{ "datasetId": "ai-product-hallucination-governance-card-v1", "schemaVersion": "1.0.0", "createdAt": "2026-07-21", "language": "zh-CN", "license": "https://creativecommons.org/licenses/by/4.0/", "licenseScope": "仅覆盖空白字段结构与 DragonAI 第一方说明，不覆盖真实用户输入、模型输出、知识库、事故、评测结果或第三方材料。", "purposeZh": "记录 AI 产品幻觉风险边界、逐声明证据、评测、控制、运行事件、复盘与发布决定。", "disclosureZh": "DragonAI 运营 AI 产品经理课程。本文件是第一方空白治理框架，不是模型安全认证、法律意见、医疗或金融判断，也不构成无幻觉保证；不含真实客户、事故或跑分，采用前须由适用责任人复核。", "claimVerdictValues": [ "supported", "contradicted", "unsupported", "stale", "not_verifiable", "not_applicable" ], "riskTypes": [ "source_conflict", "unsupported_assertion", "citation_mismatch", "fabricated_reference", "stale_fact", "scope_overreach", "uncalibrated_confidence", "deterministic_error", "other" ], "releaseDecisionValues": [ "continue_testing", "limited_release", "hold", "rollback", "retire" ], "evaluationStatusValues": [ "not_evaluated", "partial", "complete", "invalid" ], "dataCompletenessValues": [ "unknown", "partial", "complete", "invalid" ], "eventStatusValues": [ "open", "triaged", "contained", "corrected", "closed" ], "incidentSeverityValues": [ "low", "medium", "high", "critical" ], "rootCauseCategoryValues": [ "model", "retrieval", "source", "tool", "prompt", "policy", "ui", "data_pipeline", "unknown" ], "controlImplementationStatusValues": [ "planned", "implemented", "tested", "ineffective", "retired" ], "rollbackReadinessStatusValues": [ "unknown", "not_ready", "drill_passed", "production_verified" ], "sections": [ { "sectionId": "risk-boundary", "nameZh": "用途、幻觉边界与后果", "requiredFields": [ "riskProfileId", "productVersion", "releaseUnitHash", "targetUsers", "taskDefinition", "allowedOutputs", "verifiableClaimTypes", "authoritativeSourcePolicy", "freshnessPolicy", "prohibitedDomains", "riskTypes", "affectedParties", "severityDefinitions", "riskTolerance", "owners", "frozenAt" ], "verificationQuestionsZh": [ "错误是否按任务后果而非单一平均率分级？", "允许来源、时效和禁止领域是否明确？" ] }, { "sectionId": "claim-evidence", "nameZh": "逐声明证据链", "requiredFields": [ "evidenceRecordId", "riskProfileId", "runId", "claimId", "outputSpan", "claimType", "requiredEvidenceLevel", "sourceEvidence", "verdict", "adjudicator", "adjudicationRuleVersion", "rationale", "disagreement", "adjudicatedAt" ], "verificationQuestionsZh": [ "每个来源的版本、时间、证据片段和支持关系是否绑定在同一对象？", "冲突、过期和无法判定项是否保留？" ] }, { "sectionId": "evaluation-plan", "nameZh": "评测集、切片与判定", "requiredFields": [ "evaluationPlanId", "riskProfileId", "modelVersion", "promptVersion", "retrievalIndexVersion", "toolVersions", "samplingConfiguration", "datasetVersion", "normalSlices", "missingEvidenceSlices", "conflictingEvidenceSlices", "staleEvidenceSlices", "adversarialSlices", "highImpactSlices", "expectedBehaviors", "prohibitedBehaviors", "severeFailureRules", "repeatCount", "judgeModelId", "judgeModelVersion", "judgePromptVersion", "judgeRubricVersion", "judgeCalibrationDatasetVersion", "judgeCalibrationThresholds", "requiredHumanReviewSlices", "humanReviewPlan", "agreementMethod", "denominatorPolicyVersion", "metricDefinitions", "zeroDenominatorTreatment", "frozenAt" ], "verificationQuestionsZh": [ "正常完成、合理克制和过度拒答是否同时测量？", "自动裁判是否校准，高风险与分歧项的人工复核是否预先定义？" ] }, { "sectionId": "control-matrix", "nameZh": "失败原因与控制矩阵", "requiredFields": [ "controlPlanId", "controlId", "controlVersion", "riskProfileId", "riskType", "failureCause", "preventiveControls", "detectiveControls", "groundingPolicy", "verificationRules", "clarificationRules", "abstentionRules", "humanReviewRules", "executionRestrictions", "fallbackMode", "dependencyFailureHandling", "stopConditions", "implementationStatus", "controlTestReferences", "effectivenessResult", "residualRisk", "controlOwner", "evidenceReferences", "effectiveAt" ], "verificationQuestionsZh": [ "每项控制是否绑定具体失败原因、版本和测试证据？", "检索失败、来源冲突和依赖故障是否有安全降级？" ] }, { "sectionId": "evaluation-result", "nameZh": "质量、克制与可用性结果", "requiredFields": [ "evaluationResultId", "evaluationPlanId", "runReferences", "evaluationStatus", "claimCount", "applicableClaimCount", "supportedClaimCount", "contradictedClaimCount", "unsupportedClaimCount", "staleClaimCount", "notVerifiableClaimCount", "notApplicableClaimCount", "claimCountReconciliationVariance", "citationEligibleClaimCount", "citationSupportedClaimCount", "citationSupportRate", "taskEligibleCount", "completedTaskCount", "taskCompletionRate", "abstentionEligibleCount", "appropriateAbstentions", "incorrectAbstentions", "clarificationEligibleCount", "effectiveClarifications", "severeFailures", "humanHandoffs", "sliceResults", "judgeEligibleCount", "judgeEvaluatedCount", "judgeCoverage", "agreementComparedCount", "agreementCount", "judgeAgreement", "humanReviewEligibleCount", "humanReviewedCount", "humanReviewCoverage", "adjudicatedDisagreementCount", "judgeCalibrationResultReference", "judgeErrorSummary", "adjudicationReferences", "metricFormulaVersion", "dataCompleteness", "limitations", "reviewedAt" ], "verificationQuestionsZh": [ "每个派生率的分子、分母、公式版本、零分母和六种声明判定是否可重算对账？", "自动裁判覆盖、人工复核、分歧裁决和高风险切片是否有执行证据？" ] }, { "sectionId": "runtime-event", "nameZh": "生产事件与用户纠错", "requiredFields": [ "eventId", "releaseUnitHash", "occurredAt", "taskReference", "inputSummary", "inputDigest", "retrievalTraceReference", "toolTraceReference", "outputReference", "outputDigest", "claimEvidenceReferences", "detectedBy", "userFeedback", "correction", "humanHandoff", "finalOutcome", "potentialImpact", "contentAccessPolicy", "privacyTreatment", "retentionUntil", "owner", "status" ], "verificationQuestionsZh": [ "事件能否通过可读摘要、受控输出引用和摘要值追溯到发布单元与证据链？", "敏感数据是否按访问政策、最小化和期限处理？" ] }, { "sectionId": "incident-review", "nameZh": "事件复盘与纠正验证", "requiredFields": [ "incidentReviewId", "eventIds", "releaseUnitHashes", "riskProfileId", "controlPlanIds", "rootCauseCategory", "affectedScope", "severity", "containment", "correction", "controlChanges", "regressionSetAdditions", "retestReferences", "correctionVerificationStatus", "residualRisk", "owner", "approvers", "reviewedAt" ], "verificationQuestionsZh": [ "修复是否转入回归集并以明确状态完成重测？", "根因是否通过受控枚举区分模型、检索、来源、工具、提示、策略、界面和数据管线？" ] }, { "sectionId": "release-decision", "nameZh": "发布、回滚与重测决定", "requiredFields": [ "decisionId", "releaseUnitHash", "riskProfileId", "decisionRuleVersion", "precommittedAt", "evaluationPlanIds", "evaluationResultIds", "incidentReviewIds", "controlPlanIds", "controlTestResultIds", "applicableSlices", "thresholds", "metricThresholdResults", "severeFailureLimit", "judgeCoverageRequirement", "unknownTreatment", "monitoringPlanId", "monitoringEvidenceReferences", "monitoringOwner", "responseSla", "humanReviewCapacity", "rollbackUnitHash", "rollbackPlan", "rollbackDrillEvidence", "rollbackReadinessStatus", "decision", "decisionReason", "contradictingEvidence", "owners", "approvers", "decidedAt", "retestTriggers" ], "verificationQuestionsZh": [ "门槛、未知值处理和决定规则是否在观察结果前冻结并绑定评测与控制证据？", "监测、人工容量、回滚单元和演练证据是否证明受限发布可执行？" ] } ], "blankRiskBoundary": { "riskProfileId": "", "productVersion": "", "releaseUnitHash": "", "targetUsers": [], "taskDefinition": "", "allowedOutputs": [], "verifiableClaimTypes": [], "authoritativeSourcePolicy": [], "freshnessPolicy": {}, "prohibitedDomains": [], "riskTypes": [], "affectedParties": [], "severityDefinitions": [], "riskTolerance": {}, "owners": [], "frozenAt": "" }, "blankClaimEvidence": { "evidenceRecordId": "", "riskProfileId": "", "runId": "", "claimId": "", "outputSpan": "", "claimType": "", "requiredEvidenceLevel": "", "sourceEvidence": [], "verdict": "not_verifiable", "adjudicator": "", "adjudicationRuleVersion": "", "rationale": "", "disagreement": null, "adjudicatedAt": "" }, "blankEvaluationPlan": { "evaluationPlanId": "", "riskProfileId": "", "modelVersion": "", "promptVersion": "", "retrievalIndexVersion": "", "toolVersions": [], "samplingConfiguration": {}, "datasetVersion": "", "normalSlices": [], "missingEvidenceSlices": [], "conflictingEvidenceSlices": [], "staleEvidenceSlices": [], "adversarialSlices": [], "highImpactSlices": [], "expectedBehaviors": [], "prohibitedBehaviors": [], "severeFailureRules": [], "repeatCount": null, "judgeModelId": "", "judgeModelVersion": "", "judgePromptVersion": "", "judgeRubricVersion": "", "judgeCalibrationDatasetVersion": "", "judgeCalibrationThresholds": {}, "requiredHumanReviewSlices": [], "humanReviewPlan": {}, "agreementMethod": "", "denominatorPolicyVersion": "", "metricDefinitions": [], "zeroDenominatorTreatment": "return_null_and_flag_insufficient_evidence", "frozenAt": "" }, "blankControlMatrix": { "controlPlanId": "", "controlId": "", "controlVersion": "", "riskProfileId": "", "riskType": "", "failureCause": "", "preventiveControls": [], "detectiveControls": [], "groundingPolicy": {}, "verificationRules": [], "clarificationRules": [], "abstentionRules": [], "humanReviewRules": [], "executionRestrictions": [], "fallbackMode": "", "dependencyFailureHandling": [], "stopConditions": [], "implementationStatus": "planned", "controlTestReferences": [], "effectivenessResult": "", "residualRisk": "", "controlOwner": "", "evidenceReferences": [], "effectiveAt": "" }, "blankEvaluationResult": { "evaluationResultId": "", "evaluationPlanId": "", "runReferences": [], "evaluationStatus": "not_evaluated", "claimCount": null, "applicableClaimCount": null, "supportedClaimCount": null, "contradictedClaimCount": null, "unsupportedClaimCount": null, "staleClaimCount": null, "notVerifiableClaimCount": null, "notApplicableClaimCount": null, "claimCountReconciliationVariance": null, "citationEligibleClaimCount": null, "citationSupportedClaimCount": null, "citationSupportRate": null, "taskEligibleCount": null, "completedTaskCount": null, "taskCompletionRate": null, "abstentionEligibleCount": null, "appropriateAbstentions": null, "incorrectAbstentions": null, "clarificationEligibleCount": null, "effectiveClarifications