People send me these decks. I have never worked out why. Somebody in the ecosystem has decided I am approachable.
Every deck contains the phrase “trained on publicly available data.” I ask what that means. They say the internet. I ask whether they downloaded the files. They explain that the model learns patterns and does not store copies. I ask again whether they downloaded the files.
They downloaded all the files. They always download the files. There is no universe in which somebody built a music model by staring intently at Spotify and thinking hard.
What follows is a build guide. It is correct. Follow it and you will have a working generative music model, plus roughly four and a half million discrete acts of copyright infringement under the Copyright Act 1968 (Cth). I have included the section numbers of the Act. You are going to want them.
Read the whole thing in order. The legal part only makes sense once you understand what the pipeline is physically doing to the audio, and I refuse to explain it twice.
And by the way, if I you think I get it wrong (which I am certain I am not), call it out. Because this is a topic that needs public discourse.
Part One: The Corpus
Everything downstream is decided here. This is the slide your founder skips because it contains the word “acquisition” and he has classified that word as Legal’s problem.
Scale
A model that produces two minutes of coherent, prompt aligned music wants north of 20,000 hours of audio. Meta trained MusicGen on about 20,000 hours. Stability trained the Stable Audio family on several hundred thousand tracks from a licensed production library. OpenAI’s Jukebox used roughly 1.2 million songs scraped off the web with metadata and lyrics attached, which was a very 2020 thing to admit in a paper.
Call it 100,000 hours if you want outputs that stop sounding like a fax machine achieving enlightenment. At four minutes a track that is 1.5 million recordings.
What that costs in bytes
100,000 hours of 32 kHz mono 16 bit PCM is 23 terabytes. That array sits in a colo for eighteen months while you train, evaluate, retrain, and eventually discover that your data loader has been silently dropping every file with a Unicode apostrophe in the filename. You find this out in month eleven. Someone says “good catch.” You begin quietly researching agriculture.
Acquisition
This is where everyone gets vague, so I will be specific. Teams get audio four ways.
1. A command line downloader pointed at a streaming platform, run across a scraped list of track identifiers.
2. Torrents. Usually a discography dump, usually FLAC, usually with the original release metadata intact, which is convenient for captioning and apocalyptic during discovery.
3. Ripping a subscription service by capturing the decrypted stream.
4. Licensed production libraries and buyout catalogues.
Option four is lawful in Australia. Options one through three are the reason you will read the second half of this post with a slightly dry mouth.
Preprocessing
Once the files are on disk:
• Decode everything to PCM. MP3, AAC, FLAC, Opus, all of it becomes float32 or int16 arrays.
• Resample to one rate. 32 kHz if you are cloning the MusicGen recipe, 44.1 kHz stereo if you want something a human will pay for.
• Loudness normalise to a target LUFS, around minus 14, with an EBU R128 implementation. Skip this and the model learns that loud means good and quiet means intro. You will have built a mastering engineer from 2006.
• Segment into fixed windows. 30 seconds with a 15 second hop is standard. Each window is written to disk as its own file.
• Trim silence and run voice activity detection, so you are not spending 8 percent of your compute budget modelling the gap between tracks.
• Deduplicate. Chromaprint or a learned embedding, nearest neighbour search, drop anything over a cosine threshold. Duplication is the single strongest predictor of memorisation.
• Separate stems if you want melody or instrument conditioning. Demucs gives you drums, bass, vocals, and other. A chromagram off the “other” stem gives you a melody conditioning signal.
Look at what steps five and six actually did. Windowing a four minute recording produces 8 new files. Source separation produces 4 more. You started with 1.5 million recordings. You now hold somewhere between 12 and 50 million derived audio files, and every single one of them contains the recognisable hook of a commercial release.
Remember that number. It comes back with interest.
Captioning
The model needs text control, so every window needs a description. Three approaches.
5. Metadata. Genre, mood, year, tempo, key, instrumentation, artist name, lifted from ID3 tags or a music database.
6. Automatic captioning. Run PaSST or a music captioning model over the audio, hand the tag list to a language model, have it write a sentence. Roughly the LP MusicCaps recipe.
7. Joint embedding pseudolabels. A CLAP style text and audio encoder aligns windows to a caption vocabulary.
Approach one is the entire reason your model responds to “in the style of Tame Impala.” There is no emergence here. A contractor in another timezone typed the band name into a CSV column, and you paid him eleven dollars an hour to construct your primary legal exposure. Your CTO will later describe this as the model “developing an understanding of genre.” He is describing a spreadsheet.
