VERA™ Verification Standard

1Overview

AI visibility without trust is temporary. Trust without infrastructure doesn’t scale. AITP™ is the infrastructure.

Trust drives recommendations.

This Standard defines how VERA evaluates, verifies, and classifies digital content to produce trusted, machine-readable verification signals. It is the load-bearing protocol document for the registry: the Governance Charter constitutes the authority; this Standard defines the work.

VERA classifications are evidence-based signals, not adjudications. They inform decisions made by humans, platforms, and AI systems, they do not replace them.

2What VERA is, and is not

VERA is

• A neutral verification infrastructure producing classification signals
• An evidence registry, cryptographically anchored to QUBIT Blockchain®
• A machine-readable trust layer for AI systems and platforms

VERA is not

• A publisher or editorial authority
• A legal adjudicator of liability
• A takedown service or content moderator
• A journalistic fact-checking organization
• A guarantor of intent or motivation behind content

Verification scope

VERA does not verify entities against the web. VERA captures what AI platforms say about an entity and compares it to the entity’s own signed canonical record. When they disagree, Procedural Verification adjudicates, and FeedSync AI™ propagates the correction. The web is not the verification surface. The AI output is.

Trust Architecture: the editorial firewall

VERA operates editorial and community properties (VERA News™ and VERA Channels™) alongside the Registry. A strict firewall separates them from verification. Content VERA publishes, community activity, and AI-citation signals contribute only to measurement (visibility, discovery, and reach). They never contribute, directly or indirectly, to verification status, Trust Score, Registry status, or any authoritative verdict.

Authoritative scoring draws only on independent inputs: identity verification, corporate and regulatory records, signed evidence, cryptographic attestations, source provenance, and genuinely third-party citations. VERA-published sources are explicitly excluded: a citation from VERA News or a VERA Channels post is not treated as independent validation. VERA does not grade signals VERA created.

3Verification model

VERA operates on a hybrid model combining automated analysis with structured human review and cryptographic recording:

AI-Assisted Analysis  +  Evidence-Based Review  +  Cryptographic Recording

The AI-assisted layer is operated by AIBlackBox™, the QuantumX-owned multi-platform router that queries six independent AI platforms (ChatGPT, Claude, Gemini, Grok, Perplexity, DeepSeek) and consolidates findings. This avoids single-vendor dependency in the verification chain.

Visual forensic analysis runs through a vision-model cascade evaluating images and video frames for AI-generation markers, manipulation artifacts, and synthetic content signatures. All AI-assisted findings are evidence inputs; classifications require human-in-the-loop review for any record above moderate confidence.

4Evidence framework

Primary evidence

• Original source validation (publisher, creation timestamp, provenance metadata)
• Verified publisher confirmation (cryptographic signing keys, official channels)
• Metadata signatures consistent with C2PA / Content Credentials standards

Secondary evidence

• Cross-platform validation across AIBlackBox-routed AI platforms (consensus or divergence is itself a signal)
• Third-party verification from named domain experts or aligned registries
• Forensic analysis: pixel-level, frequency-domain, generation-marker, and metadata-tamper detection

Contextual evidence

• Timeline consistency (does the claim fit known events, dates, locations?)
• Event verification against public-record sources
• Narrative alignment (does the artifact support or contradict surrounding context?)

Each VERA record discloses which evidence categories were applied. Classifications relying on only one tier are constrained to lower confidence levels.

5Classification taxonomy

VERA records assign one of six classifications. Multiple classifications can apply to the same artifact (e.g., AI-Generated AND Contextually Misleading) and are recorded jointly when so.

• Verified Authentic, provenance and content integrity confirmed via primary + secondary evidence
• Contextually Misleading, artifact is authentic but presented in a context that materially distorts its meaning
• Altered / Manipulated, artifact has been edited from an authentic original; modifications materially change meaning or fact
• AI-Generated / Synthetic, artifact was generated by AI systems (regardless of intent)
• Deepfake / Fabricated, AI-generated artifact specifically designed to impersonate a real person, event, or entity in a deceptive manner
• Unverified / Under Review, insufficient evidence to classify; record open for further evidence

AI-Generated vs. Deepfake, the line

A piece is AI-Generated when synthesis is the verification finding. It becomes Deepfake when synthesis is combined with deceptive impersonation: depicting a real, identifiable person, event, or entity in a way that could reasonably mislead viewers into believing the depicted content is authentic.

