“How to prove your song isn’t AI before a distributor pulls it.”
— the practical literature reaching working musicians, 2026
The inversion is complete.
By June 2026, at its peak, more than one new upload in two arriving at Deezer was fully AI-generated: some 90,000 tracks a day. The response has been detection, flagging, takedowns, royalty clawbacks. An artist releasing music today is presumed synthetic, and must assemble a defense: archived sessions, witness emails, absence scans against proprietary watermarks.
Disclosure is not neutral either. An AI flag now carries a price: on TIDAL, a recording identified as wholly AI-generated earns nothing and cannot be sold direct to fans. The honest declaration the system demands is also a confession that costs money.
WHAT THE PLATFORM SEES
an isolated file — undefendable
WHAT THE STUDIO KNOWS
a topology of making — evidence by construction
DETECTION
“Does this file look synthetic?”
Population-level inference. An arms race by construction — false positives against legitimate work, and a counter-industry already selling artifact removal to defeat it.
DECLARATION
“Trust me, I made it.”
Self-reporting. Unverifiable when honest, free to fake when dishonest, and structurally biased against the declarant.
ACTA
“Here is the signed trajectory of how it was made.”
Attestation of continuity, anchored to the individual baseline. No verdict on an isolated file — a work situated inside the artist’s own documented trajectory.
There is no style model anywhere in the protocol. Style is not proof.
Not a detector, which infers from a finished file. Not a label, which declares. Not a post-hoc verification service, which certifies after the fact that a named person can demonstrate a relationship to a finished master. The master is an output; its hash identifies the file. Neither contains the tracks, the gestures or the time of the session. The session is the witness, and ACTA records that session while it is being made, with a key the artist holds.
Three layers of evidence, one layer of output. Each layer is independently useful; the probative force is the bundle.
Signature at the point where sound becomes data: the analog-to-digital conversion. The interface computes rolling hashes of the incoming stream, signs them with a device key, timestamps the result. Signing at capture already exists in cameras and in some microphones; ACTA ties it to the artist’s own record of how the work was made. An attestation closer to the point of capture is a research objective, not a guarantee of the current alpha.
The production session is already the process record — it has simply never been signed. A generated track imported as a finished clip and a composition built layer by layer leave structurally different topologies. Nothing needs to be detected; the shape of the session is the evidence. Events are extracted into a canonical vocabulary, built into a Merkle tree, and the root is signed with the artist’s key.
Signed snapshots accumulate in a registry held by the artist — locally — and are anchored with RFC 3161 timestamps. The record lives in the registry, not in the released file: stripping its metadata does not erase it. A transcoded copy no longer carries the same fingerprint; linking it back to the registry takes the original export, or a link the artist declares. A single session can be fabricated; a multi-year externally anchored trajectory cannot be backdated. The longitudinal baseline is not a bonus feature — it is the security mechanism.
DDEX AI fields, AI Act transparency notices, and any future regulatory disclosure are generated automatically from the attested log. The artist never fills a checkbox again — the declaration becomes a projection of evidence rather than a confession. The label is a feature of the tooling, not a burden on the artist.
{
"acta_version": "0.1",
"events": [
{ "t": "2026-07-06T14:12:03Z", "type": "audio_capture",
"capture_attestation": "acta:l0:9f2c…" },
{ "t": "2026-07-06T14:31:44Z", "type": "midi_input", "track": 4 },
{ "t": "2026-07-06T15:02:10Z", "type": "import_external",
"flag": "unverified_import" },
{ "t": "2026-07-06T15:20:57Z", "type": "generative_render",
"tool": { "watermark": "synthid" } },
{ "t": "2026-07-06T18:05:33Z", "type": "bounce" }
],
"merkle_root": "…",
"signature": { "alg": "Ed25519", "key_id": "artist:eapr-…" },
"anchor": { "type": "rfc3161" }
}A generative model inside a session is an instrument — like a synthesizer or a mastering chain — and it is traced as one. The distinction preserved is evidentiary, not moral: augmentation under substantial creative control versus autonomous generation. That is the line on which rights now depend.
The studio should produce proof by default.
We do not ask drivers to weave their own seatbelts.
We build them into the car.
No platform.
No distributor.
No model vendor.
Signing keys and registry remain with the artist. Local first. No platform, label or vendor in the custody chain.
Merkle paths make partial proof native. You can prove a track was built layer by layer, with real takes, on a given date — and, once capture attestation exists, through an attested converter — without revealing a note, a stem, or a collaborator’s name.
LOOPBACK REPLAY
Generated audio played through an attested converter acquires a capture signature — true, and anticipated. It produces a session with no antecedent MIDI, no takes, no incremental construction, sitting inside a trajectory it must also counterfeit. No single layer is absolute proof; the bundle is the claim.
SESSION FORGERY
Faking one project file is easy. Faking a sequence of externally anchored snapshots whose timestamps precede the forgery’s motive is not. The registry raises the cost of fraud from minutes to years.
WHAT ACTA DOES NOT CLAIM
Not a detector. Not a certificate of artistic merit. Not a legal determination. ACTA produces process evidence aligned with the substantial-creative-control standard — tribunals and offices weigh it. Validation requires deployment with consenting artists, under pre-registered protocols.