The Siren

Living models of artistry for the AI era: judgment, the craft beneath it, and the coordination that keeps them coherent.

Abstract

A music catalog has priced exactly two things for a century: masters — the recordings — and publishing — the compositions. Both are records of what an artist already made. Neither captures how the work was made, how the artist would decide next, or how any of it should be governed once machines, not people, become the main thing reading it.

The Siren builds living models of artistry that AI and ML systems can read and reason on. This paper describes the two authored assets those models are built from — the Siren Graph, a continuously-updating memory of an artist’s judgment, and Siren Acoustics, a captured model of an artist’s sonic craft — and the coordination layer that keeps a fleet of agents working over them coherent across long horizons, without heavy reliance on general-purpose models or fragile access to the open web.

Neither asset generates music. Both are human-authored, provenance-tracked inputs — the licensable inputs that any downstream system reasons on. The Siren sits on the rights-holder side of the line.

Two more layers of the catalog
A living memory of an artist’s judgment
Inferred from work, not from interrogation
The craft, captured — not the preset
Agents that sweep the space a human cannot
Coordinating the work without a central planner
Standing on the music-knowledge research
Building for a closed-platform era
On the rights-holder side of the line

1. Thesis — Two more layers of the catalog

Masters are what an artist made. Publishing is the composition underneath. Both are finished, static, and already fully priced. What the industry has never held as an asset is the layer above the output — the judgment that chose it — and the layer beneath it — the craft that produced its sound.

The Siren maps an artist’s judgment, sound, career, and world into distinct, ownable, licensable layers. Judgment, career, and world resolve into one queryable state we call the Siren Graph. Sound — the reproducible engineering of tone — resolves into Siren Acoustics. Together they are a third and fourth layer of the catalog: not what the artist made, but how they decided and how they sounded, as things a machine can read.

This is a forward thesis. Catalog valuation still prices only masters and publishing today. The claim of this paper is that as machines become the primary consumers of artist-sound and artist-judgment, these layers become the assets that the previous two always implied.

2. The Siren Graph — A living memory of an artist’s judgment

The Siren Graph is the whole queryable state of an artist at any moment — press, charts, popularity, scene, collaborators, career, life — organized around one core: the artist’s decision function. It is not a style model of what the artist tends to produce. It is a model of what they chose, what they rejected, and why.

At the center of the graph sit the only taste-bearing elements: the decisions, the values behind them, and the constraints that bound them. Around that core is the world the judgment reacts against — works, influences, collaborators, career events, performances, media, market conditions, eras — and the sources that make every claim citable.

Because the graph holds not just the choices but the rejected alternatives and their rationale, it can be run forward. Given a projected field — where the scene, the market, and the moment are moving — the decision function can be evaluated against moves the artist never made, surfacing the ones their own revealed preference would stretch toward, reject, or explore. The projection is speculative and marked as such; it is inference over a memory, never a recording.

3. Judgment capture — Inferred from work, not from interrogation

A decision function inferred from output is only credible if it is captured without leaning on the artist. Capture is layered by cost to the artist. The primary layer reads finished works for their signature. A second layer reads the differences between versions — the takes cut, the mixes held, the features dropped — because rejection is where taste is most legible. A rare third layer surfaces first-person rationale only when the inference is uncertain and a direct source resolves it.

This is abductive work: reasoning from fragmentary evidence back to the most likely account of how a decision was made, in the tradition of inference-to-the-best-explanation rather than pattern-matching on surface output. Every belief in the model carries a confidence and a provenance. The graph does not assert; it cites. This is the difference between a memory a system can depend on and a generated guess it cannot.

4. Siren Acoustics — The craft, captured, not the preset

Siren Acoustics is the sound asset: an artist’s authenticated sonic craft — the tones, the EQ moves, the processing and mixing technique, the way they engineer a sound rather than any single static preset. It is the doing layer, made ownable and reproducible.

The hard problem underneath it is fidelity. A physical piece of audio hardware — a compressor, a preamp, a filter, a tape machine — is a continuous, interacting system. Its behavior is not a lookup table of knob positions; it is the way every control bends the response of every other control, across the full range of signal fed into it. Sampling a device at a handful of settings captures a shadow of it. The number of interacting parameter combinations is far beyond what any engineer could exhaustively dial and audition by hand.

5. Agentic capture — Agents that sweep the space a human cannot

Siren Acoustics captures a physical unit by running a fleet of agents across it. Each agent drives the hardware through parameter combinations — and combinations of combinations — that no human would reach by hand, measuring the device’s true response at every point rather than interpolating between a few sampled settings.

