.jpg){kind=link}
Personal data
Customer: Why did you use my data to target this ad?
Advertiser: Which data source and permissions supported that use?
Layer 4 · Technical Conformance
You shouldn’t have to play detective.
A burger has ingredients. Advertising has data, creative assets, permissions and decisions. When AI acts for a company, the record should connect what it was allowed to do, what people were told and what it actually did.
An AI ad agent is software that creates ads, selects audiences or spends budget. It can hit its target and exceed its authority. Layer 4 supplies the evidence needed to test that gap.
See the proposed record ↓
Photo: cyclonebill / Wikimedia Commons. CC BY-SA 2.0. Resized. Original.
From intention to evidence.
Preserve evidence to test actual behavior against authorized intent, permissions and disclosures. This covers personal data, creative work and agent decisions—not only the finished ad.
Changing behavior
Customer: Why did the message or offer change?
Advertiser: Did the agent stay within its approved audience, task and budget?
Accountable owner
Customer: Who can put it right?
Advertiser: Who approved the action, and could they stop it?
Layer 1 defines intent and boundaries. Layer 3 defines workflow controls. Layer 4 supplies checkable evidence. Layers 5 and 6 use it for accountability and ongoing assurance. The receipt, seal and directory remain proposals for discussion.
One mark across ads, videos and articles.
Each mark below opens the fictional record. The design shows a possible access point; it does not certify consent, safety or approval.
Social ad
Example Brand · Sponsored
A familiar product. A checkable history.
CONTENT RECORD · AI ASSISTED →Video ad
Example Ad · Video 3
Illustrative player, not a playable video.
CONTENT RECORD · AI ASSISTED →Article / content
What went into the grinder?
Knowing where something came from is a starting point. Permission, changes and responsible decisions need evidence too.
Illustrative editorial placement.
CONTENT RECORD · AI ASSISTED →
A visible mark, with room for the record.
The mark has its own space. Click this seal to inspect the actual presentation’s production disclosure and file fingerprint.
Human presence, approval authority and signed credentials are proposed controls; none is verified by this demonstration seal.
Example Ad, Video 3
An actor challenges the use of their likeness. Start with the record, then hear each party. The sample facts below are invented; they are not a verified approval.
Human × AI assisted
Illustrative receipt · brand authorization proposed
- Source · where it came from
- Background made with an AI image tool on 2 Oct 2026. Product photograph supplied by the fictional brand.
- Permitted use · what it can be used for
- In this example, the actor allows this version of their likeness in social-media ads shown in the US through 31 Mar 2027. No face or voice alterations are allowed. Check the underlying agreement and current permission before use.
- Responsibility · who can answer for it
- Approver #4471 · Content Integrity Reviewer · sample approval at 10:42. The organization would retain the person’s identity, presence evidence, delegated authority and ability to reject or stop the use. This is a pseudonymous reference, not an untraceable person.
- Changes · what happened
- Background replaced. The actor’s face and voice are unchanged in Example Ad, Video 3. A later edit would require a new version and review.
- Verification · how to check it
- In a production system, a fingerprint would bind the record to this exact file. Public checks would show scope, status and signatures; an authorized auditor would inspect protected identity, rights and decision evidence. Those checks are not implemented for this fictional ad.
What does each person need?
The receipt supports a decision. It does not assign blame automatically or replace a fair investigation.
Collect → Review → Verify → Issue
Capture what went in, what was permitted and what the ad agent actually did.
1. Collect
Approved task and limits. Audience-data supplier and permitted use. Images, voices and music rights. Creator disclosures, tools, existing signed origin/edit records, watermark and AI-detection results.
2. Review
Compare proposed targeting, content and spending with purpose and permissions. Record checks, exceptions and approval or rejection.
3. Verify
For human approval, check identity, presence, authority and power to reject or stop. Link actual actions and method changes to inputs, approvals, versions and time.
4. Issue
Record what was authorized, disclosed and done. Include evidence links, scope, date and accountable owner. Record automation explicitly.
Carry and recheck: preserve method changes, human check-ins and interventions across handoffs. Recheck after changes, withdrawal or expiry. A person who can only watch is not an approval control.
Credentials, watermarks and detection do different jobs.
Content Credentials: signed provenance records about origin and edits. Watermarks: signals embedded in media; a supported signal may help recover a record. AI-detection results: estimates, not proof of origin, permission or truth.
Link credentials to permission and decision evidence. Neither a signature nor a watermark makes an ad’s claims true. C2PA explainer · Record recovery and soft bindings.
Change one fact. See what should happen.
From evidence to action
- Protect: assess harm and pause disputed use when appropriate.
- Preserve: keep the ad, approval, consent, placement, and edit logs.
