Canonical: https://thonly.org/research/every-face-signed-this-cut · Licence: CC0 1.0
Note. This paper discloses a mechanism that is designed and not built. It publishes the claims and withholds the build specification (§14 says what is withheld and why). A quick prior-art census ran before drafting, and §2 reports it in full, including the two conjuncts it killed, those it narrowed, and two industry specifications that a separate adversarial reader found afterwards and that narrow the record half of the claim. Its pre-registration was pushed before the first query to a private repository, so the push record exists but no reader outside it can inspect it. A full census ran the same day, its pre-registration published to a public repository before its first query, and §2.5 reports it in full.
Version 2, the same day. Invented co-stars and extras need no press: only a lent face, a real person's likeness, is signed, and the credits mark every role lent or invented (§9.3, §9.4). Drafts are listed to the actor and deleted at release (§8.6), the manifest is published with the film (§9.6), and a film's license never licenses the cast's likenesses (§10.6).
Version 3, the same day. The full census (§2.5); the sources earlier marked not re-read, re-read, and one figure corrected (§2.4); the actor's own checker as the default (§8.1); a film that survives a departure (§11); no real child's likeness (§4.2); §10 shortened.
Version 4, the same day. A remake is a new work, and a licensed screenplay never carries a face (§5.2, §10); claim 11; the first film's predictions are registered (§17).
Companion works: The Press Travels With the Message (the passkey signature over exact text, recorded in a public append-only log, and what such a signature does and does not show), B-Links (signed provenance for a shared file) and The Song That Is Not His (generation in the voice of artists who can no longer consent, the case this paper is the living counterpart of).
The commentary on the first book of the Abhidhamma tells how the teaching reached the world. For three months the Buddha taught it without a break to an assembly of gods, and the commentator stops to ask a practical question: how did a body that lives on food keep going? His answer is that the Buddha kept the time. When the hour for the alms round came, he made a mind-made double, a nimmita-buddha, and resolved: "Let this one's way of taking the robe and the bowl, its voice and its bearing, be like this; let it teach this much Dhamma." Then he took his own bowl and went for alms [1].
The commentator then asks who could tell the two apart. The mighty gods could; the lesser gods could not, "because between the Perfectly Awakened One and the created Buddha there is no difference in the rays and the rest: in rays, in voice, in words, there is no difference between the two." [1]
Two things in that passage are the whole problem this paper addresses. The double did nothing its original had not resolved: it taught this much, in this manner, for this time. And most of those watching could not tell, by looking, which one they were watching. The resolve was real and it was invisible.
A film made with a person's digital likeness has both properties and loses the first. Generated video can now put a real person's face and voice into scenes they never stood in, and a viewer can no longer tell, by looking, whether the person agreed. The person's resolve, if they gave one, lives in a permission setting inside one company's app or in a contract no viewer will see.
This paper specifies a record of that resolve that the viewer can check. Each person whose likeness appears in a film signs twice with their own passkey: once over the screenplay, when they are cast, and once over the exact final file, before it is released. The signatures go into a public append-only log. A viewer holding the film computes its fingerprint and checks, for every depicted role, that a passkey bound to that role's name pressed over this cut. And the likeness itself is stored so that it renders only under such a press. The double teaches this much, in this manner, and the lesser gods can now tell.
(The passage is cited as a recognition, tier: flourish. §13.4 runs the deletion test: nothing in the mechanism depends on it.)
This document is published as a defensive publication to establish prior art. Its contents, including every mechanism, the analysis and the enumerated claims of §15, are dedicated to the public domain under CC0 1.0 Universal. The authors assert no patent rights in anything disclosed here, will not seek patent protection for it, and will not assert any patent against any person who practices it. The dedication covers the mechanism under any name; the marks named in the Trademark Notice are reserved separately and are not licensed.
A quick prior-art census was run before drafting (§2.2), and a full census the same day, its pre-registration published before the first query (§2.5). Each result is reported in full: the declared aperture and date, every conjunct that was killed or narrowed with its prior art cited, and the survivor. The enumerated claims of §15 are that survivor and nothing wider. No live patent surfaced in either census, but patents were reached only through a general web index, so the absence is weak, and nothing here is a clearance of any kind.
Three earlier disclosures of the authors are cited as the authors' own prior art, each with its proof-backed date (§2.3): The Press Travels With the Message (the passkey press over exact text in a public append-only log; Bitcoin block 975913, 2026-11-15), B-Links (signed provenance for a shared file; block 959585, 2026-07-25) and The Song That Is Not His (block 962355, 2026-08-14). What those papers already disclose is not re-claimed here. A timestamp proves that a text existed no later than its date and nothing about authorship, originality or the validity of any claim.
This defensive publication discloses a consent mechanism for AI-generated video that uses real people's digital replicas (likenesses of face and voice). Existing practice records a performer's consent as a permission setting inside one platform, as a project-specific contract, or as a reference to terms held by a consent authority; the provenance signatures that travel with a file are applied by its makers, and an industry identity assertion lets a named actor sign a file's binding in a role such as creator or editor. None records, for a generated film, the depicted performer's own signature over the words they agreed to and then over the exact pictures released. The mechanism here does. (1) At casting, each performer signs, with their own passkey, a statement over the screenplay text, the role, the render style and the agreed shares; any change to that text is a new offer that must be signed again. (2) Before release, each performer is given the exact final file (with the list of every rendition that will be served) and signs a statement over its cryptographic fingerprint; a single missing signature keeps that cut from release. Both signatures are recorded in a public append-only log as fingerprints salted per party, so the log alone pairs no records and names no one; the film carries the statements in a release manifest, so that anyone holding the released bytes can recompute their fingerprint and check that, for every lent role, a passkey listed on that role's public name page signed the screenplay and then this exact file. Invented co-stars and extras belong to no one, need no press, and are credited as invented. (3) The stored likeness model is encrypted under a key that only the performer's authenticator can derive, released for one work at a time, so that a party holding only the stored ciphertext cannot render it. A census found the parts public and did not find the composition. The paper claims that composition and nothing wider, and states what it does not do: a check covers released bytes, never a re-encoded copy; it names lent roles, never every face in the frames, and never who the faces belong to; it shows consent as given at release, never consent now; a signature shows a passkey pressed, never that a particular person did; and the operator who runs the press page and the render remains trusted, because one frame of a released film is enough for current tools to render a person with no stored model at all.
