CuriationLEGAL

Curiation · Legal

How Curiation Uses Your Music

Companion to the Library Agreement (v1.1)

A plain-English companion to Section 3A of the Library Agreement. If the two ever disagree, the Agreement controls — but nothing here is a softer version of anything there.

We analyze your music so buyers can find it. We do not, and will not, build anything that makes music.

Those two sentences describe opposite technologies that unfortunately share a verb. Here is the difference, because it is the whole answer.

When your track is uploaded, our system listens to it and produces an embedding — a list of numbers describing where that recording sits in acoustic space: its texture, energy, instrumentation, era, mood. Think of it as a fingerprint, or a set of coordinates.

You cannot turn an embedding back into music. It is a one-way transformation, like a hash. It has no melody in it, no audio, nothing that could be reassembled into a cue. Its only function is comparison: this track is near that track, and far from this other one.

That is what makes search work. When a buyer types "warm 70s soul with horns," or drops in a reference track and asks for something close to it, the system is comparing coordinates — not consulting a model that has memorized your catalog.

  • Semantic search. Buyers describe what they need in words — "warm 70s soul with horns" — and get the right cues, without needing to match your tagging vocabulary.
  • Automatic tagging. Genre, mood, instrumentation, tempo, key, era — applied consistently across your catalog, including gaps in your existing metadata.
  • Similarity and the map. Acoustic clustering that lets a buyer explore by sound rather than by keyword, and find neighbors of a cue they already like.
  • Ranking. Learning from what buyers search, preview, and license, so the right cue surfaces sooner.

We're telling you this so the permission you're granting is clear, not to claim it exists today:

  • Reference-track search ("sounds-like"). A buyer drops in a temp track, a competitor's spot, or any clip, and gets acoustically similar cues from participating catalogs. This is how supervisors actually work, and it's the feature most likely to put your music in front of someone who would never have typed the right words.
  • Structure detection. Sections, cue points and natural edit points, so a buyer can jump to the drop instead of scrubbing — and so cutdowns get easier to generate.
  • Intake quality checks. Loudness, duplicates, corrupted files, missing metadata — caught before a buyer sees them.
  • Provenance screening. Flagging fully AI-generated uploads, so the human-made promise stays true for everyone on the platform.
  • Claim-release support. Faster identification of your recordings when a buyer hits a false Content ID claim. Today, claim release runs through you — we route the buyer's notice, you release it in your own Content ID or Rights Manager account.

Written into the Agreement as binding restrictions, not as a policy we could quietly revise:

  • Never train, build, or run any model that generates or outputs audio. Not with your music, not with anyone's.
  • Never build a system your recordings could be reconstructed from. Our models produce search results — rankings, tags, similarity. No output of any Curiation model is, or contains, audio.
  • Never sell, license, or share your music — or the data derived from it — with anyone as training material. No dataset business, no AI-company deals, not for any price.
  • Never synthesize or imitate a named artist's voice or style.
  • Never build anything that competes with your catalog.

Our models learn across all participating catalogs. That is not a loophole — it is the point. A search engine that only understood one library could not tell a buyer that your 1972 soul instrumental is the closest thing to the reference they just dropped, because it would have nothing to compare against.

What crosses between catalogs is understanding of sound, never audio, never files, never anything a competitor could use to make a cue like yours. Your recordings stay yours, stay on your storefront, stay priced by you.

The analysis runs in your favor too, and it is yours to keep:

  • Every tag, enriched metadata field, structural marker, and quality report we generate for your catalog is available to you to export and use anywhere, including off Curiation, at no charge. Most libraries pay a vendor for exactly this.
  • Consistent metadata means your catalog is findable by buyers whose vocabulary doesn't match yours.
  • Being discovered by sound rather than by keyword is how a small, distinctive catalog competes with a large, generic one.
  • Demand signal: what buyers searched for and didn't find is information about what to commission next.

If you leave, your catalog comes out of the searchable marketplace and your recordings — audio and stems — are removed from our storage. You take your files, your metadata, our enrichment, and your transaction history with you; your export rights run during the term and for 90 days after it ends.

Being straight with you about one thing: the derived representations — the embeddings and feature vectors described above — may be retained after you go. They are one-way and non-reversible: no audio can be reconstructed from them, they are never used to train anything that outputs audio, and they are never sold, licensed, or shared with anyone as training data. But we would rather tell you plainly that they persist than imply a deletion we don't perform.

Questions: cupchurch@curiation.fm · Haven Contemporary LLC, d/b/a Curiation