Sovereign Studio
Create. Own. Prove.
NGARi Studio records, generates, edits, arranges and delivers — film, music, stills, design — on hardware you own. It decides the rights position before anything renders, signs every artefact at the moment it exists, and hands you a chain of custody that anyone can verify offline.
Why this matters now
Three things changed at once, and they all point the same way. The tools got good, the licences got conditional, and the disclosure rules arrived.
Your work is the leverage
A creator's advantage is the body of work and the record behind it. When the tools are hosted, the prompts, the takes and the history sit on infrastructure someone else operates and under terms someone else writes.
The licence decides the release
Open weights are not the same as usable weights. The most capable open image and video models are non-commercial, gated above a revenue threshold, or limited by territory — one of them is not licensed for the EU, the UK or South Korea at all.
The record is what buyers ask for
A festival, a label or a distributor increasingly wants to know how a work was made, and whether a real person's voice or likeness is inside it. European synthetic-media marking duties are already in force. “Trust me” is not a deliverable.
What Studio is
A recording studio, a generator, an editor, an arranger and a delivery desk in one application — and underneath all of it, one chain of custody.
- Recordsfrom any interface. A live meter while you play, monitoring through the interface's own outputs, and overdub against a previous take. The take is signed at the moment of capture, and the manifest names the device, the rate and the exact bytes
- Generatesmusic, video, stills and speech from open-weight models installed on your machine, under a licence that permits the use you declared
- Decides the rightsbefore anything renders: whose voice, whose likeness, whose style, on what basis — with a written rationale either way
- Refusesa cloned voice or a real person's likeness without recorded consent, and training on someone else's work without consent. It does not substitute a model behind your back
- Edits the wordstranscribe a take and every line becomes editable. Delete a sentence and the audio obeys — cross-faded, so the cut does not click. No waveform required
- Splits and arrangesa finished mix into drums, bass, other and vocals; stack those on a timeline; set gain, mute and solo; comp the best region from each of several takes
- Programmesa drum machine with a kit generated from code, swing that delays every off-beat, per-step tuning, and a piano roll for chords and basslines
- DeliversMP3, FLAC, M4A, OGG or WAV, and Standard MIDI Files that open in any DAW — with the swing encoded in the timing, so the groove survives the export
- Measuresloudness to ITU-R BS.1770, true peak, phase and stereo width, against the published targets a platform will actually normalise to
- Signs everythingeach artefact carries its own SHA-256, its sources, and an Ed25519 signature. Stems point back at the master, the master back at the take
It carries the lineage of a body of work the way a keeper of lineage does: accurately, and answerable for it. What it will not do is invent one.
One chain, and it leaves the box
Sovereign usually means the work never leaves — a wall as well as a virtue. So Studio proves the work instead, and lets it travel.
- The take is the roota performance is the one asset no model produced, and it is signed before it is stored. Nothing downstream can claim a lineage it does not have
- A transcript is a readingnot the take. It records the hash of what it read, and carries the confidence signals that separate transcription from invention — Whisper does not fail on music, it invents fluent text, and this says so
- A stem is a derivationit records the master it came from, by hash
- A master names its sourcesevery track that went into the sum, hashed
- A mixdown of your own samples proves ita beat made with your one-shots records which ones, by hash
- Verify it yourselfthe public key, the manifest and an inclusion proof are all served at a stable URL, and a single command checks the whole set offline — no account, no API, no trust in us required
The chain is recorded in an append-only Merkle log, the tree head is signed, and that head carries an RFC 3161 timestamp from an external authority. What it does not carry is an independent witness — the log's operator is also its signer, so it cannot detect its own equivocation. We name that rather than imply otherwise.
Where Studio sits
The question worth asking of any creative stack is not how good the model is. It is where the work runs, who holds it, and what you can prove afterwards.
| Where it runs | Who holds the work | What you can prove | |
|---|---|---|---|
| Cloud creative suites | Someone else's infrastructure | The vendor, under the vendor's terms | Whatever the vendor's dashboard reports |
| Hosted generative AI | Someone else's infrastructure | The vendor — including, on some tiers, the output itself | Little, and it changes with the tier |
| Self-hosted agent harnesses | Your machine | You hold the agent, and it defaults to a cloud model | Nothing about provenance, and no air-gap claim |
| NGARi Studio | Your machine | You — models, prompts, takes and record | The full thread, with the rights decision and the manifest |
The third row is the one worth pausing on. A new class of self-hosted agent harnesses arrived in 2026 and compressed the price of an agent runtime to nothing, which is genuinely useful. But they self-host the agent, not the intelligence: they default to a cloud model provider and make no air-gap, provenance or attestation claim. Studio is built for the layers they do not enter.