There are three kinds of face a film can show. A performed face belongs to someone who stood in front of a camera. An invented face belongs to no one: a generator made it, and any other invented face would have done as well. A lent face belongs to a real person who did not stand in front of any camera for this film, and whose likeness a model rendered into it.
The order is not historical. Performed faces are not the past and lent faces are not the future; all three will be made side by side. One film may hold more than one kind: a lent lead beside invented co-stars and an invented crowd. What separates them is who stands behind the face, and what a viewer can know about it. For a performed face, the viewer knows that a person was there. For an invented face, the viewer knows that no one was. For a lent face, the viewer knows nothing at all, because the picture of a person who agreed and the picture of a person who did not are the same picture.
The studio this mechanism is designed for is THonly.com. The name is pronounced the only, and in the institution's usage it is a grammar given to everyone: each person is the only one of themselves, and everyone is that equally. Read that way, the name rules something out. An invented face is interchangeable by construction. A lent face is the only one of its kind: it can stand in for exactly one person, and, under this design, only by that person's own press. Every lent face signed this cut is the studio's name made checkable.
The studio makes films as gifts. Films are released free to the public at a set date; viewers may thank the cast afterwards; the studio charges for the compute that renders a film and takes no share of what a film earns. That frame (§10) is context for the mechanism, not part of the claims.
Consent to the use of one's likeness exists today in several forms, and none of them gives a viewer the depicted performer's own signature over the released pictures.
WHERE CONSENT LIVES TODAY WHAT A VIEWER CAN CHECK
──────────────────────────────────────────── ──────────────────────────────────
a permission setting inside one app nothing (the app's word)
a project contract between the parties nothing (the contract is private)
a provenance signature on the file that a STUDIO or TOOL made it
(applied by the maker)
an identity assertion by a named actor that a named actor's credential
(an industry standard, 2025) signed the file, in a role such
as creator or editor
a consent reference (an industry draft, 2026) that terms exist at a consent
authority; the signer does not
declare the consent itself
──────────────────────────────────────────── ──────────────────────────────────
THIS PAPER: each lent performer's own that, for every lent role, a
signature over the screenplay, then over the passkey bound to that role's name
exact released file, in a public log pressed over the words and then
over THIS file
A permission setting answers the question may this platform render me? It does not answer did I agree to this scene, in this film, as released?, and it answers nothing for a viewer outside the platform. A contract answers the second question for the parties and for a court, never for a viewer. A maker's provenance signature answers who made this file?, which is the wrong person: a studio can sign, truthfully, a deepfake of someone who never agreed (§2.2, row (g)). The two industry specifications come closest and are cited in §2.2: one lets a named person sign a file's binding, the other points from a file to consent held elsewhere. Neither asks the depicted performer to sign the words before the render and the bytes after it.
A person who agrees to appear in a generated film agrees, at best, to a screenplay. What is released is not a screenplay. It is pictures a model made, and a model invents: a gesture the script did not call for, a costume, a crowd, an expression, a line read with a meaning the words did not carry. Consent to a script is not consent to a render. A model weighing whether to license his face in today's market put it plainly to a reporter: "If I license my face, I have no idea how AI might eventually change it or whether it could portray me in degrading ways" [5]. The design therefore asks for consent twice, and the second time over the exact bytes that will be released, not over a description of them.
The design does not promise anyone a living. A lent face does no acting; the model does. A face can appear in any number of films at once, and invented faces cost nothing, so a stranger's likeness is worth little in a market. The census found one report of what such a market pays: two face-licensing marketplaces that supply AI-generated dramas pay between 15 and 700 US dollars to license a likeness, and one charges 15 to 74 dollars an episode (§2.4). The design is for people who want to make films with the people they love, and it lets whatever a film earns be shared among the cast on terms fixed before it is made. Nor does it promise to protect anyone's face: it records consent, and it does not prevent deepfakes (§7.4); the studio's own copy never says otherwise.
Three bodies of practice surround this design.
Census record: depth quick, paper every-face-signed-this-cut.
Pre-registration: pushed on 2026-11-19 at 02:40 Pacific time, before the first query, to a private repository; the push record is the repository host's, and it cannot be inspected from outside. Run date: 2026-11-19.
Declared aperture (fixed before searching): the commercial web (AI video apps, likeness-licensing marketplaces, avatar and voice platforms, vendor documentation, press, the 2023 SAG-AFTRA digital-replica terms); non-commercial practice (community and amateur film, fan films, documentary and theater consent practice); a keyword search of the Google Patents index including published applications, through a general web index. English only. Not run: Chinese, Japanese and Korean (where a market in licensed digital humans is live, so the full census must search it natively), academic indexes, standards (the provenance-signature mechanism is inherited from an earlier census of the authors and not re-searched), law reviews, and Khmer. Controls: a known product for the web (the permissioned-likeness feature of row (a)) and a known patent for the patent index (US10853717B2). Both were seen.
Conjuncts, in kill-first order, with predictions registered before any query.
(The letters are as registered; the order is the kill-first order of the predictions.)