Part Two: The Tokenizer, Which Is A Photocopier With A Research Budget
You cannot run a transformer over 1.4 million raw samples per 30 seconds. You need compression. The answer is a neural audio codec.
Architecture
A residual vector quantised autoencoder. Three parts.
Encoder. Strided 1D convolutions with residual blocks, SEANet style. For 32 kHz input at a 50 Hz frame rate your total downsampling factor is 640. Six hundred and forty samples in, one latent vector out.
Quantiser. Residual vector quantisation. Take the latent, find the nearest entry in codebook 1, subtract it, find the nearest entry in codebook 2 for the residual, subtract, repeat. MusicGen’s codec runs 4 codebooks of 2048 entries. Four codebooks at 50 Hz is 200 tokens per second. At 11 bits a token that is a 2.2 kbps representation of music, which is an elegant piece of engineering that you are about to use for crime.
Decoder. The encoder in reverse. Transposed convolutions back to waveform.
Training the codec
Losses, summed with tuned weights:
• L1 or L2 reconstruction on the raw waveform
• Multiscale mel spectrogram reconstruction across several STFT window sizes
• Adversarial loss from a multiscale STFT discriminator
• Feature matching on discriminator intermediate activations
• Commitment loss holding encoder outputs near their assigned codebook entries
Initialise codebooks with k means over a batch of encoder outputs. Update with an exponential moving average. Use quantiser dropout so the codec degrades gracefully at lower bitrates. Watch codebook utilisation like it owes you money, because collapse is the default outcome and it is completely silent. Your reconstructions will sound fine while three of your four codebooks are dead. The Descript Audio Codec solved most of this with factorised, L2 normalised codes and snake activations. Use their approach. Yours will be worse.
What the loss function says
Read the loss function again. Slowly.
The objective is to make the decoder output indistinguishable from the encoder input. You are training a network whose sole reason for existing is to reproduce a commercial master with as little perceptual difference as achievable. You then measure how well it did that, and you use the measurement to make it better at doing it.
Anybody who says “the model just learns patterns” after reading that is lying to you or has never opened the file. The pattern is the song. A human being wrote that objective function on purpose, in Python, at work, for money, and then went to a standup and said the codec was converging nicely.
Part Three: The Generator
Two architectures matter. Pick one.
Option A: Autoregressive transformer over codec tokens
The MusicGen approach. A decoder only causal transformer predicting the next codec token. 300 million to 3.3 billion parameters, scaling with how much of somebody else’s money you have.
The complication is that each frame carries 4 tokens, one per codebook, and they are hierarchically dependent. Three ways to deal with it.
• Flattening. Serialise all 4 codebooks into one stream. Correct, and it multiplies sequence length by 4. A 30 second generation becomes 6000 steps and your inference bill becomes a topic at the board meeting.
• Parallel prediction. Predict all 4 codebooks at each step independently. Fast. Ignores the dependency between codebooks entirely. The audio sounds like someone dropping a stack of plates down a stairwell with a very nice reverb on it.
• Delay pattern. Offset codebook k by k steps, so predicting codebook 2 at step t happens with codebook 1 at step t already known. One step per frame, 1500 steps for 30 seconds, dependencies intact. This is what MusicGen does. The other two are there so you can feel like you made a decision.
Option B: Latent diffusion
The Stable Audio approach, and where the field has gone.
Train a VAE compressing audio into a continuous latent, typically 64 channels at around 21.5 Hz for 44.1 kHz stereo. Then train a diffusion transformer to denoise in that latent space. Current builds use flow matching or rectified flow instead of the original DDPM formulation, because it trains more stably and distils down to a handful of sampling steps.
Variable length generation runs on timing conditioning. Feed the model a start offset and a total duration as conditioning signals and it learns where intros, drops, and endings go. That one detail separates a model that emits 30 seconds of ambient wallpaper from a model that emits a song with a shape. Most demos you have heard skipped it.
Option C: Symbolic
Train on MIDI. Tokenise with a REMI style event vocabulary of note on, note off, time shift, velocity, bar markers. A 100 million parameter transformer does a credible job. It runs on one GPU. It will impress nobody at a demo day.
Read Part Ten before you dismiss it.
Part Four: Conditioning
Text conditioning is a frozen text encoder, usually T5, feeding cross attention layers in the generator. CLAP style joint audio and text embeddings work better for timbre and production style, because they were actually trained on audio.
