Technical Specification: implementing verifiable AI supervision

Most AI governance frameworks define principles. Few define how those principles are implemented at a technical level.

The </AI> Protocol introduces a structured approach to transform human supervision into verifiable, auditable, and interoperable evidence. This page outlines the core technical components required to implement the protocol in real-world systems, not as a theoretical model, but as a practical architecture built on existing, proven technologies.

Why a technical specification is necessary

Governance principles without technical implementation are declarations. They may satisfy policy requirements on paper, but they cannot produce the evidentiary artifacts that auditors, regulators, and courts actually examine.

The </AI> Protocol is built on the premise that human oversight must be technically structured to be verifiable. This means defining not only what should happen, a human reviews an AI output – but precisely how that event is captured, preserved, and made available for independent verification.

This page defines those technical components.

Core components of the protocol

The implementation of the </AI> Protocol is based on four fundamental components that work together to produce verifiable evidence of human supervision.

• Identity anchoring: each supervisor must be linked to a verified, time-stable identity baseline through DAPI. Without this, oversight cannot be reliably attributed to a specific accountable person.
• Content integrity: each reviewed AI output must be preserved through cryptographic hashing (SHA-256) through systems such as ContentProtector.eu. Without this, the reviewed content can be altered after the fact, rendering the oversight record unreliable.
• Qualified timestamping: each event must be sealed with a timestamp that certifies the exact moment supervision occurred. Without this, the temporal dimension of oversight – critical in legal and compliance contexts – cannot be established.
• Public verification: each event must be traceable through a public registry and verification interface. Without this, oversight remains an internal claim rather than an externally verifiable fact.

Each component is necessary. The absence of any one of them weakens the evidentiary chain.

The Human Oversight Event structure (HOE)

The Human Oversight Event is the minimum unit of verifiable supervision. It is the evidentiary record that links a specific human identity to a specific AI output at a specific moment in time.

Each event is structured as a standardized data object. The reference implementation below shows the core fields required for a complete and verifiable oversight record.

{
  "event": "HumanOversightEvent",
  "protocolVersion": "CWC-AI-SUP-1.0",
  "supervisorID": "DAPI-EU-XXXXX",
  "contentHash": "sha256:7d55a1...",
  "timestamp": "2026-03-21T10:00:00Z",
  "decision": "approved",
  "governancePolicy": "internal-ai-policy-v1",
  "verificationURL": "https://www.certifywebcontent.com/supervised-ai/registry/"
}

Each field serves a specific evidentiary function:

• event – identifies the record type for interoperability and automated parsing
• protocolVersion – links the record to a specific, archived version of the protocol specification
• supervisorID – the DAPI-verified identity of the responsible human supervisor
• contentHash – the SHA-256 fingerprint of the reviewed AI output at the moment of review
• timestamp – ISO 8601 UTC timestamp of the oversight event, qualified and tamper-evident
• decision – the outcome of the human review: approved, modified, escalated, or rejected
• governancePolicy – reference to the internal policy or governance rule under which the review was conducted
• verificationURL – public registry endpoint where the event can be independently verified

Hashing and content integrity

Content integrity is the technical foundation of the oversight record. Without it, any claim of supervision is vulnerable to challenge: how can it be proven that the content reviewed is the same content now under examination?

Each reviewed AI output is transformed into a cryptographic fingerprint using SHA-256. This algorithm produces a unique 256-bit hash value for a given input. Any modification to the original content – even a single character – produces a completely different hash, making alteration immediately detectable.

This ensures that:

• the reviewed content cannot be altered after the oversight event without detection
• the exact version reviewed can always be reconstructed and compared against the hash on record
• evidence remains consistent and verifiable across systems, jurisdictions, and time

SHA-256 is a widely adopted standard, recognized in digital forensics and legal contexts across multiple jurisdictions. Its use ensures that the evidentiary record produced by the protocol is interoperable with existing legal and technical verification infrastructures.

Content integrity preservation is handled operationally through ContentProtector.eu, which manages hashing, timestamping, and evidentiary archiving for AI outputs enrolled in the oversight workflow.

Identity anchoring through DAPI

An oversight event is only as strong as the identity behind it. If the supervisor cannot be reliably identified, the evidentiary value of the entire record is compromised.

DAPI – Digital Authenticated Person Identity – establishes a certified baseline identity for the designated Human Supervisor. This identity baseline is:

• certified at a specific point in time, creating a stable and auditable reference
• verifiable independently, without requiring any confirmation from the organization
• persistent as a structured record that remains defensible over time
• linked to every subsequent oversight event performed by that supervisor

Without identity anchoring, oversight events may record that someone reviewed an output, but cannot prove who. In legal and regulatory contexts, anonymous or unverified supervision is treated as equivalent to no supervision.

Identity anchoring through DAPI is therefore not an optional enhancement. It is the prerequisite for the entire evidentiary chain.

Reference: dapi-certification.com

Qualified timestamping and event sealing

The temporal dimension of an oversight event is critical. Disputes and audits frequently turn on the question of when a decision was made – whether the relevant information was available, whether the applicable policy was in force, whether the review preceded or followed a known event.

Each Human Oversight Event is sealed with a qualified timestamp that certifies the exact moment at which supervision occurred. This timestamp is tamper-evident: any attempt to modify the event after sealing will invalidate the timestamp, making the alteration detectable.

Qualified timestamping allows:

• accurate reconstruction of decision timelines in post-incident analysis
• validation that the oversight event occurred within the applicable governance period
• temporal defensibility in regulatory, administrative, and judicial contexts
• sequencing of oversight events to demonstrate process compliance over time

Under eIDAS and equivalent frameworks, qualified timestamps carry specific legal weight in European jurisdictions, making them appropriate for compliance contexts governed by EU law.