| Row | Predicted | Found | Their first public date | Ours, earliest (proof-backed) | Verdict |
|---|---|---|---|---|---|
| Control (web): a permissioned likeness feature in a video app | found | seen (row (a)) | — | none | Seen. |
| Control (patents): US10853717B2 | found | seen, by a site-restricted web query (not through the patent index's own search) | — | none | Seen. The NOT FOUND rows below stand as weak nulls |
| (a) verified, permissioned likeness | killed | Sora "cameos" (OpenAI, launched 2025-09-30), as reported: "You record a short video following on-screen instructions to verify your identity and capture your likeness"; "you decide who can use your cameo (just friends or everyone)"; "OpenAI can't generate anyone from a prompt or photo unless they've submitted a verified cameo first." [3] | 2025-09-30 (press) | none | KILLED. Not claimed |
| (b) the likeness owner is paid | killed or narrowed | ElevenLabs voice payouts: "Earnings are calculated based on usage by paid users", the rate set by the owner's chosen notice period [4]. ActID and New Claw (face-licensing for AI dramas): ActID's "Prices range from 99 to 500 yuan ($15 to $74) per episode, with ActID taking a 10% commission"; New Claw "has imposed a minimum price of 500 yuan ($74) per image" [5] | undated; 2026-07-27 (press) | none | NARROWED. Payment to a likeness owner per use or per episode is public. A share of what the work earns was not found (weak) |
| (d) terms fixed in a signed agreement before generation | killed at mechanism | Respeecher: "Without a signed, project‑specific agreement, we simply don't start any digital replica work"; terms to "Spell out pay for AI voice work clearly — session fees, reuse, and any extra rates" [6]; fixed splits agreed before production are ordinary in film and music | 2026-08-13 | none | KILLED at mechanism. Not claimed |
| (c) an open call for one work | narrowed | ActID and New Claw are catalogs of faces that creators browse and buy [5] | 2026-07-27 | none | CONTRAST. A browsable catalog, the opposite of a call. An author's call answered by likeness holders for one work: not found (weak) |
| (h) thanks split by signed shares, no platform cut | narrowed | no query was aimed at it | — | — | NOT RUN. Not claimed |
| (e) two consents: script, then exact final file | narrowed | Respeecher: a signed project agreement before work, and advice to "give the actor or their representative the right to listen to AI‑generated lines before final approval" [6]. Sora: the owner can "see all videos featuring your likeness including drafts" and "delete any video or revoke access anytime" [3] | 2025-09-30; 2026-08-13 | the passkey press over exact text, block 975913 (§2.3) | NARROWED. Consent before work, review of output and an after-the-fact veto are public. Two cryptographic signatures by the depicted person, the second over the exact final bytes: not found (weak) |
| (g) each depicted person's signature, checkable by viewers | narrowed | Truepic and Revel.ai: a generated likeness video "signed as an AI generation with Truepic's C2PA transparency tools and the creators, Revel.ai" [7]; the depicted person signs nothing. Found after the census by the adversarial reader (§13.1), re-read: the CAWG Identity Assertion 1.1 ("DIF Ratified Specification 16 May 2025") may be "repeated any number of times to convey information specific to distinct named actors", a signature "reflective of the named actor's authorization of or active participation", over a list that "MUST include a hard binding assertion"; its recommended roles are creator, contributor, editor, producer, publisher, sponsor and translator, and other values may be used [13]. The CAWG Consent Assertion draft ("Version 1.0 Draft 12 July 2026") "carries a signed, machine-readable reference from a digital asset to an authoritative source of consent" for a "subject" or "performer", identifying the asset by its hard binding, and "The signer does not declare the consent itself" [14] | 2025-05-16; 2026-07-12 | the passkey-signed file record in a public log, block 959585 and block 975913 (§2.3) | NARROWED. A named person signing a file's binding is public, and so is a file pointing to a performer's consent held elsewhere. The depicted performer's own signature over the screenplay before the render and over the exact final bytes after it, in a public append-only log, checkable by a viewer: not found (weak) |
| (f) likeness unusable without a per-work signature | not found | one query, no candidate. The nearest public forms are a policy gate (row (a): the platform "can't generate anyone … unless") and, found by the adversarial reader and not re-read, the confidential-computing practice of releasing a model's key only after hardware attestation | — | none | NOT FOUND as stated (weak, one query); key release gated on attestation is a public practice and narrows Variant A. The adversarial reader also showed that the guard protects less than its wording claimed (§7.4) |
| Composition: all conjuncts in one system | not found | none | — | — | NOT FOUND (weak: quick, English only, patents through a web index) |
Mechanism width and composition width. At mechanism width the design is commons: a verified, permissioned likeness in a video app; payment to the likeness owner; a signed project-specific agreement before replica work; review of output and an after-the-fact veto; a provenance signature on a media file; a named person's identity assertion over a file's binding; a reference from a file to a performer's consent held by an authority. At composition width: not found, in the aperture above, on 2026-11-19. Not found means not found in this aperture on that date; it never means new.
Predictions that missed. Row (c) was predicted to narrow and found only catalogs, the opposite design, so an open call is closer to unclaimed than predicted; it is still not claimed apart from the composition. Row (g) was predicted to narrow; the quick census found only the opposite assignment of the signature (the maker signs) and reported it not found, and the adversarial reader then found the two industry specifications that do narrow it. The census missed them because they live in a standards body's own pages, a neighborhood the quick depth did not search. Row (h) was never searched: a coverage miss, not a finding. A Chinese marketplace surfaced through English press although Chinese was not run, which says the market is live and must be searched natively.
The survivor, in one sentence, as narrowed after the adversarial pass. Each credited performer presses their own passkey twice, over the screenplay and then over the exact final file and its listed renditions; the presses are recorded with per-party salts in a public append-only log and their statements travel with the film, so that anyone holding the released bytes can check both presses for every lent role; and a party holding only the stored likeness cannot render it without such a press. §15 claims that, and nothing wider.
The Press Travels With the Message [8] already discloses a passkey signature (a WebAuthn assertion with user verification) over exact text, recorded in a public append-only log, carried with the message and checked in the reader's own browser, together with the limits that bind this paper too: a press shows that a passkey pressed, never that a person did, and a name binding adds accountability to a name, never identity (block 975913, 2026-11-15). B-Links [9] already discloses a signed provenance record for a shared file (block 959585, 2026-07-25). The authors' open specification of signed statements [10] reserves, since 2026-11-13, an agreement pressed once by each of two parties over the same text with salted fingerprints, and, since 2026-11-19, a seat holder's own record of a film's premiere. The Song That Is Not His [11] is the opposite case: generation in the voice of artists murdered with no estate and no record of intent, where consent is not obtainable, and not merely unobtained (block 962355, 2026-08-14). None of these discloses a depicted person's signature over a film's screenplay and its final bytes, or a likeness that renders only under such a signature.
Census record: depth full, paper every-face-signed-this-cut.
Pre-registration: published to a public repository (thonly/publications, commit 0003d63, timestamps/census/b-double-full-prereg.md) on 2026-11-19 at 10:43 Pacific time, before the first query, and stamped with OpenTimestamps; the push record is the repository host's. Run date: 2026-11-19.