- Hear: let the actor, creator, reviewer, and brand respond.
- Decide: identify who controlled the disputed step and explain the outcome.
- Repair: correct, withdraw, re-review, or reinstate; record the response and any appeal.
An insurer can inspect this trail against the actual policy. A receipt does not create insurance coverage.
The visitor said no. Did the system listen?
Someone declines tracking, but a marketing tool still collects data. The team compares the person’s choice with the system’s behavior, corrects the mismatch, and tests again.
That is a practical alignment test: compare what someone allowed with what the system actually did. A policy or approval record alone cannot show that the live system followed it.
| Question | Evidence to retain | Who acts? |
|---|---|---|
| What was allowed? | Consent choice, purpose, scope, timestamp, and policy version. | Privacy owner confirms the rule; the person can challenge it. |
| What actually happened? | Observed tags/events or ad placement, edits, version, and timing. | Technical owner checks the behavior. Separate observation from suspected cause. |
| Where did they differ? | A reproducible mismatch, affected uses, and known evidence gaps. | Marketing, privacy, and engineering review together; hear the affected person. |
| Who corrects it? | Named fix owner, change made, rationale, and completion date. | The team controlling that configuration remediates; leadership supplies authority and resources. |
| Did the correction hold? | Test result, subsequent observations, and checks after releases. | Recheck and reopen if the mismatch returns. Keep the response and remedy in the record. |
The ad version of this test: Was the actor’s likeness used only as permitted? Did the published version match the one approved? Did the distribution stay within scope? If it changed, who reviewed it again?
Illustrative example: permission, observed behavior, correction, and a check that the correction worked.
Disclosure is already a legal obligation in some markets.
EU and US state rules impose different marking and disclosure duties. A label can answer “was AI used?” without establishing permission, a person’s authority, or who controlled a disputed use.
- EU AI Act, Article 50: in effect since 2 Aug 2026. Machine-readable marking for covered AI outputs and disclosure of deepfakes. Older systems have a marking grace period until 2 Dec 2026.
- New York synthetic performers: in effect since 9 Jun 2026. Disclosure in covered ads using AI-generated synthetic performers.
- California SB 1050: effective 1 Jan 2027. Clear disclosure in covered ads prominently featuring synthetic performers.
Regulatory references as of 11 Oct 2026. Scope and exceptions apply. The record supports review; it does not establish legal compliance on its own.
Penalties and disputes
Keep the applicable rule, disclosure decision, consent scope, and delivery evidence. Assess actual obligations by jurisdiction and use.
IP and likeness
Source and permission records help a rights holder challenge misuse and help a creator or company show agreed use.
Insurance
Optional generative-AI exclusions can affect advertising-injury coverage. Compare the evidence with the actual policy; the receipt does not make an ad insurable automatically.
Alignment after approval
The record should follow changes and delivery. An earlier approval cannot establish that every later placement or edit stayed in scope.
An AI LUMAscape linked to evidence.
A LUMAscape groups companies by their role in advertising. The proposed evidence map would link each participating brand, agency, AI tool, data supplier and adtech partner to requirements and scoped evidence. A company’s listing would not be a blanket compliance certificate.
Layer 1 defines intent. Layer 3 defines controls. Layer 4 supplies evidence. Layers 5 and 6 use it for accountability and ongoing assurance. The record connects these commercial roles:
Creators and tools
Disclose origin and edits. Retain license and ownership evidence.
Brands and agencies
Require permission and approval for the exact version. Preserve the record through handoffs.
Platforms and ad tech
Carry and check the record; expose status and routes to challenge.
Auditors and verifiers
Test assertions against protected evidence and explain uncertainty.
Insurers and buyers
Evaluate evidence and risk against actual contracts and policies.
People and rights holders
Inspect use, challenge misuse, and seek correction without surrendering public privacy.
The proposed incentive: purchasing, distribution, and risk review should value checkable proof. This is a discussion proposal, not an existing certification or a promise of lower premiums or better reach.
The committee must decide what proof is sufficient for each risk, who verifies it, how long it is kept and who answers for a missing handoff. Any assessment must show its requirement, reviewer, scope, date and result. Missing evidence is not the same as proven nonconformance. Small creators need a proportionate way to participate.
The everyday examples behind the proposal
Diamonds: same carat does not answer mined or lab-grown, origin, or quality. Art: a convincing Picasso needs supporting provenance and authenticity evidence. Digital trading cards: issuer, ownership and transfer history support a claim about the collectible. Elevators: an inspection certificate identifies a check, issuer and validity period. Used cars: “one careful owner” invites a request for service and accident history. Hamburgers: mixed ingredients need a reconstructable source trail. These analogies explain the instinct; they are not engineering specifications.