Classifier free guidance is mandatory. Drop the conditioning signal 10 to 20 percent of the time during training so the model learns the conditional and unconditional distributions. At inference, run both and extrapolate:
logits = uncond + guidance_scale * (cond - uncond)
Start around a guidance scale of 3. Push it to 7 and you get excellent prompt adherence, audible artefacts, and a dynamic range of about two decibels. Your growth lead will prefer this version. Your growth lead should not be permitted near the sampler.
Melody conditioning takes a quantised chromagram off the separated instrumental stem as an extra conditioning stream. That is the mechanism behind “make it sound like this artist but follow this tune,” a request which is doing several kinds of work at once and which we will revisit under s 195AC.
Part Five: Training
100,000 hours at 200 tokens per second is 72 billion tokens. A serious language model corpus, assembled entirely out of other people’s masters.
Configuration that works:
• AdamW, betas 0.9 and 0.95, weight decay 0.1
• Cosine learning rate schedule, linear warmup over the first few thousand steps
• Gradient clipping at 1.0
• bf16 mixed precision
• Flash attention, gradient checkpointing, fully sharded data parallel or ZeRO stage 3
• Effective batch size in the millions of tokens via gradient accumulation
Budget tens of thousands of A100 hours for a 1.5 billion parameter model at that data scale.
Your bottleneck is the data pipeline. It is always the data pipeline. Decoding and augmenting audio on the fly starves the GPUs and you spend forty thousand dollars a month heating a building in Oregon. Pretokenise the entire corpus with the frozen codec, write the token streams to sharded binaries, stream those. Utilisation goes from 30 percent to 90 percent and you stop having the dream about the fans.
Part Six: Inference
Autoregressive sampling: top k around 250, temperature 1.0, classifier free guidance on the logits. Generate tokens, run them through the codec decoder, get audio.
Diffusion sampling: DPM Solver or a flow matching Euler sampler. 50 to 100 steps for quality, distilled to 4 to 8 steps for anything you serve at scale.
Long form generation uses a sliding window. Generate 30 seconds, keep the last 10 as a prefix, generate the next 30 conditioned on it, crossfade the overlap. Structural coherence past about 90 seconds remains hard. That is why every demo you have sat through was 45 seconds long and ended on a fade. Nobody fades out because it sounds good.
Part Seven: Evaluation
• Fréchet Audio Distance between embedding distributions of generated and reference audio. The original VGGish implementation correlates badly with human judgement. Use CLAP or PANNs embeddings.
• KL divergence over tag distributions from a pretrained classifier such as PaSST.
• CLAP score, cosine similarity between prompt embedding and generated audio embedding. Measures whether the model did the thing you asked.
• Human evaluation. MUSHRA or pairwise preference. Everything above is a proxy and everyone in this field knows it.
Now look at that list and tell me what is absent.
None of those metrics detect memorisation. FAD improves as your outputs resemble the training distribution, so a model regurgitating training tracks verbatim scores beautifully and turns your evaluation dashboard a pleasant green.
The standard evaluation suite for this field cannot detect the one failure mode that determines whether your company exists in three years. Everybody in it knows that. I find that clarifying about the people involved.
Part Eight: Memorisation
Every legal question below turns on this.
A model trained with a maximum likelihood objective assigns high probability to sequences that appeared in training. Sequences that appeared repeatedly get very high probability. Carlini and colleagues extracted verbatim training images out of Stable Diffusion, and duplication rate in the corpus was the dominant factor.
Music corpora are heavily duplicated. The same master shows up as the album cut, the single edit, the 2011 remaster, the 2019 remaster, four compilation appearances, a radio rip, and forty two uploads from a channel called Chill Vibes Radio 24/7. Hash based deduplication catches none of it, because a different encoder produces a different hash for audio that is perceptually identical. Your dedup pass reported 3 percent. Your dedup pass was wrong.
How to actually test:
8. Generate a large sample set across your prompt distribution.
9. Embed every generation and every training window in a shared audio embedding space.
10. Approximate nearest neighbour search from each generation into the corpus.
11. For every hit over threshold, do exact alignment. Cross correlate the generation against the matched window and look at the peak.
12. Report the tail. The mean is always fine. The mean is a lovely place where nothing bad has ever happened. The tail is where the lawsuit lives.
Run this before you ship. Decline to run it and understand what you have chosen: a collecting society, a label, or a detection vendor runs it instead, on your public model, after your revenue figures have become discoverable. Their tooling is better than yours, because it is the whole of their business and you do it as a chore between sprints.