What it covered. Only what the claims rest on. Conjuncts killed by the quick census were cited, not re-searched. Aperture (fixed before searching): the commercial and non-commercial web; Google Patents by full text with a walk from the nearest hit, through a general web index; academic indexes through a general web index; standards bodies (C2PA, CAWG, IETF, W3C); Chinese, Japanese and Korean. Not run: Khmer, native patent offices, law reviews beyond what search surfaced, and any reading of a patent's claims. One control per source group, named before the push: web (the permissioned-likeness product), patents (US10853717B2), academic (Carlini et al. [15]), non-English (China's Deep Synthesis Provisions [22]), all seen. The standards group missed its control twice (its index reached the CAWG specification only through secondary documentation), so every absence that group reported is void; the CAWG Identity Assertion 1.1 was instead read in full by the authors, and it has no concept of a depicted person or performer, no public log and no guidance on unlinkability [13].
| Row | Predicted | Found (re-read at source) | Their first public date | Verdict |
|---|---|---|---|---|
| (l) before the release press, the actor's device hashes the cut with a checker the operator did not write, and the press proceeds only if it agrees | narrowed | nothing for a film cut; the general practice of checking a hash on a second device was not searched | — | NOT FOUND (weak) |
| (e′) two signatures by each depicted performer: the screenplay before generation, the exact final file before release, gating release | narrowed | A studio making AI-customized dramas lets performers take part "从剧本预览到数字形象细节确认" (from script preview to confirming the digital image's details), with a final sign-off by the talent's team, under a signed authorization contract; nothing cryptographic [17]. A 2026 proposal has a performer "issue a cryptographic authorization" by scope, term, territories, permitted transformations, revocation and payout, "signed by a wallet the performer controls" and "timestamped on-chain", and anchors a master file's hash on-chain; the performer does not sign the master, and the authorization names no single work [18] | 2026-03-17; 2026-03-26 | NARROWED. A performer-signed authorization on a public ledger, and a studio's review from script to final image, are public. Two performer signatures, over the screenplay before generation and over the exact final bytes before release, gating release: not found (weak) |
| (g′) the signatures in a public append-only log as per-party salted, unpairable records, the statements traveling in the film's manifest | narrowed | Salting log entries for privacy is proposed public practice: "This can be mitigated by salting the hash, though there's no built-in support for that" [21]; the on-chain anchoring of [18]; the CAWG assertions of §2.2 [13][14] | 2022-05-28 | NARROWED. Per-party salted, unpairable records of performers' signatures, with the manifest traveling with the film: not found (weak; the standards group's absences are void) |
| (i) every role credited lent or invented; an invented face rejected at render if it matches an enrolled likeness | narrowed | A video platform's likeness detection: a person enrolls with "a government-issued ID and a brief video of your face", and the system runs "a one-time search of newly uploaded videos", after upload, never at generation [19]. China requires that the person whose face or voice is edited be informed and give "单独同意" (separate consent), effective 2023-01-10 [22] | 2023-01-10 | NARROWED. Matching against enrolled faces, and separate consent to editing a face, are public. Crediting every role lent or invented, with a block at render: not found (weak) |
| (f) a likeness model decryptable only with a per-work key derived by the performer's authenticator | narrowed | Passkey key derivation for encryption is published practice: the PRF output treated as "Input Keying Material", a data key "wrapped" by a key-encryption key, and "a new, unrelated key from the same masterKey by simply changing the info string" [20] | undated | KILLED at mechanism. Applied to a person's likeness model and bound to a casting statement: not found (weak) |
| (j) an actor's signed exit ends distribution, and a recut without that role is released on the remaining cast's new signatures | not found | nothing | — | NOT FOUND (weak) |
| Composition | not found | none | — | NOT FOUND (weak) |
Leads the scouts returned that were not re-read, and are therefore not relied on: a render service that admits only faces a user owns; a consent-registry pipeline that blocks generation without consent; a university prototype that checks a likeness's permitted scope before generating; two likeness registries, one zero-knowledge and one for theater actors; an IETF draft for a Merkle log beside content provenance; and three patents on device-signed frames for the authenticity of captured video, which address capture, not consent.
The survivor, in one sentence. Each depicted performer of a generated film presses their own passkey over the screenplay before generation and over the exact final file's fingerprint, computed also by a checker the operator did not write, before release; a missing press keeps the cut from release; and the presses sit in a public append-only log as per-party salted fingerprints, with the statements traveling in the film's manifest. The claims of §15 rest on it. Element (c) of claim 1 is a known technique applied to a new object, and the key derivation of claim 2 is cited as public.
Predictions that missed. (f) was predicted to narrow and was killed at mechanism. (e′) met a performer-signed public authorization closer than predicted. The general practice behind (l) was assumed and never searched. The standards group could not see its own control. Not found means not found in this aperture on that date; it never means new.
| Role | Who | Holds |
|---|---|---|
| Holder | a person who has enrolled their own likeness | a passkey; a public name page (@name) listing that passkey; their likeness model, encrypted |
| Author | a person making a film | the screenplay, the casting call, the render budget |
| Actor | a holder cast in a lent role (the depicted person, the performer) | their two presses for that film |
| Studio | the operator of the render service (here, THonly.com) | the encrypted likeness models at rest; compute; the release |
| Log | a public append-only log of signed statements | salted fingerprints of every press, in order |
| Viewer | anyone holding the film | the file, its release manifest, and a checker |
HOLDER AUTHOR STUDIO LOG VIEWER
│ │ │ │ │
① │ live capture ──────────────────────────────►│ likeness model, │ │
│ (tied to passkey) │ │ ENCRYPTED; key only │ │
│ │ │ the holder can open │ │
② │ │ casting call ────────►│ (no index of faces) │ │
│◄────────────────────│ offer: screenplay, │ │ │
│ │ role, style, shares │ │ │
③ │ PRESS 1 over the offer text ───────────────────────────────────►│ salted │
│ (any edit = new offer, new press) │ │ fingerprint │
④ │ releases the work key, bound to PRESS 1 ──►│ RENDER (this work │ │
│ │ │ only) │ │
⑤ │◄──────────────────────── exact final file ──│ │ │
│ fingerprint computed ON THE HOLDER'S DEVICE │ │ │
⑥ │ PRESS 2 over that fingerprint ─────────────────────────────────►│ salted │
│ (no PRESS 2 from any actor → this cut │ │ fingerprint │
│ is not released) │ │ │
⑦ │ │ │ RELEASE: the file + │ │
│ │ │ a MANIFEST of every │ │
│ │ │ statement text ─────────────────────►│
⑧ │ │ │ │◄── check ─────│
│ │ │ │ recompute │
│ │ │ │ fingerprint; │
│ │ │ │ each role: │
│ │ │ │ PRESS 2 over │
│ │ │ │ THIS file by │
│ │ │ │ a passkey │
│ │ │ │ its @name │
│ │ │ │ page lists │
@name page lists, over the fingerprint of the bytes in hand (the master file or a listed rendition); and a casting statement by the same passkey over the screenplay offer that the release statement refers to, earlier in the log's order.A likeness is enrolled by its holder and by no one else. The holder makes a live capture (a short guided video with face and voice) on their own device, in the same session in which their passkey presses an enrollment statement. The capture is never accepted as an upload of existing photographs or video, because an upload can be of anyone. The enrollment statement binds the likeness model's fingerprint to the holder's passkey, and the holder's public @name page lists that passkey.