The agents do not sweep blindly. They concentrate their runs where the device behaves non-linearly — the regions where controls interact, where the response bends, where a naive sample would be most wrong — and thin out where the behavior is predictable. The result is a model of the instrument’s actual sonic behavior across its whole operating space, built from authored measurement of the real unit, at a resolution and coverage a manual capture cannot reach.

This is capture, not generation. The agents are an instrument trained on a real device; the output is a faithful model of authenticated hardware, with provenance back to the physical unit that was measured. Nothing here invents a sound the device could not make.

6. Coordination — Keeping the work coherent without a central planner

A single artist model is the work of many agents running over a long horizon — capturing sound, reading versions, updating the graph, checking a proposed use against taste. Orchestrating that with one central controller is brittle: the moment a step fails or the field shifts, a rigid plan breaks.

The Siren coordinates the way social insects do, and the way recent multi-agent research has shown works for music — through the shared environment rather than a central planner. Agents leave traceable, typed facts in a shared memory; later agents read those traces and adapt. Coordination emerges from the marks left in the work, not from a controller dictating every move. Each fact carries its source, its confidence, and its scope, so the memory is auditable end to end and one agent’s finding can be trusted, questioned, or superseded by the next.

This is what lets The Siren run long, autonomous jobs — an end-to-end capture, a catalog-wide re-read, a licensing check across a body of work — that stay coherent even when individual steps fail or the surrounding data moves.

7. Standing on the research — Music-native, not general-purpose

The Siren is built to be domain-native rather than a wrapper around a general-purpose model. It draws on and extends the public state of the art in three areas, and cites that lineage plainly:

Music knowledge representation. The field now has mature, interoperable ways to describe artists, works, performances, and provenance — flexible metadata models designed for the music domain, aligned with established library-science entity distinctions and with open general-knowledge bases for scale. The Siren’s living memory builds on that interoperability so its graph can speak to the rest of the music-data world rather than reinventing it.
Stigmergic multi-agent coordination. Recent work has shown that coherent, long-form musical structure can emerge from a swarm of agents coordinating through shared environmental traces, with no central planner — and that federated, typed, provenance-tagged memory lets agents on different platforms reason over a shared foundation. The Siren’s coordination layer is grounded in that line of work.
Abductive process modeling. Reconstructing how a creative decision was made from the evidence it left is inference-to-the-best-explanation, not classification. The Siren’s judgment capture is built on that stance.

Naming the lineage is deliberate. The defensible work is not any single borrowed idea; it is the artist-grounded system that unifies judgment, sound, and coordination into licensable assets — and the private craft of how those assets are actually built and kept honest.

8. Building for a closed-platform era

The open web is closing to machines: APIs deprecate, anti-bot measures tighten, and terms increasingly forbid the scraping that earlier tools assumed. A system that depends on fragile external access is a system that breaks.

The Siren is built to not depend on that. Its inputs are consented and first-party, its memory is internal and provenance-secured, and its agents coordinate over that internal fabric rather than over a live feed of the open web. This is both an ethical stance — license the inputs, don’t scrape them — and an engineering one: a music-native foundation that keeps working as the platforms around it close.

9. What this is not — On the rights-holder side of the line

The Siren does not sell AI-generated music. Every asset is human-authored, with provenance back to a real work, a real decision, or a real device. The Siren sits on the rights-holder side: it licenses the authored inputs — real sound-craft and real judgment — that downstream systems must license to operate legally.

The models are living because they update continuously, not because they resurrect anyone. We map a practice — judgment, craft, work, and how they evolve — not a person. We don’t clone the artist. We license their judgment and their craft.

Terms — The load-bearing vocabulary

Living model of artistry — A continuously-updating representation of an artist’s judgment, sound, work, and world that an AI or ML system can read and reason on.

Siren Graph — The judgment asset: the whole queryable state of an artist at a moment, with the audited decision function at its core.

Siren Acoustics — The sound asset: an artist’s authenticated, reproducible sonic craft, including agent-captured models of the physical hardware behind their tone.

Decision function — The model of how an artist chooses — what they select, reject, and why — inferred from work rather than interrogation.

Stigmergic coordination — Agents working through traces left in a shared, provenance-tracked memory, so the work stays coherent without a central planner.

Run it forward — Evaluating the decision function against a projected field to surface the moves an artist would stretch toward, reject, or explore — inference over a memory, never a recording.

License the inputs, not the output.