Part Nine: The Australian Audit
Now the part you scrolled to.
The setup
Australia does not have fair use.
I am going to sit here for a second, because roughly nine out of ten people building AI products in this country believe otherwise, having assembled their understanding of copyright from American podcasts.
Australia has fair dealing, which is a closed list of specific purposes: research or study, criticism or review, parody or satire, reporting news, and legal advice. A purpose absent from that list is not a defence. There is no residual category and no vibes clause. Section 200AB adds a flexible exception available to libraries, archives, educational institutions, and people with a disability. You are none of those things. You are a Series A.
In October 2025 the Attorney General confirmed there will be no text and data mining exception. On 15 July 2026 the Prime Minister said it again at the University of Sydney, announced a mandatory national AI framework with legislation expected in early 2027, and described training on Australian music without the artist’s control as theft. The Attorney General’s consultation on licensing models is running now, with statutory licensing, collective licensing, and voluntary regimes on the table.
There is no exception, and there is none coming. The Prime Minister said the quiet part into a microphone at a university. Plan accordingly.
What a song is, legally
A recorded track is not one piece of property in Australia. It is at least three, owned by different people, on different clocks.
Subject matter
Provision
Typical owner
Relevant right
Musical work (the composition)
s 32, Part III
Composer and publisher, reproduction administered by AMCOS
s 31(1)(a): reproduce in a material form, publish, perform in public, communicate to the public, make an adaptation
Literary work (the lyrics)
s 32, Part III
Lyricist and publisher
s 31(1)(a), same list
Sound recording (the master)
s 89, Part IV
The maker, usually the label. Performers co own recordings made on or after 1 January 2005 under s 22(3A) and s 97(2A)
s 85(1): make a copy, cause to be heard in public, communicate to the public, commercial rental
Section 13(2) makes the right to authorise any of those acts part of the exclusive right itself. Section 14(1) makes doing any of those acts to a substantial part legally identical to doing it to the whole thing.
Substantial part in Australia is qualitative. In EMI Songs Australia Pty Ltd v Larrikin Music Publishing Pty Ltd [2011] FCAFC 47, two bars of “Kookaburra Sits In The Old Gum Tree” turning up in a flute line in “Down Under” was held to be a substantial part.
Two bars. A flute line. In a song from 1981. The Full Federal Court agreed, and Larrikin took a percentage of Down Under for the rest of its natural life.
Your training window is thirty seconds long and contains the chorus.
The audit
Step 1, acquisition. Writing the file to disk is a reproduction in a material form of the musical work under s 31(1)(a)(i), a reproduction of the literary work if there are lyrics, and the making of a copy of the sound recording under s 85(1)(a). Three infringements per track. At 1.5 million tracks, 4.5 million primary infringements achieved before anyone has written a line of model code.
Ripping a stream protected by an access control measure engages s 116AN, circumvention of a technological protection measure. Separate statutory cause of action, separate remedies, with criminal provisions at s 132APC and s 132APE covering circumvention devices and services. Breaching a platform’s terms of service is contract law, and it incinerates any story you were preparing about innocent infringement.
Fair dealing for research or study under s 40 does not help. De Garis v Neville Jeffress Pidler Pty Ltd (1990) 37 FCR 99: a commercial press clipping service was held to fall outside research or study within the meaning of the provision. Commercial product development is the same shape with worse margins.
Sections 43A and 43B cover temporary reproductions incidental to a communication or to a technical process of using a work. A corpus living on your array for eighteen months is not temporary in any sense the Act recognises, including the sense your general counsel is currently hoping for. Section 43B also requires the underlying use to be non infringing, which returns you to the top of this paragraph.
Step 2, preprocessing. Decoding to PCM is a reproduction. Resampling is a reproduction. Loudness normalisation is a reproduction. Writing 8 windows a track is 8 reproductions of a substantial part. Source separation produces 4 more, and transcribing a melody to a chromagram or MIDI is an arrangement or transcription, which is an adaptation of a musical work under the s 10(1) definition and is separately reserved to the owner.
Recall the number from Part One. Between twelve and fifty million derived files, every one of them a separate act.
Step 3, captioning. Scraping lyrics reproduces the literary work. This is exactly what the Munich Regional Court found against OpenAI in November 2025 over song lyrics in ChatGPT. Building artist names into your conditioning vocabulary adds s 195AC, false attribution of authorship, and s 18 of the Australian Consumer Law if the output is presented in a way that implies an association.