What enrollment binds, and what it does not. It binds a passkey to a capture. It does not bind the capture to the face of the person holding the passkey. Feeding a swapped face into the camera stream, through a virtual camera, is a known and growing attack on remote liveness checks, so the live-capture requirement is a rule kept by the capture app, not a property. A person could enroll someone else's face, and the record would then truthfully show that their passkey pressed over a film showing that other person. This is why the check reports credited names and never claims to know whose faces are on screen.
Enrollment does not license anything. It creates the encrypted likeness and nothing more. No author can see it, search it or browse it (§5). It confers no rank, no counter of films and no measure of demand.
A likeness can be rendered lifelike, as a cartoon, as a three-dimensional figure or in any other style. Consent attaches to whether the person can be recognized, never to the style. A cartoon of a person is still that person. Every style therefore takes the same two presses, and the casting statement names the style, so that a person who agreed to be drawn is never released as a photograph.
An author posts a call for a role in one work: the role, its share of the screenplay, the render style, the proposed shares, the release date. Holders who see the call (because the author shared it with them, or because they chose to read open calls) may answer it. There is no catalog of faces. No author can browse enrolled likenesses, search them by appearance or sort them by popularity. A face enters a film only in answer to a call, by its holder's own act.
The design question this answers is is there an index? A catalog of faces is an index of people sorted by what buyers want, and it turns the people in it into stock. The census found exactly that design in the market (§2.2, row (c)).
A likeness is lent to one work. The lend is not transferable: an author cannot assign a cast member's consent to another author or another film, and cannot accumulate rights to a face. A second film needs a second call and two new presses. So does a remake. When an author remakes a screenplay with a new cast, or someone licensed to do so makes a sequel or a prequel, the result is a new work: no casting statement carries over, an actor from the earlier film appears in it only by pressing again, and the earlier film stays exactly as released, with its record. A remake's page never compares the two casts and never names anyone who was replaced.
The casting statement is one exact text, written by the author and shown to the actor in full. It carries:
The actor presses it with their passkey. The log records a salted fingerprint of the statement; the text travels with the parties and later with the film's manifest.
Any change to the text, to the screenplay, the role, the style or the shares, is new bytes, and new bytes are a new offer. The actor's earlier press does not carry over to it. This is a property of signing exact text, not a rule anyone enforces: a press over one text verifies against that text and no other.
Until the work key is released for rendering (§7), the actor may withdraw their press by a second statement. Which came first, a withdrawal or the release of the key, is decided by the log's order, never by the minute recorded in either.
The aim is that the studio holding a person's likeness cannot render it without that person's press, and that this be true because of how the likeness is stored, not because the studio promises it.
The likeness model is stored encrypted. The key that decrypts it is wrapped so that only the holder's own authenticator can unwrap it: a passkey can derive a secret from a value the relying party supplies (the WebAuthn pseudo-random function extension does exactly this), and the derivation is available only during a press on the holder's device, its output returned to the script of the page that requested the press. At the casting press, the holder's device derives a work key bound to the casting statement's fingerprint and an expiry, and releases it to the render service for that work alone. Without a press there is no work key, and the stored model is ciphertext.
AT REST AT A CASTING PRESS AFTER THE RENDER
───────────────────────────── ───────────────────────────────── ───────────────────────────
likeness model = ciphertext holder's device + passkey → work key expires;
studio holds no key that work key, bound to the model returns to
opens it H(casting statement) + expiry ciphertext at rest
→ render service, this work only
PROPERTY (vs the studio ───────────────────────────────── RULE: that the service kept
at rest) the service holds PLAINTEXT no copy while it had one
while it renders (§7.4)
The guard was first stated as making the likeness unusable without a press. The adversarial pass (§13.1) showed that against the operator it is trust, not a property, at four points, and that the stored model matters less than the wording suggested.
When the author has a cut they intend to release, every actor in it is sent the exact file to be released. Two programs compute its fingerprint (a cryptographic hash of the bytes) from the bytes delivered to the actor's device: the press page, and a checker the actor chose that the operator did not write, such as an open-source tool or one the actor installed themselves. The press proceeds only if the two agree. That makes the guard a property: a studio that showed one cut and released another would release a file whose fingerprint matches no actor's press, and a press page that reported a false fingerprint would be contradicted by the actor's own checker before anything was signed. An actor who uses only the press page is trusting its operator, and the record says which way the fingerprint was computed (Variant G, §12.7).
The release statement names the work, the role, the casting statement it rests on (by fingerprint), the fingerprint of the final master file, and the list of fingerprints of every rendition the studio will serve: each quality level, and for streaming the per-segment fingerprints, produced before the press. The actor's device computes the master's fingerprint; the rendition list is the operator's report, and the actor presses over it as reported. The log records a salted fingerprint.
A cut is released only when every lent role in it has a release press over that cut's fingerprint. Invented roles need none. If any actor does not press, by declining or by not answering, that cut does not clear. The author may make another cut, without that role or with that role changed, and a new cut is new bytes: every remaining actor presses again, because consent was to the old file, not the new one. A recut that only removed one scene still changes the bytes.
A description of a cut can be true and still misleading: the scene where you leave the room says nothing about how the model drew the leaving. A fingerprint describes nothing and leaves nothing out. The actor agrees to exactly the file they watched, and the viewer can check exactly that.
The release press comes after the render, so it cannot be what permits the render; the casting press is. Between the two, an author may render many drafts, retakes and alternate cuts showing the actor, and the model may invent in each of them. The design therefore lists every render that uses an actor's likeness, draft or final, to that actor and to no one else, as the permissioned-likeness product found in the census does for its owners (§2.2, row (e)), and deletes the drafts when the film is released or when the casting key expires, whichever comes first. Both are rules kept by the studio. Variant F (§12.6) states the alternatives, including a fresh key per render, which would make the listing a property at the cost of a press per render.