Who supplied what—and how we used it.
Jay’s article supplies the narrative. The technical contribution attributed to Sean supplies the proposed checks and privacy model. This table traces the rest of the inputs. Contribution does not mean verified approval.
Some discussion notes do not identify the speaker. Those stay joint or unspecified; we do not invent individual credit. The fictional ad case is separate from the actual production record.
| Source / contributor | Input supplied | Details used | Role in the deck / demo | Evidence status |
|---|---|---|---|---|
| Jay Mandel | Mystery Meat, Mystery Data — supplied article | Hamburger / grinder analogy; diamonds, art, used-car history; source ≠ truth; carrying evidence forward. | Opening narrative and limits of the receipt. | Named author in supplied article. |
| Jay Mandel | Directions and revisions in this conversation | 10-minute format; practical accountability; stakeholder empathy; IP protection; badge terms; content-record seal and example site. | Editorial direction, case framing, visual design, and demo requirements. | Direct instructions; final human approval not recorded. |
| Sean Meier / VRFD | Technical contribution described in the supplied discussion | Creator disclosure; source collection; watermark and Gen AI / deepfake signals; private organization-bound biometric methods; organization-scoped approver ID; C2PA carriage. | Proposed Collect → Review → Verify → Issue flow and privacy boundary. | Attributed technical proposal; speaker-by-speaker authentication not supplied. No controls verified here. |
| VRFD | VRFD-Mockups.html — supplied mockups | Visible label and presence-receipt designs. | Design reference for opening evidence from a mark. | Mockups, not proof of deployed capabilities. Exact authorship not established from the file. |
| Joint / working notes | Layer 4 technical conformance presentation and supplied receipt text | Source, permission, approval, changes, public checks / auditor evidence; fingerprint, authority, presence; open questions. | Receipt fields and stakeholder investigation example. | Working proposal. Unlabeled discussion is not assigned to a specific speaker. |
| Joint / working notes | Layer 4 companion / rules notes supplied 11 Oct 2026 | Existing standards, experimental controls, drafts, gaps, regulation and insurance references. | Background research map; separation of existing components from proposed end-to-end system. | User-supplied research notes; not every cited item independently verified. Statements that no system exists are not established by this demo. |
| cyclonebill | Hamburger (5).jpg — Wikimedia Commons | Hamburger photograph. | Visual for mixed inputs and provenance. | CC BY-SA 2.0; resized; source link and license below. |
| NWRGeek | Massachusetts Elevator Inspection Certificate — Pass Card | Real framed inspection-certificate photograph, 2012–2013. | Visual analogy for inspector, date, authority, and validity. | CC0; resized; historical certificate, not a current inspection. |
| C2PA | C2PA / Content Credentials explainer and specification | Provenance, asset binding, edit history, signatures, and verification limits. | Technical reference for carrying records; supports distinguishing provenance from truth. | Primary standard documentation; this demo does not issue C2PA credentials. |
| ChatGPT / OpenAI | Assistance in this conversation | Drafting, synthesis, PDF / website code, case wording, fingerprint implementation, and drawn seal. | Production assistance; fictional scenarios and implementation. | AI assistance disclosed. Not an independent auditor or human approver. |
| Example case — invented | Example Ad, Video 3; approver #4471; actor consent | Sample background edit, permission scope, approval time, and stakeholder dispute. | Shows how records could support fair review and corrective action. | Fictional values. No actual actor consent, approver identity, or biometric check. |
| Published artifact | This exact presentation PDF and record.json | SHA-256 fingerprint; disclosed workflow, sources, and verification gaps. | Allows matching a file copy to the published record. | Byte-level check implemented; no signed approval, identity proof, or presence proof. |
| Operational alignment | Illustrative consent scenario | Compare permission with observed behavior, assign correction, and recheck. | Explains how teams use evidence to hold decisions accountable. | Illustration, not a verified deployment or third-party endorsement. |
| Rawpixel / public-domain source | Diamond photograph, image 5913653 | Faceted diamond product photograph | Visual on the workstream progress page | Source labels image CC0 / public domain; responsive display. Source. |
| Institute for Advertising Ethics | Summit I report: The Intention Advantage, 22 Apr 2026; prepared by Andrew Susman | Consumer choice; agents adapting beyond initial instructions; human check-ins; organizations retain responsibility. | Broader data and agent-behavior scenarios; authorized / disclosed / actual comparison. | User-supplied institutional report, which states it reflects editorial synthesis. This page does not imply IAE or IEEE approval of the prototype. |
| Layer 1 workstream | Intent-alignment working material supplied during deck development | Purpose and boundaries before performance; downstream controls must be testable. | Defines the boundary between intent, workflow controls and Layer 4 evidence. | Working framework; not an adopted standard. |
| Jay Mandel and Sean Meier / working presentation | Layer 4 Trust Stack — Review Ready; 12 Oct 2026 presentation, six slides / ten minutes | Five evidence fields; personal data, changing behavior and accountable ownership; Collect / Review / Verify / Issue; committee decisions. | Current alignment reference for this companion page and linked PDF. | Working proposal for discussion. A named presenter is not a recorded final approval. |
| NBA Collect by Topps | Official digital trading-card image | Digital card issuer and ownership analogy. | Source image in the presentation. | Official product image used as an illustrative reference; no endorsement or certification is implied. |
Links and license details
- Hamburger — cyclonebill · CC BY-SA 2.0 · resized.