Step 4, codec training. See Part Two. Your loss function is defined as reproduction fidelity. A round trip through the trained codec emits a perceptually equivalent copy of a commercial master.
Section 10(1) defines a reproduction in a material form to include any form of storage, visible or not, of the work or a substantial part of it, and it applies whether or not the work can be reproduced from that storage. The 2006 amendment deleted the requirement that the stored form be capable of reproducing the work. That provision was drafted before any of this existed and it reads like it was drafted last Tuesday specifically at you.
Read s 10(1). Then open your codec checkpoint. Then form a view.
Step 5, generator training. Gradients computed on tokens derived from unlicensed copies. Whether the resulting weights are themselves an infringing reproduction is the live global question, and two courts answered it opposite ways within nine months on the evidence in front of them.
Getty Images (US) Inc v Stability AI Ltd [2025] EWHC 2863 (Ch): the English High Court held a model is an “article” for secondary infringement purposes, and held the Stable Diffusion weights were not an infringing copy, because on the evidence the weights did not store the claimant’s images. Getty abandoned its training and output claims mid trial, which tells you something about how the evidence was going.
Case 42 O 763/25, decided 31 July 2026: the Munich I Regional Court found for GEMA against Suno and prohibited four separate acts across six compositions. Reproduction for training carried out in the United States. Reproduction by memorisation inside the model. Communication to the public by offering the model. Reproduction and communication in the outputs. The court took jurisdiction over the US training acts through a venue rule available to collecting societies, applied US law to those acts, and rejected fair use. Penalties run to 250,000 euros per breach. Under appeal.
Both courts asked one question and answered it on the record before them. London had evidence the weights held nothing. Munich had evidence the weights held six songs. Which of those describes your model is a factual question about your model, and Part Eight tells you how to answer it this week for the price of some GPU time and a bad afternoon.
Step 6, serving. Making the model available online to users in Australia is a communication to the public under s 31(1)(a)(iv) for works and s 85(1)(c) for sound recordings, to the extent that what goes out contains a substantial part. Section 22(6) puts responsibility for a communication on whoever determined its content. That is you, not the user.
Users prompting in the style of a named artist generate more exposure, and s 36(1) makes you liable for authorising infringement. Section 36(1A) directs the court to weigh the extent of your power to prevent the act, the nature of your relationship with the user, and whether you took reasonable steps including compliance with relevant industry codes. Roadshow Films Pty Ltd v iiNet Ltd [2012] HCA 16 turned on power to prevent, and iiNet won that case because a general purpose ISP could not prevent much.
You control the weights, the prompt filter, the output filter, the logs, and the terms of service. Your power to prevent is total. Raise that at the next standup and watch the room discover an urgent need for coffee.
Step 7, outputs. Substantial part again, qualitative, per Larrikin. Plus the adaptation right for musical works.
Then moral rights, which are personal, inalienable, and held by individual humans under Part IX. Attribution under s 193. Protection against false attribution under s 195AC. Integrity of authorship against derogatory treatment under s 195AJ and s 195AK. Performers hold moral rights in their recorded performances too.
Your terms of service contain a clause in which everybody waives all moral rights worldwide in perpetuity. I have not read yours. I do not need to. It came from a template, and the template came from a US firm, and the US does not do moral rights properly. In McCallum v Projector Films Pty Ltd (Liability Hearing) [2026] FCA 173, Shariff J held that blanket waivers of exactly that kind are unenforceable in Australia, unsupported by the text, context and purpose of Part IX and contrary to public policy. What actually works is specific, informed consent under s 195AW, directed at identified acts and identified works.
Go and read your clause. It is a mood board.
Remedies
Section 115(2) gives the owner damages or an account of profits. Section 115(4) adds additional damages, directing the court to weigh flagrancy, the need to deter, the defendant’s conduct after being put on notice, and any benefit accruing from the infringement.
Every element on that list is satisfied by a scraped corpus, a circumvented protection measure, and a growth chart you presented at a conference with your name on the badge.
Section 132AC creates an indictable offence for commercial scale infringement prejudicing the copyright owner, carrying 550 penalty units or five years imprisonment for an individual. Directors who direct the conduct are exposed personally as joint tortfeasors. Section 115A lets rights holders obtain orders blocking offshore online locations, a playbook this country’s rights holders have exercised repeatedly and enjoy.
The offshore argument
Copyright is territorial. Infringement under the Act requires the act to be done in Australia. Train in Oregon, serve to Australia, and acquisition and preprocessing sit outside the Act. This is the argument your advisor is going to sell you at a rate.
