Open standard · Draft 0.1 · October 2026

An open standard for generative models

Credit and protection from one mechanism.

A draft standard for operating generative models in the open, starting with image generation and extending to language models. Each requirement names its reference implementation, most already open source, and each serves two ends at once: crediting the people whose work a model draws on, and preventing the harms of using that work without them.

1. Summary

Generative models are built from people’s work and can be turned against those same people. The two problems are usually handled apart: attribution and payment as a commercial question, harm as a question for moderation. This standard treats them as one. The record that shows whose work shaped an output is the same evidence that stops a forgery, traces it, and pays the person it drew on.

It is written for the open-source world first. Its reference implementations Veil8, VeilGarde and Braid are published under GPL-3.0, and Veil itself is licensed the same way, so any model host or developer can adopt the same controls, read them, and improve them in the open.⁠[2][3][4] The system it builds on was introduced in the first part of A System for A.I. Safety.⁠[1]

In the requirements, “must” and “should” carry the meanings given in RFC 2119.⁠[6] This is a draft: numbers and wording will change as the reference code and the review around it mature.

The case

2. Why open source should lead

As the post that introduced this system puts it, open-source communities are where real innovation happens, and they need safe systems to operate in.⁠[1] They are also where controls are hardest to add after the fact: once a model’s weights are downloaded, a host’s rules no longer reach them, as the policy brief documents.

A standard for that world has to be open in the same way. Its checks run on the machine that runs the model, without sending anyone’s images or writing to a third party. Its reference code is readable, so every claim about it can be checked. And its improvements stay public: under GPL-3.0, anyone who distributes a modified version of the reference code publishes its source.

Done this way, an open standard reaches where regulation alone does not. It puts attribution and protection inside the models people download, not only inside the services they visit, and it lets the companies that sell generative media compete on how well they credit and protect the people their models learn from.

3. Two ends, one mechanism

Every requirement here is chosen so that one mechanism does both jobs.

The check during generationPart I

Credit. Regions that match a registered photo are credited to the person who registered it.

Protection. The same match, past a threshold, halts a reproduction before it is finished.

The run recordPart I

Credit. It is the basis for royalties: whose work, how much, at what score.

Protection. It is the evidence of what a service produced, and when.

The signaturePart I

Credit. It ties a released image to the run that credits its sources.

Protection. It traces a misused image to its producer, and an unsigned image of a registered person stands out.

Shares and citations per tokenPart II

Credit. Each contributed body of writing is paid by its share of what the model wrote.

Protection. The same pass that attributes a token tests whether a strand learned the text, so reproduced writing shows even where its index leaves it out.⁠[4]

The standard

4. How Veil closes the loop

Once a model provider installs Veil, an image moves through six steps, and the last feeds the first. Steps 2 to 4 run inside the provider’s service; the others belong to the people images depict, and to anyone who meets an image later.

  1. Register

    The person

    Photos are cut into 32-pixel tiles and encoded one way, under a key that stays on the person’s machine. The registry holds strings no one can turn back into pictures.

    Veil8
  2. Check

    Inside the provider

    Between denoising steps, the forming image is encoded the same way and compared with the registry. A reproduction halts the run.

    Veil
  3. Credit

    Inside the provider

    Each run keeps a record of whose registered work it drew on, how much, and at what score: the basis for royalties.

    Veil
  4. Sign

    Inside the provider

    What is released is signed. The signature adds nothing to the pixels, holds through a crop or re-encode, and is filed with the run.

    VeilGarde
  5. Detect

    Anyone

    An image in circulation is checked. A signature traces it to its run; a registered likeness with no signature is a risk signal.

    VeilGarde
  6. Answer

    People, platforms, courts

    Findings support takedowns, claims and royalties, and what got through tightens the check at step 2.

Runs inside the model provider’s service, once Veil is installed Runs with the person, or with anyone who meets the image later

The loop closes only where providers install it. An image with no run record behind it cannot be traced, and as signing spreads that absence becomes a signal in itself. The aim is an industry that is tamper-evident, one that will treat “an absence of provenance credentials within media as a risk signal.”⁠[1]

5. Part I: Image generation

A conforming image generator meets these requirements. Each names its open-source reference.

  1. IG-1

    Register without collecting

    Protected images must be stored only in a one-way encoding that cannot be turned back into pixels, under a key held by whoever registered them. A query written under another key must be refused, not searched.

    Reference: Veil8, which writes each 32-pixel tile as 1,024 keyed two-letter words; with the key and a trained decoder, an attacker recovers only a smear of average colours.⁠[2]

  2. IG-2

    Check while generating

    The generator must compare its forming image with the registry at intermediate steps, not only once the image is finished, and must publish how often it checks.

    Reference: Veil

  3. IG-3

    Halt on reproduction

    When a registered image is being reproduced past the provider’s published threshold, the run must stop and record why.

    Reference: Veil’s gate, whose thresholds are policy the provider sets.

  4. IG-4

    Credit by contribution

    Every run must keep a record of which registered work it drew on, in what share and at what score, for long enough to settle royalties and claims.

    Reference: Veil’s royalty sheet, kept for every checked step.

  5. IG-5

    Sign what is released

    Every released image must carry a provenance signature filed with its run record. Signing must not alter the image, and the signature should hold through a crop or a re-encode.

    Reference: VeilGarde, whose signature is read from each tile’s fragility under sixteen fixed warps.⁠[3]

  6. IG-6

    Offer detection

    The provider should let anyone check an image against its signatures, and should treat an image of a registered person without a signature as a risk signal, not as clearance.

