The first judgment: relevance, importance, and uncertainty

Ripwire’s useful answer is never just the name at the top of a list. The repository boundary, approximate graph, structural ranking, lexical retrieval, and disclosure markers have to be read together. This checkpoint connects those ideas into one practical judgment: what does a ripwire result establish, what does it merely suggest, and what evidence should an agent gather before acting?

That judgment is necessary because the tool deliberately compresses several different kinds of work into a small response. Compression is valuable only when the reader can tell which question was asked and where the answer stops. A high rank is not a universal certificate. It is evidence whose meaning depends on the path that produced it.

One result can answer two different questions

The default map and the task lens are both built from the same repository, but they are not trying to retrieve the same thing. The structural map asks which symbols are load-bearing in the indexed graph. A task query asks which symbols or documents are relevant to the words and concepts in front of the agent. The first is about architecture; the second is about the current investigation.

That difference changes how a coding agent should use the output. If an agent is landing cold and wants to understand the repository’s central machinery, a structurally important symbol is a strong orientation point even if its name does not appear in the task. If the agent is trying to locate the implementation of a feature, lexical relevance is usually the better first filter. A body comment may contain the crucial mechanism while the function name remains generic. Conversely, an exact identifier may be the fastest route to a known definition even when the body contains a great deal of unrelated vocabulary.

The project’s own co-change measurements make the distinction concrete. PageRank achieved 3.8% recall at five on a historical task of finding the other files changed with a commit, while lexical retrieval over bodies reached 40.3%. Adding structural rank to the lexical signal made that result worse. This is not a failure of PageRank as a structural measure. It is a failure to ask one measure to answer a different question. Importance in a call graph and relatedness in a change description are different relationships.

The same principle applies to confidence. Ripwire’s confidence and margin_pct describe how sharply the ranking separates its head from the remaining positive matches. A high margin means the returned set has a material score cliff, or that the positive matches have already been exhausted. It does not mean that the tool has understood the user’s intention with human-level certainty. A flat ranking should be treated as a starting set even when its top row looks plausible; a sharply separated ranking is stronger evidence, but still evidence about the indexed corpus and the chosen retrieval model.

A literal mention can add another useful signal. If a task names a path, module, or Type.method, ripwire can lift the corresponding artifact. That is often appropriate for an issue report that already contains a location. But the lift is a score adjustment, not a guarantee that the artifact becomes rank one. On a flat head, the mentioned item may remain below another candidate.