Why we do not publish the model

By Yoshi Moon, Chief Executive Officer, RadioPro LLC

Every few weeks someone writes to ask how AuralVerse works. Researchers ask for the architecture. Prospective partners ask for the weights so they can run the model themselves. Journalists ask what it was trained on. The answer to all three requests is the same, and this note sets it out once so that we can point to it.

RadioPro LLC publishes three things about AuralVerse: the economic analysis of the model, which has been public since the 1.4 release; the release notes for each numbered version; and the standards we evaluate our models against. We do not publish the architecture, the weights or the training material, and we decline requests for them. The model is deployed inside our group of companies and affiliated networks, and access to it runs through a confidentiality agreement. What follows is why.

The model is the company's principal asset, and the law treats it as one

The legal category for what we withhold is the trade secret. United States federal law defines one as "all forms and types of financial, business, scientific, technical, economic, or engineering information, including patterns, plans, compilations, program devices, formulas, designs, prototypes, methods, techniques, processes, procedures, programs, or codes," provided the owner "has taken reasonable measures to keep such information secret" and the information "derives independent economic value" from not being generally known. (18 U.S.C. 1839(3)) The European Union's definition in the Trade Secrets Directive is to the same effect, and it reaches information that is secret "in the precise configuration and assembly of its components," which is what a curated training corpus is. (Directive (EU) 2016/943, Article 2(1))

The World Intellectual Property Organization put the point directly in its 2026 guide for smaller companies: trade secret protection "could be particularly effective if third parties do not have direct access to the AI system and only the AI's output is shared outside an organization. For example, weights and parameters of an AI model could be protected as confidential information." (WIPO, Learning Machines, 2026) That is a description of how AuralVerse is deployed. The protection exists only as long as the secrecy does. Practitioners have warned that "trade secret protection over model parameters or training samples may be forever lost if new threat models emerge that make it possible to reverse engineer such information." (Maisel and Stewart, Law360, March 2024) The reasonable measures the statute requires are therefore not a formality. They are the asset.

The largest developers in the field do the same

We are sometimes told that a research company should publish its methods. The companies that define the state of the art do not. OpenAI's technical report on GPT-4 states: "Given both the competitive landscape and the safety implications of large-scale models like GPT-4, this report contains no further details about the architecture (including model size), hardware, training compute, dataset construction, training method, or similar." (OpenAI, GPT-4 Technical Report) Anthropic's model card for Claude 3 describes the training data only as "a proprietary mix." Google's model card for Gemini 2.5 Pro names an architecture family and nothing more.

In music the pattern is the same. Google DeepMind's announcement of Lyria 2 says that all output is watermarked and says nothing about how the model is built. ElevenLabs states that Eleven Music "is trained on licensed stems and music" and stops there. Stability AI says Stable Audio 2.5 is "trained on a fully licensed dataset" and gives no architecture. When Suno replaced its models with a licensed generation in September 2026, its product chief declined to say how they work. The companies that publish most about their models publish what the model was licensed to learn from and how its output is marked. None of them publish the model.

Access to the output is already an attack surface

The reason for caution is not hypothetical. In March 2024 a team led by Nicholas Carlini described "the first model-stealing attack that extracts precise, nontrivial information from black-box production language models," recovering a layer of a commercial model "given typical API access" for "under $20 USD." (Carlini et al., Stealing Part of a Production Language Model) In February 2026 Anthropic reported "industrial-scale campaigns" to extract its model's capabilities through roughly 24,000 fraudulent accounts and over 16 million exchanges, and in September it described a single campaign that peaked at nearly 3 million exchanges a day. (Anthropic, February 2026; Anthropic, September 2026)

A model that can be queried at will can be copied in part. A model whose weights are published can be copied in full. AuralVerse is available neither way. It is deployed where RadioPro LLC controls the environment, to parties who have signed for it, and that is the measure the statute calls reasonable.

What you disclose about training can be used against you

The year's copyright decisions make a second case for discretion, and it is the one our legal team weighs most. In Bartz v. Anthropic the judge's order was built from the company's own records of what it had downloaded and from where, and it concluded that "there is no carveout, however, from the Copyright Act for AI companies." (Order on fair use, 23 June 2025) Two days later, in Kadrey v. Meta, a different judge wrote that "in many circumstances it will be illegal to copy copyright-protected works to train generative AI models without permission," which means that companies "will generally need to pay copyright holders for the right to use their materials." (Order, 25 June 2025) In Munich, GEMA won against OpenAI in November 2025 and against Suno in July 2026 on the basis that works reproducibly inside a model are reproductions of them. Both judgments are under appeal. Neither turned on how the models were built. Both turned on what came out and what went in.

Our position on inputs is simple. AuralVerse performs lyrics and compositions that rightsholders bring to it. The work it renders is licensed by the fact of being brought. We do not describe the material the model learned from beyond that, and we do not intend to.

What the law actually requires us to say

Discretion is not the same as silence, and the law sets the floor. The European Union's AI Act requires providers of general-purpose models to publish "a sufficiently detailed summary about the content used for training." The Commission's own template states that it "does not require disclosure of the details for the specific data and works used to train the model," that "different levels of detail are required in the Template depending on the source of data considered," and that "the precise manner in which providers curate the data and train their models is preserved." (European Commission, Explanatory Notice and Template, July 2025) Where that duty applies to a model we place on the European market, we will publish a summary at the level the template requires. We will not publish more.

The Act's transparency rules for synthetic audio apply from 2 August 2026 and require that output be "marked in a machine-readable format and detectable as artificially generated or manipulated." (Regulation (EU) 2024/1689, Article 50(2)) The platforms moved the same way this year. Apple Music told its partners in August that AI transparency tags will be required wherever AI "was used to create a material portion of the content," and YouTube Music began labelling songs "created primarily using generative AI" on 14 October. We support all of it. A label on a recording tells a listener what they are hearing. It tells a competitor nothing about the system that made it. That is the line we draw, and it is the same line the regulators and the platforms have drawn.

The argument on the other side, stated fairly

Rightsholders' organisations want more. A coalition of European creative-sector bodies including IFPI and CISAC called the Commission's template "a betrayal of the EU AI Act's objectives" and asked for Article 53 to be enforced "in a meaningful way." (Joint statement, 30 July 2025) CISAC's director general said in Athens last October that ingesting works without consent "is not training. This is theft," and set out three demands: "transparency, consent, and remuneration." (Gadi Oron, October 2025)

We agree with two of the three without reservation. Consent and remuneration are the basis on which AuralVerse operates: the composition is the rightsholder's, it is performed because they asked, and the economics of that arrangement are the subject of the analysis we published. On transparency, we think the demand is right and the object is wrong. A rightsholder is entitled to know whether their work was used. They are not entitled to the architecture of the system that used it, and no law gives them that. We will answer the first question to anyone with standing to ask it. We will not answer the second to anyone.

How partners get access

Publishers, composers and production companies that want to work with AuralVerse write to us through the contact form. Technical detail about the model is shared only under a confidentiality agreement, and only to the extent the work requires. Requests for weights, architecture or training material, from any party for any purpose, are declined, and we would rather say so here than in a hundred separate replies.

RadioPro LLC is a private company, and we intend to keep the one thing that makes it worth being one.