# Moses Yebei - repository estate > 1712 public repositories, analysed and related to each other. > Provenance is stated per fact rather than claimed for the site as a whole. > The file census, architecture pass, dependency scan, advisory scan, hygiene > audit, domain classification and eight-category grade are EXTRACTED: measured > from a tree, a history or a manifest, reproducible from the same inputs, no > model involved. The semantic similarity edges are INFERRED: cosine distance > between neural embeddings (nvidia/nemotron-3-embed-1b) of a text built from each > repository description, topics and generated summary. Treat them as a strong > hint, not as a measurement. Every edge in the relation layer carries which of > the two it is. Repositories that have not been analysed are marked as such > rather than being reported as clean. ## Data - [Schema](https://moses-y.github.io/data/schema.json): what every field in the index means. Read this first; the record keys are single letters. - [Index](https://moses-y.github.io/data/index.json): all 1712 repositories with language, domain, kind, capabilities, hygiene severities and semantic coordinates, plus 4024 similarity edges. Large - prefer the neighbourhood files below for traversal. - [Relations manifest](https://moses-y.github.io/data/relations.json): edge types, thresholds and counts for the relation layer. - [Clusters](https://moses-y.github.io/data/clusters.json): 44 groups of near-duplicate work covering 116 repositories, each with member grades and which one is the keeper. - [Cluster report](https://moses-y.github.io/data/clusters.md): the same 44 groups written out as prose, cross-domain ones first, each with its keeper and how much of it is unaudited. Read this instead of clusters.json if you want the argument rather than the records. - [Neighbourhood](https://moses-y.github.io/data/kin/1004131322.json): one file per repository id, about 1 KB each: nearest semantic kin and repositories sharing its dependency stack, with the shared packages named. Fetch https://moses-y.github.io/data/kin/.json to walk from one repository to its neighbours without loading the index. - [Grades](https://moses-y.github.io/data/grades.json): 1682 graded repositories, mean 61.3, with the per-category score, weight and the findings charged against each. - [Hygiene](https://moses-y.github.io/data/hygiene.json): the raw audit findings behind the grades. - [Dependencies](https://moses-y.github.io/data/deps.json): declared dependencies per repository by ecosystem. - [Search](https://moses-y.github.io/data/search.json): inverted index, token to position in the index records. ## Answering a question with these files 1. Resolve a repository to its numeric id once, from /data/index.json or /data/search.json. Ids are stable across renames. 2. Fetch https://moses-y.github.io/data/kin/.json, about 1 KB. It carries both edge types with their provenance, so "what else is like this" costs one small request rather than a parse of the 1298 KB index. 3. Repeat step 2 for any neighbour id. Traversal never needs the index again. 4. For a group rather than a neighbour, /data/clusters.json names the keeper; /data/clusters.md is the same thing as prose. Prefer a neighbour that appears in both the stack and semantic lists of a kin file: the extracted edge corroborates the inferred one, which is the strongest signal here that a repository is a usable starting point. ## Pages - [Projects](https://moses-y.github.io/projects.html): every repository with its measured facts. - [Code Graph](https://moses-y.github.io/code-brain.html): every repository by domain, language or meaning (similarity edges drawn), down to its modules. - [Briefings](https://moses-y.github.io/blog/): a written analysis per repository at /blog/.html. - [Site map](https://moses-y.github.io/sitemap.html): everything, grouped by domain. ## Notes - Domains: Web & Interfaces 574, AI & Data 562, Systems & Infra 385, Knowledge & Content 72, Agent Skills & Plugins 61, Mobile 57, Other 1. - Two edge types relate repositories, and they differ in how much they can be trusted. "stack" is EXTRACTED: IDF-weighted overlap of dependencies both repositories declare in a manifest, and each edge names the packages it was drawn from, so it can be checked and disagreed with. It answers "built the way you build things". "semantic" is INFERRED: cosine distance between embeddings, with no evidence attached beyond the number itself. It answers "probably solves the same problem". A pair scoring on both is the strongest signal that one repository is a usable starting point for the other, because the extracted edge corroborates the inferred one. - Clusters in clusters.json are built from semantic edges only, so they inherit that provenance: they are INFERRED. Method is Louvain modularity over the thresholded semantic edges: a group is a set of repositories linked more densely to each other than to the rest of the estate. That is a stronger claim than the connected components this previously used, which chained through bridge repositories, but it is still a claim about an embedding and not about code. A large group is a set of closely related projects, not a set of duplicates. - Stack edges exist only for the 599 repositories that declare dependencies in a manifest this pipeline reads. - Grades cover 1682 of 1712 repositories. An absent grade means not yet audited, not a bad grade. - Generated 2026-11-19.