6Confidence levels

Every classification carries a confidence level mapped to a numerical band derived from agreement across evidence tiers and AIBlackBox platform consensus.

• Low (60% to 74%): one evidence tier supports classification; conflicting signals from cross-platform validation
• Moderate (75% to 89%): two evidence tiers support classification; majority consensus across platforms
• High (90% to 97%): all three evidence tiers align; strong cross-platform consensus
• Verified (98%+): full evidence alignment AND Verification Council confirmation; cryptographically anchored as authoritative

Records below 60% confidence are not published as classifications, they remain in the Unverified / Under Review state until additional evidence is gathered.

7Process flow

1. Submission. Public, hash-authenticated. Evidence link required. Submitter identity preserved cryptographically; not always disclosed publicly.
2. Intake & deduplication. Submission hashed and checked against existing records. Duplicate evidence merges into existing case file rather than creating a new entry.
3. AI-assisted analysis. AIBlackBox queries six AI platforms; vision/audio forensic cascade runs where applicable. All outputs become evidence inputs, not classifications.
4. Evidence review. Human-in-the-loop reviewer evaluates AI findings against primary, secondary, and contextual evidence. Synthesizes into a classification candidate.
5. Classification. Reviewer assigns one or more classifications from §5 with supporting evidence summary.
6. Confidence scoring. Confidence level computed per §6. Records below 60% remain Under Review.
7. Anchoring. Record cryptographically signed and anchored to QUBIT Blockchain®. Public registry entry created.
8. Public availability. Record visible at vera.veraregistry.com/entry/{id} with full evidence summary, classification, confidence, and chain hash.

8Service-level commitments

• Standard classification: 48 to 72 hours from submission to published record
• Complex or contested cases: 7 to 14 days, with provisional Under Review status posted within 72 hours
• Council escalations: 21-day decision SLA
• Takedown / dispute requests: initial response within 48 hours; resolution per §10
• Acknowledged outage / degraded service: public status posted within 4 hours via the Press Room channel

9Transparency requirements

Every published VERA record discloses:

• Classification(s) from §5
• Confidence level and numerical band from §6
• Evidence summary (which tiers from §4 were applied)
• Verification method (AI-assisted only / human-reviewed / Council-confirmed)
• Submission timestamp and anchor timestamp
• QUBIT Blockchain® transaction hash for cryptographic verification
• Version history of any updates to the record

Records are immutable: updates create new versions linked to the original. The version chain is itself anchored to QUBIT, so historical state can always be reconstructed.

10Dispute and appeals protocol

Subjects of any VERA record may submit a dispute at any time. Disputes are reviewed in three tiers:

Tier 1, Standard review

• Submission of dispute with counter-evidence (free, no V-Verify required)
• Re-evaluation by an independent reviewer not involved in the original classification
• Decision within 7 business days; counter-evidence appended to record regardless of outcome

Tier 2, Verification Council escalation

• If Tier 1 result is contested, dispute escalates to the Verification Council (per Governance Charter §3.B)
• Council issues binding decision within 21 days
• If V-Verified counter-evidence is provided, original classification may be marked Disputed, V-Verified Counter-Evidence

Tier 3, Public commentary

• Disputed records remain visible with all dispute history attached
• Public commentary on disputed records is permitted via the Press Room channel
• No record is ever silently deleted; jurisdictional removals (GDPR, court order) are stamped and version-history-preserved

Appeals contact: disputes@veraregistry.com

11Accuracy framework

VERA publishes accuracy metrics quarterly:

• False-positive rate target: <2% across all classifications
• False-negative rate target: <5% across all classifications
• Inter-reviewer agreement target: >85%
• Annual independent audit: classification methodology and dispute outcomes audited by an external party (selected and disclosed by the Verification Council)

Metrics are published at vera.veraregistry.com/transparency with the underlying methodology disclosed.

