Back to Details

jev-office-gate

Compare original and translation side by side

🇺🇸

Original

English
🇨🇳

Translation

Chinese

Jev Office Gate

Jev Office Gate

Use Jev as a narrow judgment layer after an office assistant has produced a deliverable. Keep the producing assistant and Jev separate unless an actual product integration is verified.
在办公助手生成交付物后，将 Jev 作为一层范围明确的判断层。除非已验证存在实际的产品集成，否则应将生成内容的助手与 Jev 分离。

Workflow

工作流

  1. Identify the deliverable and its release criteria. Preserve the user's requested format and office tool.
  2. Extract bounded review items. For citation review, each item needs a claim and an excerpt copied from the source. A URL alone is not evidence.
  3. Run deterministic checks outside Jev when relevant: file opens, links resolve, quoted text exists, dates and numbers match, formulas calculate, required sections exist, and document layout renders in the target application.
  4. Use Jev only for narrow questions with predefined choices. Prefer claim support, evidence disposition, contradiction/overclaim risk, requirement coverage, or routing priority.
  5. Apply explicit thresholds and route uncertain or consequential items to a human. Never treat model probability as calibrated truth without task-specific evaluation.
  6. Return a release table with the original item, deterministic findings, Jev probabilities, decision, reasons, and required human action. Keep the original generated file unchanged unless the user asked for edits.
  1. 确定交付物及其发布标准。保留用户要求的格式和办公工具。
  2. 提取边界明确的审查项。对于引用审查，每个审查项都需要一项主张和一段从来源复制的摘录。仅有 URL 不构成证据。
  3. 在相关情况下，在 Jev 之外运行确定性检查：文件能否打开、链接能否解析、引用文本是否存在、日期和数字是否匹配、公式能否计算、必需章节是否存在，以及文档版式能否在目标应用中正确呈现。
  4. 仅将 Jev 用于选项预先定义的明确问题。优先处理主张支持度、证据处置、矛盾或过度声称风险、需求覆盖情况及路由优先级。
  5. 应用明确阈值，并将不确定或后果重大的项目路由给人工复核。除非已针对具体任务进行评估，否则绝不能将模型概率视为经过校准的真实性度量。
  6. 返回发布表格，其中包含原始项目、确定性检查结果、Jev 概率、决策、理由和所需人工操作。除非用户要求编辑，否则保持原始生成文件不变。

Citation gate

引用门禁

Use scripts/jev_office_gate.py for repeatable claim-to-evidence review. Read references/input-schema.md when preparing its JSON input.
Default review questions:
  • Does the supplied excerpt support the full claim?
  • Is the supplied source suitable as a formal citation, supporting context, unusable, or in need of manual review?
  • Does the supplied claim/excerpt pair show an obvious semantic conflict or overclaim risk?
The default thresholds are starting points: support below 
0.70
, risk above 
0.60
, disposition confidence below 
0.50
, or any disposition other than 
formal_source
routes the item to review. A selected contradiction or 
cannot_use
result becomes an automatic rejection only when its option probability is at least 
0.80
. Calibrate these values on representative examples before automating release.
使用 scripts/jev_office_gate.py 进行可重复的主张到证据审查。准备其 JSON 输入时，读取 references/input-schema.md。
默认审查问题：
  • 所提供的摘录是否支持完整的主张？
  • 所提供的来源适合作为正式引用或支持性上下文，还是不可用或需要人工审查？
  • 所提供的主张与摘录组合是否显示明显的语义冲突或过度声称风险？
默认阈值仅作为起点：支持度低于 
0.70
、风险高于 
0.60
、处置置信度低于 
0.50
，或处置结果不是 
formal_source
时，该项目会进入审查流程。只有当所选 
contradiction
或 
cannot_use
结果的选项概率至少为 
0.80
时，才会成为自动拒绝项。在自动化发布之前，应使用有代表性的示例对这些值进行校准。

Other office decisions

其他办公决策

For requirement coverage, row classification, or routing, read references/decision-patterns.md. Keep the state small and the answer space fixed. Split compound decisions into separate questions.
对于需求覆盖情况、行分类或路由，请读取 references/decision-patterns.md。保持状态规模较小且答案空间固定。将复合决策拆分为独立问题。

Boundaries

边界

  • Jev does not fetch pages, test URL reachability, authenticate a source, or know whether an excerpt was copied correctly.
  • Do not call it a fact checker. Report that it judged the supplied state and choices.
  • Use code for exact arithmetic, dates, identifiers, string matching, link checks, file integrity, and spreadsheet formulas.
  • Avoid asking one question to combine relevance, freshness, authority, and factual support. Split these dimensions or handle deterministic parts in code.
  • Treat company statements as self-disclosure, not independent verification.
  • For legal, medical, financial, compliance, or externally binding decisions, Jev may prioritize review but must not be the final approver.
  • If Jev is unavailable or credentials are absent, preserve the same review table and mark the judgment columns for human review. Do not pretend the API ran.
  • Jev 不会抓取页面、测试 URL 可达性、验证来源身份，也无法得知摘录是否被正确复制。
  • 不要将其称为事实核查器。应说明它仅判断所提供的状态和选项。
  • 对于精确算术、日期、标识符、字符串匹配、链接检查、文件完整性及电子表格公式，请使用代码。
  • 不要用一个问题同时涵盖相关性、时效性、权威性和事实支持度。应拆分这些维度，或在代码中处理确定性部分。
  • 将公司声明视为自我披露，而不是独立验证。
  • 对于法律、医疗、财务、合规或对外具有约束力的决策，Jev 可以确定审查优先级，但不得成为最终批准者。
  • 如果 Jev 不可用或缺少凭据，请保留相同的审查表格，并将判断列标记为需要人工审查。不要假装 API 已运行。

Credentials

凭据

The helper reads 
TYPESAFE_API_KEY
by default and also accepts 
JEV_API_KEY
. Never place an API key in a document, prompt, source file, or saved review artifact.
辅助脚本默认读取 
TYPESAFE_API_KEY
，同时也接受 
JEV_API_KEY
。绝不要将 API key 放入文档、提示词、源文件或已保存的审查产物中。