在病毒基因组中，在整个分类体系中保持一致地命名蛋白质 家人。

公共序列数据库保存了数万个病毒基因组，其蛋白质 名字由成千上万个不同的提交者组成，所以是同一个蛋白质 有数十种拼写，许多甚至没有实用名称。低范 模块用一个受控集合取代了它：一组PSSM配置文件 还有一个JSON，声明分类单元拥有哪些蛋白质、它们有多长以及具体长度 匹配必须自信。

这个仓库包含了构建新病毒模块所需的一切 分类单元，运行它，看看它是否真的有效——再加上代理技能 走过这一切。

每个已发布页面及其URL均被收录在 ARTIFACTS.md。 覆盖审计、PSSM注册、字符串崩溃和词汇饱和 十个分类单元的页面。它们托管在 claude.ai 上，而非这里;该仓库存放 在reports/下提供源代码HTML及链接。

构建一个注释串主要是判断：哪些注释字符串代表相同的蛋白质， 哪些集群是真实的，哪些特征数据无法支持。**技能 存在是因为这些判断很容易在某些方面出错，没有任何后续影响

LowVan module kit

Name the proteins in a viral genome, consistently, across a whole taxonomic family.

Public sequence databases hold tens of thousands of viral genomes whose protein names were written by thousands of different submitters, so the same protein appears under dozens of spellings and many carry no useful name at all. A LowVan module replaces that with one controlled set: a collection of PSSM profiles plus a JSON declaring which proteins a taxon has, how long they are, and how confident a match must be.

This repository is everything needed to build such a module for a new viral taxon, run it, and find out whether it actually works — plus the agent skill that walks through doing so.

Every published page and its URL is indexed in ARTIFACTS.md. Coverage audits, PSSM registries, string-collapse and vocabulary-saturation pages for ten taxa. They are hosted on claude.ai, not here; this repo holds the source HTML under reports/ and the links.

Building one is mostly judgement: which annotation strings mean the same protein, which clusters are real, which features the data cannot support. The skill exists because those judgements are easy to get wrong in ways nothing downstream detects.

Where the work lives

This repository is the source of truth. build/ and annotate/ are the pipeline and the annotator; skill/ is the method; modules/ holds the built modules; reports/ holds the published pages.

CEPI-dxkb/Viral_Annotation is downstream. It will be updated from this work eventually, but that is the maintainer’s call and happens only once everything here is correct — it is not a target to push to as modules land.

Until then it should not be treated as an input either. At the time of writing it has no internal_stop support at all, so a module relying on a readthrough is silently cropped against it. Do not build against a fresh clone of it, and do not treat patches/ as the current state: patches are diffs against pristine upstream and drift behind build/ and annotate/ as work accumulates. patches/ exists to make the eventual downstream update reviewable, not to describe what the kit does now.

To set up a runtime directory for the annotator, copy annotate/* and build/* over it first. See modules/README.md.

Does it work?

The Rhabdoviridae module built with this kit, measured against every Rhabdoviridae genome BV-BRC holds between 1 and 50 kb — 25,237 records collapsed to 1,356 distinct genomes, one per 95%-identity cluster:

reports/ holds four standalone HTML pages with the full evidence. GitHub will not render them in place — download the file, or use a raw-HTML viewer.

| Report | What it answers |