A research project by Hilal Agil
A catalog has priced the same two things for a century — the recordings, and the songs underneath them. Both are records of what an artist already made. The Siren maps the two nobody has held: how an artist decides, and how they sound — and turns them into assets the artist owns.
How an artist decides
A living memory of an artist’s judgment — what they chose, what they turned down, and why. Not a record of what they made, but a model of how they make the call, built from the work itself and every version behind it. Because it holds the reasons and not just the results, it can be run forward: pointed at where the scene is heading to surface the moves that artist would lean into, reject, or explore next.
How an artist sounds
Their real sound-craft, captured faithfully from the instruments and gear behind it. Real studio hardware has no single setting you can copy — every control bends how every other one behaves. So a fleet of agents sweeps the actual unit across its whole range, measuring what it truly does. What you license sounds like the hardware because it is the hardware, modelled from the real thing, with a clear line back to the device.
As machines become the main thing reading an artist’s work, how you decide and how you sound stop being byproducts. They become the work — and worth owning.
How it holds up
Every part traces back to a real work, a real decision, or a real instrument. Nothing is invented, and nothing is generated in an artist’s name. The Siren does not make music — it makes the licensed inputs an honest AI system has to work from.
Judgment reads sound, but sound never stands in for judgment. Only real decisions carry an artist’s taste; the sound, the market, the press are context the taste reacts against. That firewall isn’t a promise — it’s the shape of the thing, so an industry’s worth of noise around an artist is never mistaken for the artist.
Everything the model believes comes with a source. It shows its work and marks what it isn’t sure of, so it can be checked rather than taken on faith — the difference between a memory a system can depend on and a guess it can’t.
A single artist model is many agents running for a long time. They coordinate through marks left in a shared, sourced memory — like an ant colony, with no central controller to break — so long jobs stay coherent even when a step fails. And the inputs are consented and first-party, not scraped from an open web that’s closing to machines. It keeps working as the platforms around it close.
The business
Masters and publishing have been the only two owned assets in music for a century. The Siren adds a third and fourth — how an artist decides, and how they sound — as things a rights-holder can actually own and license.
The Siren sits with the artist and the label, not against them. As AI becomes the main consumer of an artist’s sound and judgment, these become the inputs any downstream system must license to operate honestly — a new line of revenue built on consent, not scraping.
A licensed check other AI can call to ask whether an output really sits inside an artist’s taste. A signal an investor can weigh when valuing a catalog. A sanctioned way to work with a historic artist’s craft. Uses the industry has wanted and never had a way to own.
We don’t clone the artist. We license their judgment and their craft.