FAQ

Short answers about how the current runtime behaves, with the safe local workflows kept separate from migration assumptions.

Does analysis change audio files?

No. SONARA, MAEST, MERT, MuQ, CLAP, and classifier scoring write SQLite data only.

Which actions can write audio files?

Three of them. MAEST genre-tag apply, Audio Doctor apply, and Audio Dedup deletion. All three stay separate from normal scan, search, and analysis.

Their gates differ, so be precise about which one you are using:

• MAEST genre apply writes the genre field for tracks with stored MAEST genres, when you start it.
• Audio Doctor apply is dry-run first, backs up each file, verifies the result, and restores the backup on failure. It takes no confirmation phrase, so --apply writes immediately.
• Audio Dedup deletion requires the phrase APPLY DELETE in the request. From the CLI you type it. In the browser the client supplies it and you answer a Yes / No confirmation dialog.

Local-first safety is the full list.

Why does the interface mix Russian and English?

It grew that way. Labels were written in whichever language was at hand as the app was built, so most headings, buttons, and tooltips are Russian while model names, tab labels, mode names, and numeric field labels such as Limit are English. There is no language setting. This documentation is English throughout, and UI language maps every control from one to the other.

What does deleting a track remove?

The catalog row, its FTS entry, and every catalog table that references it by track_id, which covers tags, likes, embeddings, fingerprints, classifier scores, and feedback. The optional Evaluation database is a separate file with no foreign key back to the catalog, so its search_session_seeds and search_result_events rows for that track are deleted in a second step. Audio Dedup deletion also clears matching Rhythm Lab label rows. Source audio is removed only by Audio Dedup deletion, never by DELETE /api/tracks/{track_id}.

Can the app migrate an incompatible database?

Normal startup refuses an incompatible legacy split database and never rewrites it. Stop the server and your database tools, then run dj-sim migrate-database --db .\data\library.sqlite --confirm 'MIGRATE SINGLE LIBRARY'. It creates a timestamped backup, then builds and verifies the replacement. It does not run analysis or reanalysis.

What follows a SONARA change?

Adapt the project to the SONARA fields you want. Reanalyze only the affected outputs, and only when you choose to. Retrain, promote, and rescore a SONARA-dependent classifier only when its feature recipe changed.

Why are classifier artifacts blocked?

Runtime scoring needs an artifact whose ordered feature recipe and inputs match your current stored data. Affected artifacts have to be retrained and promoted. Incompatible scores are never silently reused.

Can I use the browser UI with a compatible database?

Yes. Build the current frontend bundle and serve it with the current backend. Database selection, library paging, analysis, search, Current Set, CLASS, LAB, Rhythm Lab launch, metadata, preview, and exact-identity writes all use the current typed responses.

Library loading is one fixed page of up to 200 tracks per request, with previous, next, and page-number navigation. There is no second row-window paginator and no incremental chunk loading inside a page. Sort direction and shuffle reorder the loaded page in the browser rather than requesting a new one.

Can the app build a set for me?

No. The current set is a list you build by hand, one track at a time, and export in the order you put them. Automatic sequencing from anchors, a mood arc, or a story prompt is a direction rather than a feature. See Project idea for the boundary.

Can I share reports or databases?

Review them first. They can contain local paths, tags, scores, and your listening decisions.