Radar does not calculate meaning
Vector similarity places related prose together and a link graph can mark notes with no inbound references as orphan candidates. Both narrow a curation queue well, but neither can justify merge or archive.
Similar notes may be accidental duplication or a principle deliberately repeated for separate souls. An orphan may be residue or a canonical source that no summary should replace.
Drift is the gap between index and reality
Find files with no index line, index lines with no file, and stale paths after rename through two-way set differences. These are deterministic enough for automatic surfacing.
Fixes still differ. A new note may require a human summary, while a line for a deleted file may permit mechanical removal. The detector classifies gaps; an approved actuator performs changes.
Attach candidate provenance
A Radar card should name the signal: cosine score, exact title collision, or zero inbound links. Threshold and index revision make the candidate reproducible and limit overconfidence.
Several signals may coexist, but do not collapse similarity 0.91 and orphan true into one risk score. They ask different questions and should remain visible in parallel.
Dismissal is evidence too
A false positive can retain reason and fingerprint so the identical cluster does not return every scan. The suppression must expire when content hashes change; otherwise an old judgment hides a new note.
Do not automatically convert every dismissal into a model-training label. Preserve free-form reason and policy version, then validate detector changes separately. Judgment history and training data are different characters.