12Cryptographic integrity

• Every record is hashed (SHA-256) and the hash is anchored to QUBIT Blockchain®
• Anchor transactions are publicly verifiable via QUBIT’s Polygon v3 public layer
• Tampering with a record produces a hash mismatch detectable by anyone
• VERA cannot retroactively alter an anchored record without producing a new version chain entry

13Relationship to standards bodies and C2PA

Content provenance and entity authority are two different questions. Provenance standards answer where a piece of media came from and whether it was machine-generated. Entity authority answers whether the entity behind a claim, a business, a person, a product, should be trusted. VERA is built for the second question and interoperates with the standards that address the first. The two are complementary, not competing.

This distinction is now operational, not academic. As of 2026, major platforms read cryptographic provenance directly: YouTube applies a permanent, unremovable AI-content label when a file carries C2PA metadata confirming it is generative. Provenance metadata is already load-bearing at the point of extraction. VERA is designed so that its verification records are machine-readable by the same pipelines.

VERA interoperates with parallel content-provenance and trust efforts:

• C2PA (Coalition for Content Provenance and Authenticity), VERA reads C2PA manifests where present and issues C2PA-compatible Content Credentials, signed by the AITP™ Cert, that extend the manifest with assertions C2PA does not itself define: VERA entity ID, Trust Score, registry hash, verification level, evidence hash, and QUBIT® anchor. Issued credentials validate against the C2PA toolchain (claim signature, content binding, and trusted timestamp); recognized-issuer status follows trust-list accreditation.
• Adobe Content Credentials, VERA respects existing credentials and supplements them with classification and entity-level trust.
• SynthID and platform watermarking, VERA treats a watermark as one provenance signal among several, never as a substitute for independent verification.
• IETF / IEEE trust-anchor specifications, VERA records align with emerging machine-readable trust signals.

VERA does not replace these standards. It operates as the verification and entity-authority layer above them, and treats C2PA as one serialization target among several, alongside JSON-LD, Schema.org, and the native AITP™ record. A new provenance standard becomes an export target, not an architectural reset.

14The AITP™ Cert

Every C2PA Content Credential VERA issues is signed by a single protocol-level credential: the AITP™ Cert. It is a standard X.509 certificate holding an ES256 (NIST P-256) key, used only at the export boundary, where a finished verification is serialized into a credential the outside world can validate. The internal trust engine remains proprietary; the AITP Cert is the seam where VERA’s verdict meets the open C2PA toolchain.

One signer, one accreditation. A VeraGraph™ entity badge, a V-Verify badge, and a verified content credential are all signed by the same AITP Cert. Because there is one signing identity for the entire ecosystem, recognition of that identity by a C2PA trust list extends trusted-issuer status to every VERA credential at once.

VERA holds the first AITP Cert. The model is federated, in the shape of the public certificate-authority system:

• Every AITP authority holds its own AITP Cert.
• A C2PA trust list registers which AITP Certs are recognized.
• Verifiers check membership. AITP defines the protocol; VERA is the first authority implementing it.

Trust posture, stated plainly. Issued credentials validate today against the C2PA toolchain: claim signature, content binding, and trusted timestamp all verify. Recognized-issuer status, the green check in a strict inspector, follows trust-list accreditation. The signature is real; the trust-list recognition is the remaining step. VERA does not claim trusted-issuer status it has not yet earned.

15Research and methodology

The VERA Verification Standard commits to evidence-based claims, pre-registered measurement protocols, and the publication of null results. The methodology paper that operationalizes this commitment for the VeraGraph™ measurement program is in preparation.

Forthcoming: VeraGraph Measurement Methodology, v1.0. Pre-registered protocols for measuring representation accuracy across AI Recommendation Engines (AIREs), including matched entity pairs, staggered publication schedules, cross-model probing methodology, statistical reporting standards, and the null-results disclosure policy. Available Q3 2026.