The released film carries, beside it, a manifest: every casting statement and every release statement, in full, with their salts. The log holds only fingerprints, each salted with a salt of its own party's, so no two records in the log can be paired, even when every actor pressed a release statement over the same file. The log alone therefore says nothing about who appeared in which film or how large a cast was; a person holding the film and its manifest can check everything.
PUBLIC LOG (anyone) MANIFEST (travels with the film)
────────────────────────────────── ────────────────────────────────────────
record 4,812,301: salted fp, time casting statement, role "Mother":
record 4,812,977: salted fp, time work id · screenplay fp · style ·
record 4,815,040: salted fp, time shares · terms · salt
... release statement, role "Mother":
(no names, no titles, no files) work id · casting fp · FILE fp · salt
... one pair per role ...
────────────────────────────────── ────────────────────────────────────────
CHECK = hash(salt ‖ statement) is in the log, pressed by a passkey that the
role's @name page lists, and the FILE fp equals hash(the file I hold)
@name page lists.The checker never reports verified, never reports consented, never reports who held a passkey and never reports that a person was present at a device. Those are not what a press shows [8][10].
A film may mix lent faces with invented ones: the author's own likeness and a friend's in the lead roles, invented co-stars, an invented crowd. Only a lent face belongs to someone, so only a lent face is signed. An invented face belongs to no one and needs no press; requiring one would make an ordinary film impossible to make, and would protect no one. The credits mark every role either lent or invented, so a viewer knows which faces the check speaks for.
The line is therefore a compression. Its exact form is: for every lent role in the manifest, a passkey listed on that role's @name page pressed over the screenplay offer and then over the fingerprint of this exact file. The design guarantees the record. It does not guarantee the person behind the passkey, and the paper uses the short line only where this sentence stands behind it.
Every figure that is not a lent role (a named co-star, a crowd, a passerby, a face in a photograph on a wall) is invented: generated, belonging to no one, and credited as invented. Two rules keep an invented face from being a real one. At render time each invented face is matched against the enrolled likenesses of people who agreed to that check at enrollment, and one that matches an enrolled likeness above a threshold is rejected, which keeps an enrolled holder out of a film they did not press for. And an invented role is never described in the screenplay as a real person. Against a person who never enrolled, no check is possible: a model can invent a face that happens to resemble a real person, and image models have been shown to reproduce individual training images, photographs of identifiable people among them [15]. The record can say who pressed; it cannot prove the negative that no unconsenting person appears. That is why the check reports lent roles, never every face in the frames.
A repost on another platform is almost always re-encoded, and a re-encoded file has different bytes, so its fingerprint matches no release press. The studio's own renditions are different bytes from the master too, which is why the release press covers the rendition list (§8.2); a viewer who downloads a listed rendition can check it. A viewer who streams cannot hash what they watch unless their player checks each segment against the pressed segment fingerprints (per-segment hashing of streamed media is public practice in the industry provenance standard, and is not claimed); otherwise, for a stream, the check is the page's word. And a check is only as good as the checker: a viewer who uses the checker the film's own page offers is trusting that page [8]. The check is exact on listed bytes, through a checker the viewer chose, and silent on everything else. Variant C (§12.3) finds the record again for an unlisted copy by perceptual matching or a watermark, and reports it as a likely match to a released film, never as a check passed.
An actor's participation is public to anyone who holds the film and its manifest, in the way screen credits are. It is not public to someone who only reads the log: per-party salted fingerprints cannot be paired with each other, or with a name or a title, without the manifest. A holder who does not want participation discoverable at all should not answer the call; the design offers no anonymous role, because an anonymous face is not checkable.
The index others can build. Checking a lent role by name requires that the manifest show which passkey, listed on which @name page, pressed. So anyone who crawls released films and their manifests can list every film a person appeared in, exactly as anyone can compile screen credits today. The institution builds no such list and never shows a person's films on their own name page; it cannot stop others from compiling one from public manifests. The default is that the manifest is published with the film, because that is what lets anyone holding a download check it, and because a credit on the film is a credit on the subject, never on the person. The casting statement says so once: the credit is public with the film. Variant D (§12.4) states the alternatives, including a manifest given only to those who play the film, and what each costs.
The census killed or did not search these parts (§2.2, rows (b), (d), (h)), and they are not claimed. They are stated briefly because the mechanism was designed with them.
A film released free at the studio's venue is licensed by its author under CC BY-NC-ND 4.0 by default: anyone may watch it and share it unchanged, with credit, not for commercial use, and not as a derivative work. The license covers the film's copyright and nothing else. The Creative Commons 4.0 licenses state that they do not license publicity, privacy or other personality rights, so a film's license never licenses the cast's likenesses; a recut, a remix or a sequel that shows a lent face needs that person's presses, as a new work. A recut made by someone else is new bytes and fails the check, correctly. This paper itself is CC0; the films the mechanism produces are not.
An actor may leave a released film by a press. Their share ends with it, and the film survives them.
The record does not change when an actor leaves. The log is append-only, so the earlier release presses stay in it and keep verifying for the earlier cut, which is no longer served. A leaving press is a separate statement; it is not a withdrawal of a release press, and checkers never report a film or a press as withdrawn, because a mark that follows every copy would announce one person's departure to every viewer forever. A check shows consent as given at release, never consent now.
The exit is not waivable. An author cannot ask an actor to sign away the right to leave: a lent face in a film its owner wants out of is the harm the design exists to prevent. What the design makes cheap is not the exit but the film's survival of it.
This replaces the after-the-fact delete found in the market (§2.4) with a press, and differs from it in one way: the delete there makes the owner a co-owner of each video inside one platform; the press here ends one operator's distribution of the cuts that show that face, and the film goes on without it.
(i) The work key is released to the studio's render service, which holds plaintext during the render (§7.2; the default). (ii) The render runs in attested confidential computing: the holder's device checks the enclave's attestation before releasing the work key, and the enclave releases only the rendered frames. (iii) The render runs on the holder's own device or one they designate, and only frames leave it. (ii) and (iii) narrow the rule of §7.4; (iii) is limited by the device.
