Novelty Estimation
CORPUS is being built in the open. Some of what you read here is live, some is still design intent. Expect it to evolve.
Novelty Estimation is the measurement behind the diversity dimension of scoring. It answers one question: how far does this contribution sit from everything already in the corpus? Because that answer directly affects what a contribution earns, this page lays the mechanism open in full.
A note on words: we say novelty, not originality. Originality carries a romantic story about lone creators that no measurement can honor; what a machine can measure is distance from what already exists. The distinction is argued at length in Originality Is a Story We Keep Telling on the CORPUS Journal.
Novelty is relational
A track is not novel on its own, any more than a number is large on its own. Novelty is a relationship between one piece of music and the body of music around it. The system therefore never asks "is this creative?"; it measures how far a track sits from a defined collection, and that collection is the corpus itself: the whole living library, tens of thousands of tracks across dozens of genres, growing as contributions come in. The breadth matters. Against a narrow catalog, a fingerpicked folk demo can look wildly novel just because little folk sits beside it. Against the full corpus, that kind of accident washes out.
Picture it as a cloud
Picture the whole corpus as one vast cloud, each track a single point in it. Tracks that sound alike sit close together: two pop songs with a similar beat land side by side; a pop song and a lone traditional instrumental drift far apart. Novelty is simply how far your track sits from the dense middle of that cloud. Land near the center and your track sounds like much of what already exists: low diversity score. Land out at the sparse edge and it adds a direction the corpus barely covers: high diversity score. Everything below is how that distance gets measured precisely. The picture is the whole idea.
How the measurement works
- Embedding. Thirty seconds of each track run through an open audio model (currently Music2Vec) that was trained on large amounts of music without any labels. From one of its inner layers we read a compact vector of a few hundred numbers. In the space of these vectors, tracks that sound alike sit close together and tracks that sound different sit far apart, across dimensions no human labeled by hand.
- Distance. For the corpus as a whole, the system fits a statistical model of the resulting cloud of vectors: its center, and how it spreads in every direction. A new track's raw score is its distance from that center, measured in units of the corpus's own spread (a Mahalanobis distance). Sitting at the edge of a dimension where the corpus already varies widely counts for little; sitting outside it in a direction the corpus never explores counts for a lot.
- Calibration. The distance passes through a calibrated curve to become a novelty value between zero and one.
There is no trained "originality classifier" anywhere in this chain, and no generative model passing judgment. Labels for originality do not exist, and any we invented would encode one person's taste. Geometry can be inspected; a verdict cannot.
Three readings per track
Newness rarely lives in a track as one undivided thing. A conventional chord progression can sit under a beat no one has heard, or a familiar groove under a strange harmonic language. Before embedding, each track is therefore split into harmonic and percussive parts, and three distances are measured separately, each against its own corpus model:
- Timbral: the overall sonic gestalt of the unsplit track.
- Harmonic: melodic and harmonic language.
- Percussive: the rhythmic profile.
These are the three axes described in The Scoring System in Detail. A track that is timbrally familiar but rhythmically out on its own produces a visibly different profile from one that is strange all the way through.
A moving reference
Novelty has a shelf life: the first record to do a particular thing is new, the hundredth is a genre, the thousandth is a cliché. The corpus model is therefore not a fixed snapshot. It pulls in embeddings for new music as it arrives, drops what leaves, and refits whenever the collection changes. A track ingested today is measured against today's corpus; the same track a year later would meet a different reference.
For your score this cuts in your favor: the novelty reading is taken once, at ingest, and the resulting diversity bonus is protected for five years before it decays toward a permanent floor. See Temporal Dynamics.
What the number is not
Distance from the corpus is not musical quality, and it is not artistic worth. A corrupt file, a field recording of a washing machine, and a visionary album can all sit far from the center, and the geometry will not tell them apart. That is why novelty is only one dimension of the score: integrity filtering happens before any scoring, and craft is assessed separately by the Production Quality Assessment.