Classifier scores and the deferred CLASS tab

Audience: Users with promoted Rhythm Lab classifier profiles. Goal: Filter, rescore, and use classifier scores without confusing them with analysis families. Type: guide

Use a classifier when you keep making the same personal judgment and ordinary tags do not capture it. Examples include a profile you define for vocal-forward versus mostly instrumental tracks, or for a specific kind of live instrumentation. Rhythm Lab learns from your labeled examples; the main app then stores a reusable score for the rest of the compatible library.

This score does not discover an objective fact about music. It reflects a local boundary learned from one profile and its labels. Use it to reduce review work, then listen to borderline and important tracks.

For a one-off sound idea, use seed or text search instead. A classifier is most useful when the same question will return across many sessions.

What you get in the main app

1. You create and label a profile in Rhythm Lab.
2. Training produces candidate artifacts that you review and promote.
3. The main app scores compatible tracks for that promoted profile.
4. The CLASS tab can filter by the stored score.
5. SET can prefer, avoid, raise, or lower the concept through a preview.

The backend reads promoted local profiles only when their manifest is valid and compatible with current scoring inputs. The React frontend has not yet been ported to v7, so the CLASS tab behaviors described below are design intent, not a verified current UI contract. Use the CLI or direct API for current scoring.

How a promoted classifier is stored

A promoted classifier uses one pointer to an immutable model and manifest generation:

models/classifiers/<artifact-prefix>/current.json
models/classifiers/<artifact-prefix>/generations/<generation-id>/model.joblib
models/classifiers/<artifact-prefix>/generations/<generation-id>/model.json

Promotion verifies and syncs both generation files before atomically switching current.json. Discovery rejects an incomplete generation or any pointer, manifest, or model hash mismatch.

The manifest describes the classifier key, labels, model id, calibration status, required inputs, and optional Hybrid signal metadata. Version 2 also requires the exact SONARA analysis signature for any SONARA-dependent feature set.

The promoted artifacts currently checked into models/classifiers/ use manifest version 1. Runtime discovery keeps them visible with a blocker, but scoring requires retraining and promotion to version 2.

Filtering

Each promoted classifier appears with a slider from 0.00 to 1.00. The library browser can filter tracks by minimum stored score for that classifier.

Missing classifier scores do not pass a positive minimum filter. In SET and Hybrid modifiers, missing scores stay neutral.

Rescoring

The play button on a classifier row resets and rescans that one classifier key. The UI calls the reset path first, then starts /api/classifiers/{classifier_key}/analyze.

Classifier scoring is database-only. It reads exactly the SONARA and MAEST/MERT/CLAP inputs declared by the promoted manifest and writes Core classifier_scores. It never decodes audio and never runs inside a SONARA or ML job.

Scoring is blocked when the promoted artifact, manifest, and track do not share the same current SONARA signature. Retrain and promote legacy SONARA profiles after reanalysis. Labels and feedback remain available.

CLASSIFIERS analysis stage

The left panel includes a standalone CLASSIFIERS checkbox. Selecting it clears SONARA and ML selections; FULL is the explicit way to combine all stages in one ordered pipeline. The single Analyze action scores compatible profiles. The UI shows manifest-specific ready/not-ready counts and blockers. Missing inputs exclude a track before the job total is formed and do not create a partial score. Empty aggregate selection means all compatible promoted classifiers.

SET and Hybrid

SET classifier preferences are signed. Negative values avoid the classifier concept, while positive values prefer it. Flow can stay flat or move upward/downward through the set.

Hybrid preview can use classifier preference and risk metadata when a promoted manifest exposes a compatible signal. The Hybrid details panel shows whether classifier support was available, fresh, stale, missing, or neutral.

When no profiles appear

Promote a profile from Rhythm Lab or place model.json plus model.joblib under models/classifiers/<profile>/. See Rhythm Lab.