(i) One release press per role over the whole file (§8; the default). (ii) One press per scene the role appears in, over each scene's fingerprint, with a final press over the assembled file. (ii) lets an actor approve most of a film and decline one scene, at the cost of more presses, and makes a recut after a departure cheap for everyone whose scenes did not change (§11).
(i) None: the check is exact on the released file only (the default). (ii) A perceptual fingerprint of the released film is published beside the manifest, and a copy is matched against it. (iii) An invisible watermark carries the work's identifier. (ii) and (iii) report a likely match, never a passed check.
(i) Per-party salted fingerprints in the log, statements in a manifest published with the film, the casting statement saying so (§9.6; the default). (ii) The manifest given only to those who play the film, by the player, and not published beside it, so that a crawler of public pages cannot compile who appeared in what; a viewer who holds only a downloaded file then cannot check. (iii) Release statements published unsalted, so the cast of any film is discoverable from the log alone. (iv) Casting statements kept off the manifest, so a viewer checks release presses only. (ii) trades checkability for privacy; (iii) trades the actors' privacy for discoverability; (iv) loses the check that the cut rests on a casting the actor pressed. A per-film key for each actor would make films unlinkable to each other and would break the binding to the actor's name, so it is not offered.
(i) The casting statement names one style (§4.4; the default). (ii) It names a set of permitted styles. (iii) It permits any style. (iii) is permitted but weakens the actor's knowledge of what they agreed to; the second press still binds the actual file.
(i) Nothing governs drafts. (ii) Every render that uses an actor's likeness, draft or final, is listed to that actor and to no one else. (iii) As (ii), and drafts are deleted when the film is released or when the casting key expires, whichever comes first (§8.6; the default). (iv) Each render needs a fresh, short-lived key from the actor's authenticator, so the actor's device sees every render it permits, at the cost of a press per render.
(i) The press page alone (trust in its operator). (ii) A checker the actor chose, one the operator did not write, alone (a property). (iii) Both, and the press proceeds only if they agree (§8.1; the default).
The thesis was attacked twice before this paper was pushed. The first pass was the author's own, during the paper-worthiness check. The second was a separate adversarial reader with no write access, so that the author did not refute, or rescue, its own claim. Neither pass killed the composition. Together they narrowed it in these ways, each now in the text.
| Attack | What answers it | Strength |
|---|---|---|
| The studio shows the actor one cut and releases another | the actor's own checker and the press page must agree on the fingerprint of the bytes played (§8.1) | property by default |
| The author edits the screenplay after the actor pressed | an edit is new bytes and a new offer (§6.2) | property |
| The author reuses a casting for a second film | the work key is bound to one casting statement (§7.3) | property |
| A breach of the likeness store | the stored likeness is ciphertext with no key (§7.3) | property |
| The operator keeps the enrollment capture, the derived key or a render | not prevented; Variant A narrows the render (§7.4) | trust |
| Anyone renders the actor from one released frame | not answered; no stored model is needed (§7.4) | none |
| Someone enrolls another person's face by injecting a swapped camera stream | the capture app's liveness check (§4.1) | rule, and a weak one |
| A holder's passkey is shared, synced, stolen or held by software | not answered; a press shows a passkey (§9.3) | none |
| A generated extra resembles an enrolled holder | rejected at render by a match against enrolled likenesses that agreed to the check (§9.4) | rule |
| A generated extra resembles, or reproduces, a person who never enrolled | not answerable by this design (§16) | none |
| An actor leaves and the whole film is lost | the author recuts without that role; the remaining cast presses again (§11) | rule-free: a new cut is new bytes |
| Drafts showing the actor are made and kept | every render is listed to the actor; drafts are deleted at release (§8.6) | rule |
| Pressure on an actor to approve, or not to leave | no reminders, no countdown, no visible waiting list, no rendered absence (§8.4) | rule; social pressure is not answered |
| A repost or a stream is presented as checked | the check fails on unlisted bytes; a likely match is never a pass (§9.5) | property of the check |
| A viewer is shown a fake checker page | a check is as good as the checker the viewer chose [8] | none by design |
| Someone compiles every film a person appeared in | per-party salts keep the log unpairable; public manifests do not (§9.6, Variant D) | partial |
A guard is a property if it holds when no one is enforcing it, and a rule if it needs someone to keep it at the moment it is tested.
| Guard | Property or rule | Why |
|---|---|---|
| A changed screenplay needs a new press | property | a signature verifies against one exact text |
| The released file is the one the actor approved | property by default | the actor's own checker and the press page must agree (§8.1); with the press page alone, trust |
| A party holding only the stored ciphertext cannot render it | property | no key exists without the performer's authenticator |
| The operator cannot render without a press | trust, not a property | the operator sees the capture, the key and the render (§7.4) |
| A work key opens one work only | property | the key is derived from one statement's fingerprint |
| The log does not pair records or reveal who appeared in what | property | per-party salts; nothing pairs without the manifest |
| No catalog of faces | rule | nothing in a database prevents an index; the studio must not build one |
| Enrollment only by a live capture of the holder | rule, weak | a swapped camera stream defeats liveness checks (§4.1) |
| No minors | rule | the form excludes them; no age signal stands behind it (§4.2) |
| No one presses for the dead | rule | a synced or shared passkey presses for whoever holds it (§4.2) |
| No reminders, countdowns or visible waiting lists | rule | an interface can always add one |
| No withdrawn mark on a film or a press | rule | a checker could add one |
| Every render is listed to the actor; drafts are deleted at release | rule | the studio keeps the list and does the deleting (§8.6) |
| An invented face is not an enrolled holder | rule | a render-time match against enrolled likenesses (§9.4) |
| The studio takes no share of a film's receipts | rule | a price list could change |
Five guards are properties: the four above and the approved-file guard, by its default. The rest are rules or trust, and the design says so rather than calling them properties.
Remove the Preamble's passage and every reference to it. The mechanism is unchanged: two presses, a manifest, a check, an encrypted likeness. The passage is a recognition of the problem's shape, tier flourish, and carries no load.
Disclosed: the composition of §15 and its variants; the statement contents (§6.1, §8.2); the release manifest's structure and the check (§9); the key structure at the level of §7.2; the guards and their classification (§13.3).
