Zetesis Scientific Due Diligence

io.github.reutavidan由 GitHub 账号 reutavidan 发布v0.2.2更新于 Oct 25, 2026

Due diligence on life-science claims: omics, trials, gene therapy. Dimension by dimension.

已验证Streamable HTTP可网页运行知识与记忆数据与分析AI 与机器学习

概览

AI 生成的概览

Zetesis 依据公开证据评估科学、生物医学、临床及 AI/ML 主张,返回带引用的来源、审阅问题与失败模式。

可以做什么
将主张归入其科学类别,返回来源目录、领域审阅者会问的问题以及失败模式分类(含反例),可限定到较早年份,只用当时已存在的证据。对 Europe PMC、ClinicalTrials.gov、openFDA、NIH RePORTER 和 SEC EDGAR 执行检索,返回去重后的证据包,每条来源都带硬性公开标识符,并附评分标准。由 Zetesis 服务端逐维度给出对主张的评级解读和总体结论。校验已签名的 Zetesis 证明,确认其主张、证据与结论自签署后未被改动。
适用场景
适合核查已发表证据是否支持某项生命科学、临床或 AI/ML 主张,例如摘要、论文、资助申请或投资材料;也适合按主张提出当年可知的信息而非事后视角来评判。还可在他人分享 Zetesis 档案时独立复核。
运行要求
托管端点无需账户、密钥或令牌。较早的 stdio Python 客户端(通过 uvx 安装 zetesis 包)需要 ZETESIS_TOKEN 访问令牌,需向 Zetesis 申请;校验功能无需令牌。需要能访问 Zetesis API 及上述公开证据来源的网络。
安装前请注意
evaluate_claim 会把主张文本发送给 Zetesis 隐私政策中列明的模型分包处理方;证据与范围工具不向模型提供商发送任何内容。主张文本不被记录,仅保留路由类别、次数等元数据。stdio 客户端需要 ZETESIS_TOKEN 密钥,并将每次调用代理到托管引擎。托管引擎是与 MIT 许可代码分开的独立服务。

安装

在 SourceWeft 中

  1. 打开 控制台中的 Zetesis Scientific Due Diligence,将其添加到工作区。
  2. 为需要使用其工具的对话启用该服务。

Web executable,通过 Streamable HTTP。 远程服务在工作区中配置后即可从网页运行时运行。

交给智能体安装

把这段提示词粘贴到 Claude Code、Codex、Cursor 等智能体,或 SourceWeft 对话中。智能体会读取该服务器的安装指南,先向你展示发布者、端点和登录方式,经你同意后再把它添加到你的客户端。密钥不会写进任何命令。

阅读 https://sourceweft.com/mcp/io.github.reutavidan%2Fzetesis/install.md,并按其中说明把这个 MCP 服务器接入我的客户端。修改任何配置之前,先告诉我它的发布者、端点和登录方式,等我确认。不要把 API 密钥写进命令或我会提交的文件;问我密钥应从哪里获取。

查看智能体所依据的安装指南

其他 MCP 客户端

把它添加到你客户端的 mcpServers 配置中。

{
  "mcpServers": {
    "zetesis-scientific-due-diligence": {
      "type": "http",
      "url": "https://api.zetesis.science/mcp"
    }
  }
}

README

Zetesis

Available on Smithery

Scientific due diligence on a claim, from inside Claude, Copilot, or any MCP host.

Give Zetesis a claim, an abstract, a paper, a grant or a deck. It routes the claim to its scientific class, then returns the questions a domain reviewer would ask, the failure patterns that caught comparable claims before, and the public evidence bearing on it, with a PMID, DOI, NCT number, NIH grant number or SEC filing reference on every source. Every identifier it hands back was retrieved. None are generated.

It can also evaluate a claim as it stood in an earlier year, restricting evidence to what existed by then, so a claim is judged on what was knowable at the time rather than on how it turned out.

Install

No account, key or token is required for any of these.

WhereCommand
Terminal chatuv tool install zetesis, then run zetesis (or pipx install zetesis)
Claude Codeclaude mcp add --transport http zetesis https://api.zetesis.science/mcp
Claude Code plugin (tools plus skills)/plugin marketplace add reutavidan/zetesis-plugin, then /plugin install zetesis@zetesis
Codexcodex mcp add zetesis --url https://api.zetesis.science/mcp
Gemini CLIgemini mcp add --transport http zetesis https://api.zetesis.science/mcp
Browserhttps://api.zetesis.science/app

Chat in the terminal

bash
uv tool install zetesis     # or: pipx install zetesiszetesis

zetesis opens a chat in the terminal with the hosted Zetesis agent. Ask a question, or drag a deck or paper (PDF, DOCX or PPTX) into the window, and watch it scope the claim, search the dated record and grade each dimension. Follow-up questions use the same evidence. /help lists the commands (/attach, /sources, /new, /web). No account or key is needed.

