organization input不足をemployee責任にしない。
consistency・comparability・prohibited proxyをcheck
similarly situated employeeにsame standard、evidence window、data source、rounding、opportunity、support、reviewer discretionが適用されたか比較し他人dataを露出せず差異調査。
- release evidence
- comparison group definition、tested dimension、material difference、explanation、reviewer、correction、monitoring decisionを記録。
- failure to prevent
- age、disability、leave、accent、caregiving、location、schedule、camera、network centrality、complaint activity、social similarityでrankしない。
accommodation・leave・complaint・retaliation contextを保護
known/possible accommodation requestをapproved confidential processへroute。protected leave、complaint、investigation、safety report、protected activityをapplicable rule/need-to-knowで扱う。
- release evidence
- performance recordはnecessary job evidenceのみ、restricted context、hold、authorized adjustment、non-retaliation reviewを別管理。
- failure to prevent
- diagnose、summaryでmedical detail要求、approved absenceをmissed output、accommodation/report/discussion/data challengeでreview低下を禁止。
employee response・correction・authorized releaseを要求
accessible formatでsummary、citation、standard、processをemployeeへ提示しresponse、disputed fact、additional evidence、correction outcome、reviewer、appeal/complaint routeをauthorized use前に記録。
- release evidence
- release recordにemployee access、response window、correction、unresolved dispute、approval、rule version、notice、recipient、retention、downstream action。
- failure to prevent
- seen、silence、signature、manager clickをagreement/rights waiverとせずemployee response上書きやAI final rating/actionを禁止。
criterion evidence stateを分離
demonstrated/partial
cited evidenceがperiod内にpre-existing anchorのall/defined partを満たす。exact dimensionとcontrary evidenceを保持。
insufficient/not assessed
coverageがconclusionをsupportしない/valid assessmentなし。absenceをpoor performanceへ変換しない。
contradicted/disputed
material source disagreementまたはemployee accuracy/interpretation challenge。両position保持しadjudicationへ。
context/accommodation hold
scope change、missing organizational input、protected leave、unresolved accommodationがvalid interpretationへ影響。context制限しaffected useをpause。
worked example：activity volumeはdelivery qualityを証明しない
hypotheticalでreal employee/validated ratingではない。AI draftは「final six weeksのmessage activityが32%低下しlow engagement」とした。しかしapproved standardはaccurate monthly forecast/documented escalationでmessage volumeではない。approved leave、data access delayがありcompleted forecast artifactはapproved error/review criteriaを満たす。
unsafe compression
prohibited proxyをperformance evidence化しapproved standard/leave contextを無視、actual artifactを削除しpercentageからcausationを示唆。
evidence-preserving correction
「forecast accuracy—4 completed artifactでdemonstrated [links]、approved error/reviewer criteria pass。escalation—partial [2 records]。required access unavailableのperiodはnot assessed。approved leave/restricted accommodationはoutput scoring除外。message activityはapproved criterionでなくremove。employee response/final ratingはauthorized review pending。」
corrected summaryはevidence fileよりshortだがreviewに必要なlink、approved standard、coverage、missing period、proxy decision、employee response status、human authorityを保持。
OpenMaxによるevidence-preserving summary coordination
OpenMaxはapproved review contract、permissioned evidence collection、source-linked draft、contrary search、coverage/context flag、independent review、consistency check、accommodation hold、employee review、correction、adjudication、authorized release、audit、monitoringをcoordinate。job/performance standard、legal scope、accommodation、sensitive context、compensation、discipline、promotion、appeal、employment decisionはhuman。
employment・fairness・privacy・accessibility boundary
- qualified local legal、HR、labor、compensation、benefit、accessibility、privacy、security、ER、record、localization、AI-governance ownerがrole、standard、data source、comparison、workflow、notice、retention、decision useをapprove。
- job-related、pre-existing、consistently applied standardとapplicable effective accommodation/accessible participation。medical/restricted factはauthorized need-to-know以外でsummaryに含めない。
- protected leave、accommodation request、discrimination/safety complaint、investigation participation、protected working-condition discussion、correction、automated evidence challengeをpenalizeせずretaliation/disparate treatment riskをqualified channelでreview。
- detailed summary、source link、consistency check、bias audit、manager approval、employee signature、metric、dashboardをlawful、accurate、fair、predictive、non-retaliatory、valid performance managementの証明にしない。actual outcome、correction、appeal、impactをevaluateしharmful useをstop。
情報源、編集方法、制限
OpenMax編集部はcurrent U.S. EEOC performance evaluation、consistent standard、disability/accommodation、internal discrimination complaint、correction、retaliationおよびNIST AI RMF human role、documentation、uncertainty、testing、independent review、impactを確認し7 checks、evidence state、activity-volume correctionを独自構成。2026年9月3日再確認。
- U.S. EEOC — Conducting performance evaluations
- U.S. EEOC — Performance and conduct standards for employees with disabilities
- U.S. EEOC — Handling discrimination complaints about evaluations
- U.S. EEOC — Treating employees consistently
- NIST — AI RMF Core
よくある質問
AIはfinal performance ratingをassign？
AIはratingをownしない。cited evidence整理/gap flagは可能だがauthorized peopleがstandard validation、context/accommodation、employee response、dispute adjudication、accountabilityを担う。
summaryは全source documentを含める？
full evidence fileはapproved accessで保持。summaryはmaterial statementをsmallest sufficient sourceへlinkしinterpretationを変えるcontrary/missing evidenceを保持。
message volumeはuseful performance metric？
defaultではない。pre-approved validated job-related measureのみをscope、opportunity、support、context、uncertainty、proxy review付きで解釈。
employeeがfactをdisputeしたら？
original statement、employee response、additional evidence、reviewer、correction decision/reason、unresolved issue、appeal/complaint routeを保持しsilent overwrite禁止。
OpenMaxで何をautomate？
evidence collection、source-linked draft、gap/consistency check、review、employee response、correction、adjudication、release、monitoringをcoordinateしstandard/employment authorityはhuman。