Public verification layer and registry

Internal records, however well structured, remain internal claims. To achieve external verifiability, oversight events must be referenceable through a public mechanism.

The CWC public registry provides this function. Each oversight event associated with a domain or content unit can be assigned a CWC verification code. This code is listed in the public registry and enables any third party to verify:

• that an oversight event exists for the referenced content or domain
• that the event has not been altered since registration
• that it corresponds to the specified decision context and supervision activity
• the current status of the record: Active, Corrected, or Retracted

This transforms oversight from an internal process into an externally verifiable fact – auditable by regulators, reviewable by legal teams, and accessible to any party with a legitimate verification interest.

In the public registry, each supervision event is represented through a minimal set of standardized metadata, designed to ensure immediate understanding and verifiability without exposing internal details. In particular, the event classification (HOE Class) and the associated evidence level (HOE Level) allow third parties to interpret the nature and evidentiary strength of the event without accessing the full technical documentation.

Registry verification is available at: certifywebcontent.com/supervised-ai/registry/

Schema.org structured metadata

For organizations publishing AI-assisted content, the protocol supports structured metadata markup that enables automated systems – including search engines, AI crawlers, and compliance scanners – to detect and parse the oversight declaration.

The following JSON-LD is a recommended implementation example for a supervised content page::

<script type="application/ld+json">
{
  "@context": "https://schema.org",
  "@type": "CreativeWork",
  "creator": {
    "@type": "Person",
    "name": "Supervisor Full Name"
  },
  "identifier": "CWC-20260321-0001",
  "publisher": "CertifyWebContent",
  "accountablePerson": "Supervisor Full Name",
  "isBasedOn": "Human supervised AI assistance",
  "url": "https://www.yoursite.com/your-page/"
}
</script>

This markup makes the oversight declaration machine-readable and integrates with existing web standards without requiring proprietary infrastructure.

Integration into existing systems

The protocol is designed to be interoperable with existing technical and governance infrastructures. It does not require replacement of current logging, compliance, or workflow systems.

The Human Oversight Event acts as an additional evidentiary layer that can be integrated into:

• AI pipelines: oversight events can be generated at defined checkpoints within automated workflows, capturing the human review step as a structured record
• Decision support systems: review and approval actions within existing tools can be mapped to oversight event fields, producing compliant records without workflow disruption
• Compliance workflows: existing audit trails can be extended with oversight event records, adding identity anchoring and integrity verification to previously unstructured review logs
• Document management systems: AI-assisted documents can be enrolled in ContentProtector.eu at the point of human approval, automatically generating hash and timestamp records
• Legal and regulatory reporting: oversight event records can be packaged as structured evidence submissions for regulatory review, audit response, or legal proceedings

Integration does not require a complete infrastructure overhaul. In most cases, existing review steps can be formalized into oversight events by adding the four core components – identity, hash, timestamp, and registry reference – to processes that already exist.

The evidentiary chain in practice

The four components described above, identity anchoring, content integrity, qualified timestamping, and public verification – form a chain. Each link depends on the one before it.

• Without a verified identity, the oversight event has no accountable actor
• Without a content hash, the reviewed output cannot be proven to be unchanged
• Without a qualified timestamp, the temporal dimension of the event is unverifiable
• Without a public registry reference, the event cannot be independently confirmed

When all four components are present and correctly implemented, the resulting record is:

• Attributable: linked to a specific, verified human identity
• Immutable: any alteration of the supervised content is detectable
• Temporally anchored: the moment of oversight is certified and tamper-evident
• Externally verifiable: any third party can confirm the record independently

This is the standard of evidence that audit, regulatory, and legal contexts require. Not a policy statement. Not a log entry. A structured evidentiary record that can be examined, challenged, and verified.

From theory to implementation

Every component described in this specification uses existing, proven technology. SHA-256 is a standard cryptographic algorithm. Qualified timestamping is defined under eIDAS. Identity verification systems are widely deployed. Public registries are technically straightforward.

The contribution of the </AI> Protocol is not the invention of new technology. It is the definition of how existing technologies should be structured, combined, and applied specifically to produce verifiable evidence of human supervision over AI systems.

This structuring – the specific combination of identity, integrity, time, and public verifiability – is what transforms a collection of technical components into a defensible evidentiary architecture.

The value of the protocol lies not in any single component, but in how they are bound together to produce a record that can be verified, challenged, and relied upon.

The missing layer in AI governance

Most systems can log decisions. Some can store context. Very few can produce verifiable evidence that a specific human, with a verified identity, reviewed a specific output, under specific conditions, at a specific moment in time, and that neither the output nor the record has been altered since.

This is the gap that the technical specification addresses.

Not just recording what happened. Making it possible to prove, to any external party, that human supervision was real, substantive, and documented in a form that survives scrutiny.

If AI governance is to move from declared compliance to demonstrable compliance, this is the technical layer that makes it possible.

#bemorehuman

In this context, being “more human” does not mean adding more people to the process.

It means making human actions visible, structured, and verifiable. It means ensuring that every moment of genuine human oversight leaves a technical trace that can be examined, challenged or relied upon.

It means turning supervision into evidence.

If supervision cannot be technically verified, it cannot be relied upon as evidence.

Further reading

To understand the supervisory event this specification implements, see the Human Oversight Event page. To explore how supervised events are transformed into structured decision records, see the Decision Attestation Layer. For the operational role responsible for maintaining this structure, see the AI Evidence Officer page. For the public protocol specification, see the CWC-AI-SUP-1.0 specification.