Withheld: the statement formats and field encodings; the key-derivation inputs, the expiry and the wrapping scheme; the enrollment capture protocol and its liveness thresholds; the render-time match threshold for extras (§9.4); the perceptual-fingerprint and watermark parameters (§12.3); the manifest's serialization; every interface. Why: the mechanism is designed and not built, and a specification for something unbuilt and unscheduled is the part most likely to be wrong. The claims are published so that the composition cannot be enclosed; the specification will be published when it is built and tested. The authors' open specification of signed statements reserves the record a seat holder keeps (§10.3) and will define its format when it is built [10].
The claims below are the survivor of the census in §2, as narrowed by the adversarial pass of §13.1, and nothing wider. Claims 2 to 11 depend on claim 1 and are not claimed apart from it: a verified, permissioned likeness inside a video app, payment to a likeness owner, a signed project-specific agreement before replica work, review of output and after-the-fact deletion, a maker's provenance signature on a media file, a named actor's identity assertion over a file's binding, a reference from a file to consent held by an authority, per-segment hashing of streamed media, and release of a model key after hardware attestation are prior art, cited, and are not claimed. The full census (§2.5) adds to that list, as prior art cited and not claimed: a performer-signed authorization on a public ledger; a studio's review of the script and the final image with its performers; salting log entries for privacy; matching uploads against enrolled faces; a legal requirement of separate consent to edit a face or voice; and deriving per-purpose encryption keys from a passkey. Each claim is disclosed for the public domain.
Claims 2 to 11 rest on narrow nulls in a quick census and a full one whose patent, academic and standards reach was through a general web index, the standards group being blind; they are correspondingly weak. Later prior art may narrow any of the eleven. Claim 11 was not searched by either census.
This paper asserts no empirical claim about behavior: it does not claim that the check will be used or that the design will reduce misuse of likenesses. The first film made with the mechanism is planned as the studio's own short, made with its founder's own likeness, family members who press, and invented co-stars, and three predictions about it were registered in the authors' public prediction register before it was made and before the mechanism was built [23]. The founder chose each band:
Each falsifier is the band's edge. The register records that the authors' earlier estimate of the cost, roughly 6 to 33 USD a minute for one take, was stated before the founder chose his band.
The commentator's lesser gods could not tell the double from the one who made it, and it did not matter, because the double taught only what it had been told to teach. The resolve was enough, and it was the Buddha's. A film cannot rest on that. Its doubles are made by a machine, for an author, out of other people's faces, and the person whose face it is may never see the film. So the resolve has to be written down, by the person whose face it is, over the words they agreed to and then over the exact pictures they were shown, in a place no one can quietly change. The viewer is a lesser god, and can now tell.
| Term in this paper | Standard term |
|---|---|
| B-Double; lent likeness | digital replica of a real person's face and voice, licensed per work by that person |
| holder | person who enrolled their own digital replica |
| actor; depicted person; credited performer; lent role | person whose digital replica appears in a work as a lent role (the industry draft's performer) |
| author | creator of a generated film who casts it |
| studio | operator of the rendering service |
| press | WebAuthn passkey assertion with user verification over exact text |
| casting statement; press one | signed per-work likeness license over the screenplay, role, style and shares |
| release statement; press two | signed approval over the cryptographic hash of the exact final video file |
| work key | per-work decryption key derived by the holder's authenticator (WebAuthn PRF extension) |
| the log | public append-only transparency log of signed statements |
| salted fingerprint | hash commitment to a statement with a salt of its own party's, stored in place of the statement so that records cannot be paired |
| invented role | generated character belonging to no real person (a co-star or an extra), credited as invented |
| rendition list | fingerprints of every encoding and streaming segment of a work, pressed with the release statement |
| manifest; release manifest | list of the signed statements, with salts, distributed with the video |
@name page | public registry page listing the passkeys bound to a name |
| the check | verification of a video file against its manifest and the log |
| render style | rendering mode of the replica (photorealistic, cartoon, three-dimensional) |
| casting call | per-work solicitation of replica licenses, answered by holders |
| leaving | the actor's signed termination of the operator's distribution of a work |
| gift release | free public release of a work at a date fixed in advance |
| thank; Thank After | voluntary payment or message from a viewer after watching |
| remake | a new work from the same screenplay, with new presses from every lent actor |
| screenplay license | a two-party signed grant of rights in a screenplay (a remake, or characters and world for sequels and prequels), never of a likeness |
| B-Ticket; seat | pre-sold, non-transferable-above-face-value admission to a premiere event |
Thon Ly conceived the mechanism's purpose and its frame: the third kind of video, starring real people's own likenesses; the author who writes a film with AI assistance and posts casting calls to friends, family and the public; the actor who signs permission for their likeness to be used in one film; the cast sharing what a film earns; the name; the observation that a likeness may be rendered in any style; the pre-sold premiere seat, first conceived years earlier as his final project for a blockchain certification; and the question of who pays to serve a film. He ruled on every design choice recorded here, including the choice to fix shares before generation, to split thanks from the first dollar, to release films as gifts, to refuse tokens for seats, to keep advertising out of viewing, and to lift a sequencing gate so that this tool may open to every user. Miss Aquarius℠, the name under which this corpus discloses its AI collaboration, proposed the two-press structure, the fingerprint computed on the actor's device and its checker, the encrypted likeness released one work at a time, the salted log with the manifest carried by the film, the treatment of declines and recuts, the variants and the census predictions; ran the census through a research agent and re-read the decisive sources before they entered this text, marking what it could not read; ran the adversarial pass with a separate reader, re-read the two industry specifications that reader found, and narrowed the claims to what survived it; checked the commentary passage of the Preamble against the Pāli; and drafted this paper. No AI system is an inventor. Thon Ly is the natural person who conceived the invention.
B-Double is the institution's working name for the service of lending one's own likeness to a film; no mark is claimed in it here, and the mechanism is dedicated to the public domain and may be implemented under any name. B-Ticket is likewise a working name. Neither name has been cleared; in Australia a B-double is a kind of truck, and a B ticket may be heard as second-class. THonly™, B-Sala™, B-Premiere℠, B-Origin℠, HeartBank® and Miss Aquarius℠ are marks of their respective holders. No mark is licensed by this publication.
abh01a.att.xml, lines 505 and 513 (romanized Pāli, VipassanaTech tipitaka-xml at commit 05d5d3c). Translations here are the authors'. Tier: commentary; cited as a recognition, not as authority for the mechanism.thonly/publications program/prediction-register.md, entries P-BD1, P-BD2 and P-BD3, registered 2026-11-19.— End of defensive publication —