360 Secure SkillHub
360 安全技能中心

GroundLens

将草稿答案、问题与多份源文档写成文件后，用本地命令行工具比对，检查数字、日期、金额、算术结果及实体指向是否与原文一致，返回 PASS、FAIL 或 REVIEW 判定并可输出逐条断言与 JSON 记录；全程不联网、不需 API key、不调用第二个模型。

javihaus
javihaus
热度 24.44
最近更新 2026-10-17 19:03:26
分类：数据依赖：代码执行 / 读写文件应用场景：数据处理 / 校对润色 / 研究

安装技能

复制下面的话术发给你的 Agent，它会通过 secure-skills 检索并安装该技能。

执行命令安装此 skill： npx secure-skills-cli add javihaus/groundlens
Metadata
name
groundlens
description
Check an answer against the user's own documents before sending it. Catches wrong figures, recomputes the arithmetic, and flags claims about the wrong company or person. Runs locally in milliseconds, no API key, no second model.
version
1.0.0
metadata
{"openclaw":{"emoji":"🔎","homepage":"https://groundlens.dev","requires":{"bins":["groundlens"]},"install":[{"kind":"uv","package":"groundlens","bins":["groundlens"]}]}}

GroundLens: check the answer before you send it

Use this skill when you answer from documents the user gave you or that you retrieved (reports, contracts, invoices, statements, regulations, notes) and your answer contains figures, dates, percentages, amounts or names.

GroundLens compares your draft with those documents and returns one of three decisions:

  • PASS: every figure it could check is in the sources or follows from them.
  • FAIL: a figure is contradicted, a calculation does not add up, or the answer talks about a different entity than the one asked about.
  • REVIEW: it could not confirm something. Not proven wrong, not proven right.

It runs on the user's machine. No account, no API key, no network, no model.

When to use it

  • Before sending any answer that quotes numbers taken from documents.
  • After you compute something from document figures (a total, a change, a ratio), to check the arithmetic.
  • When the user asks about one company, person or product and the documents mention several.

Skip it for small talk, opinions, or answers with no figures and no documents.

How to run it

Write the draft answer, the question and each source to files, then:

groundlens verify \
  --question question.txt \
  --answer answer.txt \
  --source report=annual_report.txt \
  --source notes=meeting_notes.txt
  • Give every source an id (report=, notes=); the result cites it.
  • Add --scope entity="Acme Ltd" when the question is about one entity.
  • Add --locale es (or de, fr, ...) when the documents write numbers in that language's format, e.g. 41,2 millones.
  • Add --json to get the full record with every claim and its source.

Exit code: 0 PASS, 1 FAIL, 3 REVIEW, 2 error.

Example output:

FAIL  policy=groundlens_default_v1  record=rec_ab201f0cf308_b0ced69fd71a
  c2   groundlens.numeric       contradicted  0.00  nearest in report: '€41.2 million'
  c5   groundlens.numeric       unsupported   0.00  nearest in report: '2025'

What to do with the result

  • PASS: send the answer.
  • FAIL: do not send it as is. Read each reason, fix the figure using the value GroundLens found in the source, and verify again. If you cannot fix it, tell the user which figure you could not confirm.
  • REVIEW: send the answer, but say plainly which figure or claim you could not confirm in their documents, so they can check it themselves.

Never present a FAIL or REVIEW answer as verified.

Limits

GroundLens checks figures, arithmetic, who the answer is about, and whether it answers the question. It does not judge the meaning of free text: a wrong statement with no number in it will usually pass. Treat PASS as "no figure problem found", not as "the answer is true".

More