- Elevator certificate — NWRGeek · CC0 · resized.
- C2PA explainer · standards reference, not certification of this demo.
Jay’s article, Sean-attributed discussion, working notes, original deck, and VRFD mockups were supplied to this conversation. They are identified here without publishing the source files or private correspondence.
Additional references named in the supplied working notes
These are the research inputs named in the companion notes. Inclusion records their role; it does not claim that the prototype implements them or that every supplied interpretation was verified.
- IAB AI Transparency & Disclosure Framework V2 — disclosure policy and risk test.
- C2PA 2.4 ingredients, actions, and asset binding; Content Credentials Verify / c2patool — provenance carriage and checking.
- Google SynthID; Reality Defender removal test (vendor report) — watermark / detector context, not detector proof in this demo.
- W3C ODRL 2.2 — permission scope and conditions.
- HAPS Sep 2026; NIST SP 800-63-4 — presence and identity research context.
- W3C VC 2.0; SD-JWT; CAWG vLEI draft; CAWG Identity Assertion 1.3 — authority and identity assertion references.
- Trust Stack Layers 1, 3, 5, and 6; Proof-of-Control working draft; IAE AIRS — governance, accountability, verifiability tiers, and incidents.
- EU AI Act Articles 50 and 99 / EC FAQ; California SB 1000, AB 2713, AB 853, SB 1050; New York synthetic-performer disclosure — regulatory context.
- ISO CG 40 47 / 40 48 — optional generative-AI insurance exclusion context; actual coverage depends on the policy.
Inspect our content record.
Here is the record for the presentation that brought you here. We can disclose the process and verify the file’s bytes today. We have not implemented the full certificate system.
Human × AI assisted
- Human direction
- Jay Mandel supplied his article, examples, and revision instructions. This records direction, not verified final approval.
- AI assistance
- ChatGPT drafted wording, wrote PDF and website code, and assembled sourced photographs and drawn vector symbols.
- Image sources
- Hamburger: cyclonebill, CC BY-SA 2.0. Diamond: Rawpixel, CC0. Digital card: NBA Collect by Topps, official product image. Elevator certificate: NWRGeek, CC0, historical 2012–2013 certificate. Images resized for the slides.
- Human sign-off
- Not recorded.
- Proof of presence
- Not performed. No biometric data collected.
- Signed credential
- No digital signature or C2PA credential issued.
Download the exact presentation Who supplied what? Read the record
Check the file fingerprint
Compare the PDF’s SHA-256 fingerprint with the published record. A match confirms the same bytes. It does not prove ownership, truth, or human approval.
Loading the published fingerprint…
No check run yet.
Your selected file stays in your browser; it is not uploaded. The record is not digitally signed, so someone controlling the website could change both file and fingerprint.
The example does not run AI/deepfake detection, watermark recovery, biometric authentication, or private auditor access. Those controls need implementation and independent testing. No private identity or biometric evidence is published here.
Keep the claim smaller than the proof.
What do the badge terms mean?
AI assisted: human-led work with Gen AI help. AI generated: Gen AI creates the core content. AI free: no Gen AI in the declared workflow. These are proposed definitions; declare the workflow and its boundaries. An AI detector returning no signal does not prove “AI free.”
What does the receipt protect?
Creators can document ownership and agreed use. Actors can inspect consent and challenge misuse. Reviewers can show what they knew and their authority. Brands can reconstruct approval and correction. Auditors and platforms can assess evidence without publishing private identities. Consumers can see the status without investigating the entire supply chain themselves.
How does this relate to Content Credentials?
A production implementation could use C2PA to carry signed provenance. Provenance can support checking history and tampering; it does not establish that an ad’s claims are true. Human presence and authority require additional evidence. Read the C2PA explainer.
Open questions: What proof is sufficient for each risk? Who verifies it? Who can challenge it? How long is evidence kept? Who answers for a missing handoff?