": null, "severeFailures": [], "humanHandoffs": null, "sliceResults": [], "judgeEligibleCount": null, "judgeEvaluatedCount": null, "judgeCoverage": null, "agreementComparedCount": null, "agreementCount": null, "judgeAgreement": null, "humanReviewEligibleCount": null, "humanReviewedCount": null, "humanReviewCoverage": null, "adjudicatedDisagreementCount": null, "judgeCalibrationResultReference": "", "judgeErrorSummary": {}, "adjudicationReferences": [], "metricFormulaVersion": "", "dataCompleteness": "unknown", "limitations": [], "reviewedAt": "" }, "blankRuntimeEvent": { "eventId": "", "releaseUnitHash": "", "occurredAt": "", "taskReference": "", "inputSummary": "", "inputDigest": "", "retrievalTraceReference": "", "toolTraceReference": "", "outputReference": "", "outputDigest": "", "claimEvidenceReferences": [], "detectedBy": "", "userFeedback": "", "correction": "", "humanHandoff": {}, "finalOutcome": "", "potentialImpact": "", "contentAccessPolicy": {}, "privacyTreatment": {}, "retentionUntil": "", "owner": "", "status": "open" }, "blankIncidentReview": { "incidentReviewId": "", "eventIds": [], "releaseUnitHashes": [], "riskProfileId": "", "controlPlanIds": [], "rootCauseCategory": "unknown", "affectedScope": [], "severity": "low", "containment": [], "correction": [], "controlChanges": [], "regressionSetAdditions": [], "retestReferences": [], "correctionVerificationStatus": "not_evaluated", "residualRisk": "", "owner": "", "approvers": [], "reviewedAt": "" }, "blankReleaseDecision": { "decisionId": "", "releaseUnitHash": "", "riskProfileId": "", "decisionRuleVersion": "", "precommittedAt": "", "evaluationPlanIds": [], "evaluationResultIds": [], "incidentReviewIds": [], "controlPlanIds": [], "controlTestResultIds": [], "applicableSlices": [], "thresholds": [], "metricThresholdResults": [], "severeFailureLimit": null, "judgeCoverageRequirement": null, "unknownTreatment": "hold", "monitoringPlanId": "", "monitoringEvidenceReferences": [], "monitoringOwner": "", "responseSla": "", "humanReviewCapacity": {}, "rollbackUnitHash": "", "rollbackPlan": {}, "rollbackDrillEvidence": [], "rollbackReadinessStatus": "unknown", "decision": "hold", "decisionReason": "", "contradictingEvidence": [], "owners": [], "approvers": [], "decidedAt": "", "retestTriggers": [] }, "sources": [ { "evidenceId": "nist-ai-rmf-genai-profile-20260721", "url": "https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-generative-artificial-intelligence" }, { "evidenceId": "openai-why-language-models-hallucinate-20260721", "url": "https://openai.com/zh-Hans-CN/index/why-language-models-hallucinate/" }, { "evidenceId": "google-cloud-grounding-overview-20260721", "url": "https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/grounding/overview" }, { "evidenceId": "google-cloud-genai-evaluation-metrics-20260721", "url": "https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/determine-eval" }, { "evidenceId": "microsoft-hax-human-ai-guidelines-20260721", "url": "https://www.microsoft.com/en-us/haxtoolkit/ai-guidelines/" } ], "title": "AI 产品幻觉治理证据卡 v1", "description": "把风险边界、逐声明证据、评测切片、控制矩阵、运行事件、事故复盘和发布门禁连接起来。", "url": "https://course.dragonai.tech/datasets/ai-product-hallucination-governance-card-v1.json", "creator": "烛龙智元内容研究组" }