What Studio is not
It is not a plugin chain. The track stack sums with gain, honours mute and solo, and reports the peak it reached. There is no automation, no bus processing, no sidechain. It is a rough mix, honestly labelled — and calling it a digital audio workstation would be a claim this page cannot support.
It does not judge your music. Studio measures: loudness to ITU‑R BS.1770, true peak, crest factor, DC offset, phase correlation. It compares those to published delivery targets and tells you a master is 1.6 LU louder than the streaming target and will be turned down. What it will not do is tell you the mix is good. That is a judgement against your intent, and nothing on this machine knows what you intended. Several products claim this. None can support it.
It does not choose your best take. Comping assembles the regions you pick and cross-fades every seam, and it reports the level step it measured at each one rather than levelling the takes for you. Which take is best is a judgement about a performance. It is yours.
The generated music is not a finished master. Music runs the diffusion branch alone: the 5 Hz planner is skipped on CPU hardware, so a render is real and it is well below a release master. Every render says so in its own result. A wired invocation is not a quality result, and this page will not describe it as one.
It is not a C2PA implementation. Each manifest carries a real Ed25519 signature that anyone holding the published key can verify offline. It is not Content Credentials: there is no C2PA assertion store and no conformance claim, and no C2PA-conformant signing credential exists for this class of hardware. And it is not a copyright clearance, a licence grant or legal advice — a lawyer decides clearance.
No output here is a claim about itself. Every artefact we have produced is listed below with its hash; there is no film, no album and no campaign asset; no throughput figure is offered as a capability; and the box has never been operated by anyone but its builders.
Three things it will not do
- Nothing renders unless it is asked tothe layer builds the exact invocation and stops. Execution is a separate, explicit decision, and it is never the default
- A refused take is never renderednot even if something downstream asks. The render layer re-checks the rights decision rather than trusting its caller
- It calls no third-party APIStudio is not connected to any commercial creative platform. Every model it runs is installed on your machine, and no data leaves it
Studio's generation layer is wired for image, video, speech and music, and for restoration, capture, stems, arrangement, sequencing, delivery and measurement. The wired set is the set it will stand behind.
What has actually been run
Six verification artefacts have been produced end to end on the machine and each is recorded with its SHA-256. Nothing else exists. There is no film, no album and no campaign asset, and no throughput figure is offered as a capability.
| Artefact | What it is | SHA-256 |
|---|---|---|
| Image | 128×128, 2 steps, rendered on CPU | 0c1d4dce07d43984… |
| Speech | 4.80 s, 24 kHz mono, voice design | cfbb87851b24c98d… |
| Video | 9 frames, 128×128, Motion JPEG | 81d8bba1216d7b63… |
| Music | 10.000 s, 48 kHz stereo, diffusion branch only | 18b9ebcb9735a4b6… |
| Restored | 128×128 enlarged to 512×512 on the GPU in 4.26 s | 9e0d958506035a8f… |
| Capture | recorded from a Focusrite 2i2, signed at the source | in the log |
Video is the slowest capability here and the honest figure is minutes per clip on CPU, not seconds. The image path took tens of seconds per step. Neither is offered as a benchmark.
Who it is for
Independent filmmakers
Features and shorts where the rights position has to survive a festival submission and a distributor's questions.
Musicians and labels
Where the voice, the sample and the training corpus each carry a consent question, and the answer has to be on file.
Studios and archives
Organisations holding a catalogue they are answerable for, who need the record to outlive the tool that made it.
Bring one production
The useful test is not a demo you can run yourself. It is one real production — a short, a single, a poster series — put through the thread on your own machine, so you can see which parts of your own work stop, and what the record looks like when someone asks.
NGARi builds sovereign AI: the model, the hardware and the record, owned by the person using them. Studio is our creative-production application. Nothing on this page is a claim about a customer, a partner or a delivered result.