Connect it

The hosted server is at https://api.zetesis.science/mcp, over Streamable HTTP. No account, key or token is required.

Claude Code:

bash
claude mcp add --transport http zetesis https://api.zetesis.science/mcp

Claude Desktop (claude_desktop_config.json):

json
{  "mcpServers": {    "zetesis": {      "type": "http",      "url": "https://api.zetesis.science/mcp"    }  }}

Any other MCP client:

ClientHow
Microsoft Copilot StudioTools, then Add a tool, then Model Context Protocol. Server URL, auth None.
ChatGPTSettings, then Connectors, then Developer mode. Add the URL.
Gemini CLIgemini mcp add --transport http zetesis https://api.zetesis.science/mcp
Codexcodex mcp add zetesis --url https://api.zetesis.science/mcp

For Gemini's settings.json, use httpUrl rather than url; the latter is SSE and will not connect. Full setup notes: https://api.zetesis.science/docs

Tools

zetesis_scope routes the claim and returns the diligence apparatus for its class: the questions a reviewer would ask, structured by substrate, methods, cohort and risk of bias, a failure-pattern taxonomy carrying the companies each pattern was derived from, and the edge cases where those patterns were wrong. A checklist that only ever fires positive teaches over-rejection, so the counterexamples ship alongside it.

zetesis_evidence runs the searches and returns a deduplicated bundle from Europe PMC, ClinicalTrials.gov, openFDA, NIH RePORTER and SEC EDGAR, every source carrying a hard public identifier, followed by the grading rubric so you grade the evidence yourself in context.

Neither of those calls a language model. They return in under a second, cost nothing to run, and send nothing to a model provider. That is usually the answer a security reviewer is looking for.

evaluate_claim produces Zetesis's own graded reading server-side. Slower, and only needed when the assessment itself is the deliverable rather than the evidence.

verify_attestation re-checks a signed Zetesis record to confirm its claim, evidence and conclusion have not been altered since signing. Needs no account.

Claim classes: genomics and Mendelian randomisation, single-cell, bulk omics, CRISPR screens, clinical trials, real-world evidence, AI clinical decision support, diagnostics, preclinical models, cell and gene therapy, structural biology.

Why the year fence matters

Ask a general model about a 2020 claim today and it answers with years of hindsight; the publication that mattered at the time is buried under everything published since.

Measured on a control claim: unfenced retrieval missed the pivotal publication entirely and scored 35% evidence coverage. Fenced to the claim's own year, the same query set retrieved it and coverage rose to 79%. So the fence is not only about honesty in retrospect. It is a retrieval precision feature.

Set as_of to the year a claim was made for anything that is not brand new.

Try it

What did the published evidence actually support about aducanumab and cognitivedecline at the end of 2019, using only sources available by then?

Then ask the same question without the year and compare. The difference is the point.

Privacy

The evidence tools send nothing to a model provider. evaluate_claim processes claim text through a model sub-processor, named along with retention terms and hosting region in the privacy policy. Claim text is not logged; only metadata (the routed class, depth, counts) is kept.

The Python client

This repository also publishes a thin stdio MCP client to PyPI, which predates the hosted server and exposes an older tool set (evaluate_claim, check_evaluation, verify_attestation, account_status). It holds no keys and runs no model; every call is proxied to the hosted engine, and it needs a token.

Prefer the hosted endpoint above. It needs no token and carries the current tools. The client remains for existing stdio setups:

bash
claude mcp add zetesis --env ZETESIS_TOKEN=zk_... -- uvx --from zetesis zetesis-client

Run from a terminal, zetesis opens the chat above. Launched by an MCP host over stdio, it runs this client, so zetesis and zetesis-client both work in MCP configs.

Tokens: https://api.zetesis.science/request-access

  • ZETESIS_TOKEN sets the token (verification works without one)
  • ZETESIS_API overrides the API base, default https://api.zetesis.science

Security

Report vulnerabilities privately to [email protected]. See SECURITY.md.