    Reference: VeilGarde’s detector, a stack of independent members whose every score is kept beside the verdict.⁠[3]

  7. IG-7

    Publish measurements

    Thresholds, and the accuracy of matching and detection, must be published with the date and the data they were measured on, failures included.

    Reference: VeilGarde’s battery and census, published with their dates, their data and the images they miss.⁠[3]

6. Part II: Language models

The same principles carry into language models, where the work is writing and code rather than pictures. Braid shows the shape: a shared pretrained base, the commons, and one strand for each contributed body of writing, called a Thread. For every token it writes, Braid reports how much each strand supplied; the shares are exact and sum to one, each strand’s part is cited to the place in its writing it came from, and every answer is logged so its shares and citations can be checked afterwards.⁠[4]

  1. Contribute
  2. Train a strand
  3. Answer
  4. Attribute each token
  5. Cite
  6. Log
  7. Pay

RaoLM, a language model in development on Braid, adds the commitments that make it consent-based: on-machine first, consent before aggregation, addresses that travel with the words, citation as architecture, and personal and general knowledge kept separate.⁠[5]

  1. LM-1

    Consent before aggregation

    A body of writing must join a model only when its owner contributes it.⁠[5]

    Reference: RaoLM

  2. LM-2

    Keep the address

    Every passage must keep an address that travels with it, so a citation resolves to the exact place it came from.⁠[5]

    Reference: Thread, which names each partition by its SHA-256 hash.

  3. LM-3

    Attribute every token

    The model must report, for each token, how much each contributed source supplied, as exact shares that sum to one.⁠[4]

    Reference: Braid

  4. LM-4

    Log for verification

    Every answer must be logged so its shares and citations can be checked afterwards, by someone other than the provider.⁠[4]

    Reference: Braid’s session log and verifier.

  5. LM-5

    Detect what the index misses

    The model should test whether a strand learned a passage, not only whether its index holds it, so reproduced writing is caught even where it was never indexed.⁠[4]

    Reference: Braid’s detection reading, beside its citation.

  6. LM-6

    Pay from attribution

    Royalties must be derived from the recorded shares and citations, per contributed source and per cited document.⁠[5]

    Reference: Braid’s credit, per Thread and per cited document.

Adoption

7. Conformance and status

A provider conforms to Part I by meeting IG-1 to IG-7, and to Part II by meeting LM-1 to LM-6. The reference implementations stand as follows.

Veil In betaThe check, the halt and the record · IG-2 to IG-4
The Mac app is in beta, and a working preview runs on this site. It matches registered pixels today, and matching a person’s likeness is in development.
Veil8 Open sourceRegistration · IG-1
Released under GPL-3.0, with the stated aim of standardizing the encoding for others building risk management in generative media.⁠[2]
VeilGarde ResearchSigning and detection · IG-5 to IG-7
Open source under GPL-3.0. Its v1 detector called 90.0% of unseen real images real but only 34.5% of unseen generated images generated, measured October 21, 2026, with work under way to reach 80% on generated images.⁠[3] Its watermark signs pictures in Veil’s beta without changing a pixel; how well a signature holds through crops, re-encodes and screenshots is still being resolved.
Braid Open sourceAttribution in language models · LM-2 to LM-5
Released under GPL-3.0. In a proof run on October 19, 2026, a trained Thread told writing it had studied from new writing every time, and its index found the source of its own writing every time.⁠[4]
RaoLM In developmentConsent and royalties · LM-1, LM-6
Described at lm.rao.nyc, whose figures illustrate intended behavior rather than record a released model.⁠[5]

8. Where it meets policy

The policy brief recommends making these controls a duty for every service that generates realistic media of people. This standard is the technical side of those recommendations.

A service-level dutyPolicy recommendation 2
Parts I and II are what a service would implement to meet it, whether it hosts open models or sells its own.
A consent-verification registryPolicy recommendation 3
IG-1 and LM-1 define what the registry holds and how it is queried without collecting anyone’s images or writing.
A presumption of non-consentPolicy recommendation 4
The registry check and run record (IG-1, IG-4) show whether a registered work was drawn on when an output was made.
A safe harbor for certified servicesPolicy recommendation 6
Certification can rest on conformance to a published, open standard, checked against public reference code.
Licensing and royaltiesPolicy recommendation 7
IG-4 and LM-6 make payment follow the record of whose work was used, image by image and token by token.

Read the brief: Likeness, consent and generative media.

Sources

  1. [1]Ritesh Pakala, “Media Safety via Veil, Veil8, VeilGarde.” A System for A.I. Safety, part 1 of 5. paka.la
  2. [2]Veil8: one-way image fingerprints for similarity matching without pixel recovery. Source code and documentation, GPL-3.0. github.com
  3. [3]VeilGarde: watermarking and AI visual risk management. Source code and documentation, GPL-3.0; accuracy as measured October 21, 2026. github.com
  4. [4]Braid: a decentralized language model that cites. Rao Studios. Source code and documentation, GPL-3.0. github.com
  5. [5]RaoLM, a braided language model. Rao Studios, in development; its figures illustrate intended behavior. lm.rao.nyc
  6. [6]S. Bradner, “Key words for use in RFCs to Indicate Requirement Levels,” RFC 2119. Internet Engineering Task Force, March 1997. rfc-editor.org