How It Works

BananaRat Studio is the AI music platform for copyright-conscious songwriters. That's a claim about process, so this page shows the process: which models we use, what the Studio deliberately declines to do, and what's stamped on every song that leaves it. Everything below describes something that ships today — most of it you can verify in your own downloads.

The model

A song here is made in two stages, and we're specific about both.

The writing — lyrics, chord charts, tablature, notation, MIDI — comes from a multi-agent composition pipeline built on Anthropic's Claude models. It works from music theory and your direction, and it outputs real notation. Your MIDI, MusicXML, and stems are then rendered deterministically from that notation — not approximated from a finished mix afterward.

The produced audio — the full track with vocals — is rendered by ACE-Step 1.5, an open-weights music model released under the MIT license. We run it on our own hardware, in our own studio; the audio generation isn't handed to a third-party service. Because the weights and license are public, anyone can inspect exactly what the base model is. We didn't train it, so we won't characterize its training data secondhand. What we can verify — and show you — is the license it ships under, where it runs, and the filtering we built around it.

What we deliberately won't do

The Clean Room filter screens what you type and what you import before anything reaches the audio engine:

  • Ask for a song in the style of a named artist, and the Studio declines — then offers descriptor sliders so you can describe the sound you're after without borrowing anyone's name.
  • Imported lyrics pass through a two-tier similarity screen: a blocklist of known song lyrics, backed by an AI classifier. A near match pauses the upload and asks you to confirm; a strong match blocks it, with an in-app appeal that goes to a human reviewer.
  • The melodies the Studio generates run through a separate similarity gate against known melodies — the output side of the same idea.
  • Every upload carries its own declaration — original work, licensed cover, public domain, or written permission — and each declaration is recorded with the song.
  • After composition, an output copyright screen checks the finished lyrics against known songs, and the result is recorded with your song — you see what it found, every time.

Honest framing: these are screens, not certainty. They will occasionally flag work that's genuinely yours (that's what the human-review appeal is for), and no filter catches everything. What they establish is that declining to imitate is built into the pipeline, not written into a policy page.

Provenance on every song

Every produced track carries AuthorMark — two independent signatures plus a session record:

  • C2PA Content Credentials — a cryptographic manifest embedded in the file: authorship claim, platform identity, timestamp, and a machine-readable "do not use for AI training" assertion. Check any track yourself at contentcredentials.org/verify or with the open-source c2patool.
  • AudioSeal neural watermark — inaudible, embedded in the audio itself, and built to survive transcoding and partial edits. A second, independent channel if the manifest is ever stripped.
  • The Witness record — a timestamped log of every creative decision in the session, feeding the Authorship Evidence gauge on every song.

This is how the Studio meets the EU AI Act Article 50 and California SB 942 machine-readable disclosure requirements — in place ahead of their August 2, 2026 date.

What counts as your work

Every song here starts from an AI-generated draft — what the gauge tracks is what you add to it. As you work, the Studio logs each creative decision: the lyrics and ideas you bring, the key, tempo, genre, and structure you set, the sections you regenerate or keep, your edits and restyles. That log — the Witness record — feeds the Authorship Evidence gauge, which weighs four kinds of contribution:

  • What you wrote and changed — your lyrics, imports, edits, and refinements. This carries the most weight.
  • How you shaped the song — the key, tempo, genre, arrangement, and restyle choices that set its form.
  • What you chose — picking among alternatives, keeping one take over another.
  • Time in the session — how long you worked it, how many passes you made.

The weighting and anti-gaming safeguards were set with input from music counsel: no single action can carry the gauge on its own, so the score reflects work you actually did, not a box you ticked. The more of the song you made, the more the record shows — a one-tap draft reads low, and the gauge says so. It exports as a split sheet a registrar or a co-writer can review. The gauge is a record of your work, not a ruling on it.

We don't train AI on your work. Your compositions, lyrics, uploads, and prompts are never used to train, fine-tune, or improve any AI model — that's a commitment in our Terms, and every track carries a machine-readable do-not-train credential.

What you get, and what stays yours

Full commercial rights come with every tier — including the three free trial songs. Release them, license them, pitch them to clients; there's no upgrade step that changes what you can do with a song you already made. Each song ships as a complete session package — produced vocals, lyrics, chord charts, tabs, MIDI, MusicXML, 48kHz stems, and a Pro Tools handoff — and the Witness record makes your authorship documentation registration-ready from the start.

Registration and royalty eligibility depend on jurisdiction and the registrar's decision. Evidence pack, not legal advice.

What we don't claim

We won't tell you any AI music tool is free of legal risk, because no one can honestly say that. Our screens are best-effort engineering: they reduce the chance of imitation, and they document that reduction — they don't make it impossible. The Authorship Evidence gauge measures documented activity, not legal conclusions. And one beta note in the same spirit: our C2PA signatures are cryptographically valid today, but until our signing certificate joins the public trust lists at general availability, verifiers may report "signed by an unknown authority, signature valid."

What we claim is narrower, and checkable: an inspectable open-weights base model run on our own hardware, filters that decline imitation by design, provenance credentials on every render, and a documented record of the human work in every song. Process you can verify — that's the promise.

Nothing on this page is legal advice.