The communication to Australian users stays squarely inside it. Sections 37 and 102 cover importing an article for commercial dealing where making that article in Australia would have infringed. Authorisation liability attaches to conduct here. And Munich just demonstrated a court reaching offshore training through a domestic jurisdictional hook, which is the sort of thing that spreads. Australia has no equivalent collecting society venue rule, so that specific route is untested here.
Untested means nobody has run it at you yet.
Part Ten: You Do Not Own The Output
This is the part that kills the business, and I have never once seen it in a deck.
Copyright in Australia requires a human author who exercised independent intellectual effort. IceTV Pty Ltd v Nine Network Australia Pty Ltd [2009] HCA 14 established the independent intellectual effort requirement. Telstra Corporation Ltd v Phone Directories Company Pty Ltd [2010] FCAFC 149 held that a compilation generated by software with no identifiable human author attracts no copyright at all. Acohs Pty Ltd v Ucorp Pty Ltd [2012] FCAFC 16 confirmed the same for computer generated source code.
A track generated entirely by your model from a text prompt has no human author. It therefore attracts no copyright in Australia. Anyone may copy it, sell it, sync it into a national television campaign, and you have no cause of action whatsoever. You cannot register it with APRA AMCOS, because registration requires an author and there is not one.
So the position is this. You infringed several million copyrights to build a machine whose output belongs to nobody the instant it exists.
Sit with that. I did. It was the best part of my week.
If your marketing tells customers they own their generations and hold exclusive commercial rights, that is a representation about legal rights made to a consumer, and s 18 and s 29 of the Australian Consumer Law are sitting there, patient and unamused.
The symbolic MIDI model from Part Three sits differently, by the way. Training on transcriptions still reproduces the musical work. It touches no sound recordings, which deletes the entire Part IV layer and the label side of that table above. Fewer rights holders, smaller table, cheaper conversation. Nobody builds it because it does not demo well, and demos are how this industry makes decisions.
Part Eleven: The Version That Is Legal
Everything technical in Parts Two through Eight survives unchanged. The parts that change are the ones nobody asked the engineers about.
Licence at the source, both layers. You need sound recording rights from the label or aggregator, and the reproduction right in the musical work from the publisher or AMCOS. A public performance licence licenses public performance. It does not license reproduction for training, and no Australian blanket licence covers AI training today. Production libraries with buyout terms that expressly name AI training are the cleanest source available. GEMA launched a licensed training dataset in Germany in July 2026. Suno agreed to deprecate models trained on unlicensed music as part of its Warner settlement in November 2025. Universal settled with Udio in October 2025 on terms reported to involve per generation royalties measured in tenths of a cent.
The industry has already priced this. Somebody has paid the number.
Maintain a provenance ledger. Every file hashed on ingestion, with chain of title, licence document, licensor, date, and scope recorded against the hash. When discovery arrives that document is the entire case. Build it on day one, because reconstructing it afterwards is impossible and every team discovers that on the same afternoon, usually a Thursday.
Get consent, not a waiver. Specific consents under s 195AW enumerating the acts and tied to identified works. Attribution and remuneration terms a performer would sign without a lawyer having to explain what they just gave away.
Instrument for memorisation. The Part Eight protocol, running continuously, with a similarity gate on the output path that blocks a generation before a user hears it. Stability’s filtering was expressly treated as mitigation in the Getty judgment.
Keep the model out of the argument. Munich found six songs in the parameters. London found nothing in the parameters. Deduplication, corpus scale, and regularisation decide which of those two paragraphs gets written about you.
Closing
The engineering here is good. The residual quantiser, the delay pattern, the flow matching objective, the timing conditioning that lets a model land an ending. Real work, by people who are better at this than the people selling it.
The legal position is arithmetic. Australia has a closed list of exceptions, a substantial part test that caught two bars of a nursery rhyme, a definition of material form covering storage you cannot see, and a government that has said no to a training exception twice in nine months. This audit took me an afternoon. Any competent solicitor produces the same list on a train.
There is one hard question in this entire document. It is whether your weights memorised the corpus. It is answerable with a nearest neighbour search you could run before lunch.
Most teams do not run it. They prefer the version where nobody knows.
I am comfortable with darkness. I have considerably less patience for people who choose it and then act surprised by the temperature.
I am also a musician and a studio engineer. I am not a neutral party here and I have not pretended to be.