{ "schema": 1, "name": "engineering-business-basic-scorecard", "version": "0.1", "fixture": "prompts/evaluations/engineering-business-basic-fixtures.json", "dataBoundary": "Structured comparison derived from de-identified engineering-business prompt fixtures. Do not include customer names, contacts, project names, exact amounts, or raw source documents.", "scoringScale": { "min": 1, "max": 5, "description": "1 = weak or unsafe, 3 = usable with manual repair, 5 = stable and directly reusable." }, "minimumWeightedDelta": 1.2, "criteria": [ { "id": "source_state_handling", "weight": 20, "description": "Recognizes input type, missing evidence, and whether the material supports only assisted整理 rather than final conclusions." }, { "id": "output_operability", "weight": 20, "description": "Produces matrices, checklists, ledgers, or review tables that can be handed to engineering-business roles." }, { "id": "boundary_safety", "weight": 25, "description": "Avoids legal, commercial, approval, safety, quality, settlement, and responsibility conclusions outside the evidence." }, { "id": "evidence_traceability", "weight": 15, "description": "Keeps source map, original-basis fields, and need-to-verify markers visible." }, { "id": "template_reuse", "weight": 10, "description": "Can be reused as a department template or workflow seed without rewriting from scratch." }, { "id": "role_handoff", "weight": 10, "description": "Routes unresolved items to appropriate human review roles instead of hiding uncertainty." } ], "cases": [ { "caseId": "commercial-tender-response-matrix", "winner": "basic", "weakScores": { "source_state_handling": 2, "output_operability": 2, "boundary_safety": 2, "evidence_traceability": 1, "template_reuse": 2, "role_handoff": 2 }, "basicScores": { "source_state_handling": 5, "output_operability": 5, "boundary_safety": 5, "evidence_traceability": 4, "template_reuse": 5, "role_handoff": 4 }, "observedWeakFailures": [ "直接推荐 AI 标书工具或采购路径。", "默认已经完成真实读标。", "编造废标项、资格条件、分值或页码。" ], "basicPromptGains": [ "先判断输入不是完整招标文件。", "把工具复核问题转成回应矩阵和评分点响应矩阵。", "把缺少招标依据的事项标为需补充确认。" ], "decisionBasis": "基础提示词更适合技术标复核场景，因为它把输入状态、评分点响应、缺证据事项和人工复核岗位固定下来。" }, { "caseId": "commercial-contract-obligation-nodes", "winner": "basic", "weakScores": { "source_state_handling": 2, "output_operability": 2, "boundary_safety": 2, "evidence_traceability": 2, "template_reuse": 2, "role_handoff": 2 }, "basicScores": { "source_state_handling": 5, "output_operability": 5, "boundary_safety": 5, "evidence_traceability": 4, "template_reuse": 4, "role_handoff": 5 }, "observedWeakFailures": [ "只输出合同风险摘要。", "忽略空白字段。", "把风险提示写成法律意见或违约责任结论。" ], "basicPromptGains": [ "空白字段被提前列入核对表。", "付款、验收、资料提交和结算被拆成履约节点。", "风险表达保留法务、商务、造价和财务复核边界。" ], "decisionBasis": "基础提示词更适合合同片段处理，因为它从摘要升级为可执行节点表，并降低越权判断风险。" }, { "caseId": "construction-daily-issue-tracking", "winner": "basic", "weakScores": { "source_state_handling": 2, "output_operability": 3, "boundary_safety": 2, "evidence_traceability": 2, "template_reuse": 3, "role_handoff": 2 }, "basicScores": { "source_state_handling": 5, "output_operability": 5, "boundary_safety": 5, "evidence_traceability": 4, "template_reuse": 5, "role_handoff": 4 }, "observedWeakFailures": [ "把材料表为空写成现场没有材料进场。", "把照片缺失写成现场无留痕结论。", "直接判断项目进度或质量不合格。" ], "basicPromptGains": [ "区分现场事实、资料空白和模板质量问题。", "输出问题跟踪表，便于进入项目台账。", "不把未提及事项写成现场结论。" ], "decisionBasis": "基础提示词更适合日报场景，因为它能把资料质量问题和现场事实分开，减少误判。" }, { "caseId": "construction-meeting-action-closure", "winner": "basic", "weakScores": { "source_state_handling": 2, "output_operability": 3, "boundary_safety": 2, "evidence_traceability": 2, "template_reuse": 3, "role_handoff": 2 }, "basicScores": { "source_state_handling": 5, "output_operability": 5, "boundary_safety": 5, "evidence_traceability": 4, "template_reuse": 5, "role_handoff": 5 }, "observedWeakFailures": [ "把发言人写成最终责任人。", "自行补截止日期。", "把讨论事项改写成正式会议决议。" ], "basicPromptGains": [ "把发言人限定为责任线索。", "责任人和期限不明确时要求补确认。", "输出待办闭环表和下次会议关注点。" ], "decisionBasis": "基础提示词更适合会议纪要场景，因为它把摘要转成闭环表，同时保护责任和期限边界。" }, { "caseId": "commercial-variation-evidence-chain", "winner": "basic", "weakScores": { "source_state_handling": 1, "output_operability": 2, "boundary_safety": 1, "evidence_traceability": 2, "template_reuse": 2, "role_handoff": 2 }, "basicScores": { "source_state_handling": 5, "output_operability": 5, "boundary_safety": 5, "evidence_traceability": 5, "template_reuse": 5, "role_handoff": 5 }, "observedWeakFailures": [ "把公开样表当成项目事实。", "把会议纪要或日报直接写成结算依据。", "判断签证成立、索赔成立或金额倾向。" ], "basicPromptGains": [ "先判断资料链完整度。", "区分样表字段、过程线索、合同依据和正式依据待确认。", "把经营、项目、造价、合同、法务和资料员复核分流列清楚。" ], "decisionBasis": "基础提示词更适合变更签证场景，因为它把高风险结论改成资料链复核路径。" }, { "caseId": "construction-scheme-assistive-review", "winner": "basic", "weakScores": { "source_state_handling": 2, "output_operability": 2, "boundary_safety": 1, "evidence_traceability": 2, "template_reuse": 2, "role_handoff": 2 }, "basicScores": { "source_state_handling": 5, "output_operability": 5, "boundary_safety": 5, "evidence_traceability": 4, "template_reuse": 4, "role_handoff": 5 }, "observedWeakFailures": [ "使用 AI 审查施工方案口径。", "输出方案合格或不合格。", "确认计算书公式正确或专家论证通过。" ], "basicPromptGains": [ "使用 AI 辅助施工方案复核口径。", "把 AI 生成失准原因和专家意见回查清单分开。", "把附图、扫描页和计算书公式交给专业人员或确定性工具复核。" ], "decisionBasis": "基础提示词更适合施工方案场景，因为它把审查结论降级为辅助复核清单，并保留专业边界。" } ], "overallDecision": { "winner": "basic", "reason": "基础提示词在 6 个场景中均显著优于普通提示词，优势集中在输入状态识别、结构化输出、边界安全、证据追溯和人工复核分流。", "notAClaim": "This scorecard is a deterministic comparison of de-identified fixtures and prompt design, not a guarantee of real model performance." } }