MIT licensed. The hosted engine is a separate service.


mcp-name: io.github.reutavidan/zetesis

来源:README.md,提交 5c0d598

工具

4

evaluate_claim

Evaluate a scientific or AI claim

Run Zetesis's graded reading of a scientific, biomedical, clinical, or AI/ML claim. Grades each evidence dimension and returns an overall reading, with every source cited by hard public id (PMID, DOI, NCT, NIH grant, SEC filing). Covers drug, omics, target-validation, diagnostic, and AI-model claims. Runs the hosted engine end to end and typically takes a couple of minutes, long enough that some clients background or time out the call. The output is Zetesis's own reading as a finished artifact. zetesis_scope and zetesis_evidence return the same dated sources and the same grading rubric immediately and without running a model, at full source depth rather than abstract level. A signed, independently re-checkable dossier is available by request at the Zetesis site. Args: claim: the claim to evaluate, in one or two sentences. context: optional background (stage, field, the decision at hand). documents: optional source text (a deck, abstract, or paper). as_of: year the claim was made. Set it for any claim that is not brand new. It restricts evidence to what existed then, which both keeps later outcomes out of the reading and sharply improves what is retrieved: left open, a pivotal trial report competes with every paper published since and gets buried.

输入参数as_ofclaim必填contextdocuments

zetesis_scope

Scope a claim for evaluation

Scope a scientific, biomedical, clinical, or AI/ML claim for evaluation. Returns three things: the source catalog, meaning where the evidence for this claim class lives; the diligence apparatus for that class, meaning the questions a domain reviewer asks and the failure patterns that have caught comparable claims before; and the scoping rubric used to work out which evidence dimensions bear on this particular claim and what to search for. Applies to a research claim, a study, a paper, an abstract, a preprint, a grant or a pitch, and to checking a paper or statistic before relying on it. Runs no model, costs nothing, and returns immediately. The queries worked out from its output are what zetesis_evidence takes as input. With as_of set, the apparatus and rubric are framed for judging the claim as it stood at the end of that year. Args: claim: the claim to evaluate, in one or two sentences. context: optional background (stage, field, the decision at hand). as_of: optional year. Set it to evaluate the claim as it stood at the end of that year, grading how it was derived at the time rather than how it turned out. claim_class: optional override for the routed claim class. The response lists every class; pass one back if the routing picked wrongly, or if nothing routed.

输入参数as_ofclaim必填contextclaim_class

zetesis_evidence

Retrieve dated public evidence for a claim

Step 2 of a Zetesis evaluation. Runs the search queries against public evidence sources (Europe PMC, ClinicalTrials.gov, openFDA, and optionally NIH RePORTER and SEC EDGAR) and returns a deduplicated bundle where every source carries a hard public id, followed by the Zetesis grading rubric so you can grade the dimensions yourself. With as_of set, retrieval is fenced to sources published, registered, or filed on or before 31 December of that year, and two fields that leak later outcomes are suppressed: a trial's present-day status, and FDA labels effective after the cutoff. That makes it possible to judge a claim on what was actually knowable at the time. Runs no model, costs nothing, and returns immediately. Takes the queries worked out when the claim was scoped. Args: queries: 3 to 6 short keyword phrases. Query 1 should be the bare name of the thing claimed about (a drug, compound code, model, or gene); query 2 that name plus at most two outcome words. Longer phrases retrieve commentary rather than the primary report. If the claim names no agent at all (a behaviour, diet, procedure, exposure or policy), anchor on the field's own technical term instead of the lay one ("time restricted eating", not "intermittent fasting"), pair it with the outcome as the field measures it, and allow one design word such as crossover or randomized in one later query. Without an entity name that design word is the only thing separating a trial report from a review of trial reports. as_of: optional cutoff year, as used in zetesis_scope. include_capital: also retrieve funding and public-filing signal.

输入参数as_ofqueries必填include_capital

verify_attestation

Verify a Zetesis attestation

Verify a Zetesis attestation, confirming an evaluation's claim, evidence, and conclusion have not been altered since it was signed. Use when someone has a Zetesis dossier or attestation and wants to independently re-check it. Args: attestation_json: the full attestation object as JSON text.

输入参数attestation_json必填

版本历史

2
  1. v0.2.2最新Oct 17, 2026
  2. v0.2.1Sep 16, 2026

评分

已满足适用检查项的 100%

该评分衡量此收录条目的完整程度,依据注册中心条目、端点与公开来源计算得出。它不是对服务器本身的评价,也不表示服务器安全或当前可用:可连通性是带日期的检查结果,评分也不包含安装量或用户评价。

  • 近 30 天内可连通必需25 分
  • 已声明认证方式必需10 分
  • 已列出工具必需15 分
  • 描述必需15 分
  • 图标10 分
  • 文档15 分
  • 许可证10 分