reachprobe

Overview ¶

Command reachprobe asks why a gold file is never retrieved.

The ceiling decomposition leaves two retrieval gaps of similar size: 13.1% of gold files sit in the response below rank 20 (a ranking problem) and 14.3% are not retrieved even among two hundred results. The second is the one no lever has touched, and the fix depends entirely on WHY the file is invisible:

semantic  — the query's vocabulary appears nowhere in the file. Only a
            better representation reaches this: a stronger embedder, or
            generated queries per document.
lexical   — the words ARE in the indexed text and the lexical channel still
            ranked the file out of two hundred. Term weighting or expansion.
uncovered — the words are in the file but NOT in what we index, because the
            body cap truncated them or no symbol covers that span. Neither
            ranking nor a better model reaches this; the indexer does.

Every share is reported against a control — the same statistic over the gold files we DID retrieve — because "identifiers appear in 60% of missed files" means nothing until you know the figure for the found ones.