← Blog

How to prove a song belongs to a character that has no face

2026-09-06 · OpenStela

Almost every demonstration of “character consistency” is a face. Upload a portrait, generate a new angle, measure the similarity, done. That is fine for one kind of creator and useless for most of the others. A virtual singer's output is audio. A VTuber's archive is hours of video in which the character is a small overlay, or not on screen at all. A comic character lives in posts. None of that has a face to measure, and a registry that can only vouch for faces is vouching for the smallest part of a character's body of work.

So the question is not “does this image look like the character”. It is “what evidence connects this work to this character” — and the honest answer is that it is several different kinds of evidence, of different strength, and they should be kept apart.

Start with what a fingerprint does and does not do

When a work is registered here, what goes into the nightly on-chain batch is not the hash of the file. It is the hash of a small record. Since September 2026 there are two of them, and they go in separately. The first is the authorship record: the work's own fingerprint, the name it is published under, the claimed creation date, what the registrant contributed if AI tools were involved, and the registration time. It exists whether or not the work is ever bound to a character — a song can be registered on its own. The second, written when the work is bound, is the binding record, which names both sides: the character's registry number and specification version, the fingerprint of the authorship record, who bound it, and when. Anyone can recompute either hash and walk the Merkle path to the chain. What the pair establishes is precise and narrow: this file existed under this name from this date, and was bound to this character at this version by this account on that date. It says nothing about whether the account had any business doing so. (The full walk-through, with the arithmetic, is in how a registered work is verified.)

Evidence that does not need a picture

All four checks on the evidence ladder work on a song, a faceless video or a text post exactly as well as on a portrait.

Timing. A derivative work has to come after the thing it derives from. The check compares the work's dates — the creation date you declare, the capture date in an image file, the publish date on the linked page — against the anchoring date of the character's specification version. “Registered after the character” is weak evidence on its own; “a file whose own metadata predates the character by two years” is strong evidence in the other direction, and the check says so.

Attribution. For a linked work, the platform fetches the page once at registration and freezes that snapshot. Out of it come the title, the author or channel name, the provider and the publish date as they stood that day. If the channel name matches the character's name or its holder's handle, the check passes; if the page names someone else, it fails, and it fails visibly. A snapshot cannot be edited after the fact, which is exactly why it is worth more than a live link.

Rights. The declaration the holder made when registering: own work, or licensed. Music is asked twice — once for the composition, once for the recording — because those are routinely held by different people and a buyer needs to know which one is on offer. This check records what was claimed, with a timestamp, so that a false declaration is a dated, signed statement rather than a vague memory.

Public exposure. How long the work has sat on the character's public page, attached to this character, in plain view. Time in public is not anyone's consent, but a year of open association is the kind of fact a court, or a cautious buyer, actually weighs.

The strongest evidence for a song is who published it

For most non-visual work the decisive question is not what is inside the file. It is which account is the official one. That is what channel verification is for. The holder declares a YouTube channel, an Instagram account, a Bandcamp page, a domain — and gets a one-time code to put in the bio, the channel description or any post. The platform fetches the page once, finds the code, and records the result with a date. The code contains a random part, so placing it proves control of the page; a plain link back to the character page would not, because a fan page can do that too.

Once a channel is verified, every work linked from it carries that evidence automatically. A track on the verified channel, a video on the verified channel, a post on the verified account — the buyer sees “published on a channel this holder controlled on this date” next to each one. The wording is deliberately narrow. It proves the code was there at that time, not permanent ownership of the account; that is why the verification expires after 90 days and is worth re-running before a sale.

What is deliberately not checked

Nothing here looks at the work and decides whether it "matches" the character. There is no marker detection, no palette comparison, no face similarity score. A character changes outfits, styles and framing all the time; a machine verdict on likeness would mostly be wrong, and where it was right it would only add a doubt no buyer had asked for. The platform records who said what, when, and where it was published. It does not grade the picture.

Reading the ladder as a buyer

Put together, a song on a verified channel with a clean timing check, a matching attribution snapshot, a two-part rights declaration and eight months of public exposure is well-evidenced. Not proven — nothing here proves authorship, and the platform never claims it does — but evidenced in a way that can be written into the contract by reference and checked again in a year. That is what a registry can honestly offer for the work that has no face: not a similarity score, but a stack of dated, independent, recomputable facts about where the work came from and who put it there.

How binding and evidence work →