The method / System One Search
A repository contains the pieces of its explanation. The harness finds them, judges their relevance and connects them.
Code retrieves, ranks and builds the graph. Jev answers small typed questions about the evidence. No LLM writes the explanation.
Inspect the questionsThe construction of an answer
Tracked files become a repository tree. A content pass extracts declarations, imports, calls, comments and other language-specific facts. Compact descriptions put distinctive facts first.
Files become searchable evidence.
A lexical index scores paths, facts and source text. A static reference graph connects files. These propose candidates before a model judgment is needed.
Recall starts with ordinary code.
For Find, TypeSafe shortlists files and verifies them against relevant source lines. For Map, it judges membership in a subject. Question text stays constant; repository data lives in the request state.
Typed judgments refine the candidates.
Explain gathers members, expands through references, and judges evidence blocks and edges. Code builds the final chart. Titles are identifiers, summaries are selected comments and arrows come from real references.
An inspectable route through the code.
Reading the atlas
A judgment is a probability, not proof. A flow shows what was found within its search depth and caps. It is a static reference view, not an execution trace.
Dynamic loading, dependency injection and client–server boundaries can leave gaps. Extractors are approximate. The trace, omitted units and source references help you assess the result.
Coverage checks report which paths a harness says it examined and which recorded neighbors remain. They do not establish semantic completeness.
Read the limitationsNow follow a real question
Questions about public repositories are logged to improve the demo. Don’t include secrets.