Read similarity scores as suggestions

Scores answer "which candidates should I inspect first under these settings?" They can tell you that one candidate ranked ahead of another inside the same search. They cannot judge musical quality or promise that a transition will work.

The practical output of a score is an audition order. Listen from the top, keep useful exceptions, and stop when you have enough candidates.

This page owns the rules for keeping score spaces apart. Other pages link here.

Eight scales that never mix

One search means one evidence source and one score space. The app produces eight distinct scales, and a value from one says nothing about a value from another.

Scale	Where it appears	What it measures
SONARA Core similarity	SONARA tab, /api/search/sonara	weighted agreement across named audio features
ML seed cosine	SIMILARITY tab and LAB, one value per family	cosine between L2-normalized embeddings of one family
CLAP text contrast	PROMPT tab with CLAP selected	prompt vector against stored CLAP audio vectors
MuQ-MuLan text contrast	PROMPT tab with MuQ-MuLan selected	prompt vector against stored MuQ-MuLan audio vectors
Audio Dedup min_similarity	Audio Dedup embedding mode	content gate over MERT, MAEST, MuQ, and CLAP vectors
fingerprint_similarity	Audio Dedup reports	native SONARA acoustic fingerprint agreement, 0 to 1
Transition risk	Evaluation transition diagnostics	weighted mean of mix-risk components
Classifier probability	CLASS filters	probability of a promoted classifier's positive label

A threshold tuned on one of these is meaningless on another. The 0.45 fingerprint floor in Audio Dedup and the 0.45 low-confidence tempo cutoff in SONARA share a number and nothing else.

Even inside one family, the question changes with the input. CLAP text search and CLAP seed search read the same clap_embeddings table, and their scores are still separate: a text score is prompt against audio, a seed score is audio against audio.

ML seed search

The SIMILARITY tab searches one embedding family at a time over MAEST, MERT, MuQ, or MuQ-MuLan. The query is the L2-normalized mean of the selected seed rows, the seeds are removed from the results, and the score is a cosine over unit vectors. Values fall in -1 to 1 in principle, and useful neighborhoods sit well above zero in practice.

CLAP is available for seed search through the HTTP API and in LAB, without an entry in the browser model selector.

SONARA feature search

SONARA search compares stored feature values under mixer weights, with optional modifier bias on top. It explains itself better than an embedding cosine, and it is still a similarity model. Raising a mixer weight changes the ranking question rather than improving the answer.

How a Custom score is assembled

Custom mode is the mode where every part is visible:

• Five mixer groups carry a weight from 0 to 5: Timbre and Rhythm default to 1.0, Dynamics and Harmonic to 0.8, and Tempo to 0.35.
• Each numeric dimension is normalized against a 2nd to 98th percentile band computed over the candidate set of that request. Values outside the band clamp to the edge, so one outlier track cannot flatten the scale.
• A vector field splits its weight across its components, so 13 MFCC values cannot outvote the rest of the timbre group.
• Nine modifiers run from -1 to 1. A modifier weight is |direction| x 2.5, and the field a modifier drives leaves the group similarity so the two do not cancel each other.
• Aggression is the one modifier attenuated by evidence confidence, and it drops out entirely when SONARA stored no aggression confidence for a candidate.
• A candidate is dropped when fewer than two numeric dimensions overlap with the seed.

Harmonic scoring uses Camelot compatibility attenuated by key confidence. Key confidence is a reliability weight rather than a similarity dimension of its own. The harmonic group in Custom mode uses lighter tonal weights than Balanced and DJ transition do.

DJ transition adds a structural term

DJ transition mode blends the feature similarity with a directional structural fit:

score = 0.8 x similarity + 0.2 x transition_fit

transition_fit compares the seed outro against the candidate intro, the two energy levels, and the stored energy-curve summary. Missing parts are left out of the mean rather than counted as zero, so a track with partial structure data is not penalized for the gap.

Text search and the contrast score

With no negatives and a single positive prompt, the score is the cosine between the prompt vector and each stored audio vector.

With a negative bank, the visible number is a contrast score:

contrast = positive - negative_weight x mean(two highest negative cosines)

The negative term is the mean of the two highest negative similarities, not the single strongest one. Asking a second negative prompt to agree before a track is pushed down measured better than the maximum at every weight for both text models. A bank holding one negative collapses the two terms into the same value, which reproduces the older behavior.

The default negative weight is 0.5, and the API accepts 0 to 2. A contrast score can go negative and is not a probability. The result rows show the positive and negative parts separately so you can see which side moved a track.

Rank fusion between CLAP and MuQ-MuLan was measured and rejected. The app never blends them, and the preset picker reports a conflict rather than mixing two banks across models.

Reference Compare scores

The LAB panel keeps CLAP, MERT, MuQ, MuQ-MuLan, MAEST, and SONARA in six separate groups for one seed track. Compare scores within one model group first. A high MERT score, a high MuQ-MuLan score, a high CLAP audio score, and a high SONARA score are related listening hints from four different measurements.

Saved LAB verdicts are manual pair-feedback labels for a specific model source. They record what you heard. They are not automatic truth labels and they do not rewrite the model score.

Audio Dedup thresholds

Audio Dedup min_similarity is an audio-to-audio content gate over the enabled stored MERT, MAEST, MuQ, and CLAP audio embeddings. MuQ-MuLan is not one of its sources. MuQ evidence alone does not satisfy the separate safe-delete corroboration rules.

The fingerprint_similarity value in dedup reports is another separate scale. At 0.45 or above it creates manual-review candidates and never authorizes deletion.

Two ranking controls that do not change the score

The HTTP search API carries two filters with no browser control:

• epsilon keeps only candidates within that distance of the best score.
• noise from 0 to 1 adds a deterministic jitter derived from the catalog and track UUIDs. It perturbs the ranking order only, and the score returned in the response stays untouched. The same request returns the same order every time.

Practical reading

• Compare scores inside the same tab and the same settings.
• Be careful after changing thresholds, weights, or prompts.
• Preview audio before adding to a set.
• Do not use one tab's threshold as another tab's safety rule.
• Use LAB when you want to compare model families by ear before trusting one for a reference track.

Related pages

• Features, embeddings, and tags
• Search with seed tracks
• Text search