Organizations, journalists, analysts, and academic collaborators interested in the methodology paper, in early access to results, or in independent replication of the measurement program are invited to contact press@veraregistry.com. Publication will be announced through VERA News and to the contact list.

The Verification Standard, the lexicon at veraregistry.com/lexicon, and the methodology paper together form the institutional record of what VERA does, in what vocabulary, under what protocol. Architecture covered by U.S. Provisional Patent Application No. 64/091,032.

16Agentic commerce and machine trust

AI search is becoming AI action. Increasingly, an AI agent does not merely tell a person which product to consider; it compares, selects, and transacts on that person’s behalf. The visibility-optimization tools of the AI-search era answer one question, whether a brand is seen by these agents. They do not answer the question that governs a purchase.

This is no longer speculative. OpenAI has shipped Instant Checkout in ChatGPT on the Agentic Commerce Protocol (ACP), an open standard built with Stripe. Google has shipped Universal Cart and agentic checkout on the Universal Commerce Protocol (UCP), an open standard endorsed across the retail ecosystem, alongside a separate Agent Payments Protocol (AP2) for spend authority. Marquee retailers are already live on both. This is the direction of the entire field, not a single company’s bet: the major AI assistants, from ChatGPT and Gemini to Claude and Grok, are all becoming agents that act on a person’s behalf. The rails now exist, they are open, and they are arriving across every assistant at once.

The binding question is trust, not visibility. When an agent is about to spend a real person’s money, ranking well in an AI answer is not sufficient. The agent needs to establish that the merchant is real, that the product is authentic, that the seller can be relied upon to deliver, and that the price and claims are genuine. That is authentication and provenance, not measurement. An agent that transacts with an unverified merchant because it appeared prominently in an AI response is a liability.

These protocols move goods and money. None of them verifies truth. ACP, UCP, and AP2 govern the cart, the checkout, and the payment authorization. Not one establishes that a merchant is genuine or a product authentic; today that gap is papered over by hand-onboarding known brands. As agentic checkout opens to the long tail of sellers, the verification question becomes load-bearing, and it must be answered by a party with no stake in the sale. A platform cannot be the impartial verifier of the merchants transacting on its own rails.

AITP is the trust layer agentic commerce requires. The same infrastructure that verifies entities and content for human decisions serves machine decisions, exposed to agents through the protocol:

• Resolution. An agent resolves a merchant to a verifiable identity, a TrustScore, and current standing before it transacts, the same resolution that AI systems use to answer a person.
• Content credentials. A product image, specification, or claim carries a C2PA credential signed by the AITP Cert, so an agent can confirm it is authentic and unaltered at the source.
• Reputation and correction. The registry gives an agent a place to check standing and correction history, not marketing copy.

Discovery is a race that erodes as every brand games it. Trust is the gate that must exist for agentic commerce to function at all, and gates are held by authorities, not by dashboards. The authority must also be neutral, independent of any platform that profits from the sale, which is why AITP is cross-protocol by design: an agent resolves a merchant and a product against the same independent trust record whether it transacts through ACP, UCP, or a successor. VERA states this application of the protocol plainly and early: visibility tools help an agent find a brand; AITP is how an agent knows it can trust a brand enough to act. This is a forward application of the Standard, built on the same verification model, evidence framework, and cryptographic integrity documented above.

17Versioning

This document is v1.2. Future versions will be released under the same version-controlled, anchored, version-history process applied to all VERA records.

v1.2 (July 2026): added Section 16, Agentic commerce and machine trust, documenting the application of the protocol to AI agents that transact on a user’s behalf.

v1.1 (June 2026): added Section 14, The AITP™ Cert, documenting the protocol-level signing credential, the single-signer model, and the trust-list accreditation posture. Section 13 updated from specifying C2PA-compatible issuance to issuing it.

v1.0: initial publication.

Prior versions (when they exist) are retrievable at vera.veraregistry.com/protocol/versions.

How to Cite This Document