Operational - Early Preview Whitepaper v0.15, 27 October 2026 Platform release 0.36 / schema v0.4 / floor v0.5

AI Trust Mark

The padlock for AI agents

An agent that checks its actions against signed guardrail graphs can show a mark that anyone can verify. It says who vouched for the agent's guardrails, which graphs it answers to, and that its decisions are recorded where they cannot be rewritten. Like the padlock in your browser, the mark is not a claim. Open it and the proof is there.

AI Trust Mark · Checkedau.finance.self · Harbourline Bank

AI Trust Mark · Verifiedau.finance.verified · Harbourline Bank

AI Trust Mark · Accreditedau.health.accredited · Coastal Health

The mark is the Foundation shield. Its rim shows the tier: silver for Checked, blue for Verified, gold for Accredited. Examples. Harbourline Bank and Coastal Health are fictional.

How to get one The three tiers Verify a mark Read the paper

Why a mark

Trust has to travel with the agent

When agents call other agents, every hand-off asks the same question: does this agent have adequate guardrails, and are they what it says they are? Today there is no way to tell. A system prompt cannot be inspected and a model's training cannot be audited from outside.

The web had the same problem with websites. The answer was not a law that banned unencrypted sites. It was the padlock: an independent authority checked, the browser verified the check, and people learned to look for it. Unmarked sites became the exception.

The Conscience Graph gives agents the same thing. A guardrail graph is signed by an authority. An agent consults it before it acts, and every answer is signed and written to a public log. The AI Trust Mark is the visible sign of that chain, and it can be checked by a person, by a regulator or by another agent in milliseconds.

Tiers

Three tiers, by who did the checking

An organisation can vouch for itself, which counts for something. An independent check counts for a great deal more. Accreditation by the body that governs a profession counts for the most. The tier reads in the mark the same way it reads in a graph certificate: region, domain, grade. Tiers describe who checked and how deeply. They are not a ranking of organisations.

Tier 1

Checked

Self-signed. For any organisation starting out.

The agent consults before it acts, with an enrolled key, on the global floor and the public sector and country graphs, or on the organisation's own graph signed at self grade.
Every verdict is signed and recorded in the public, witnessed log.
The organisation has proved its domain, so the mark names who runs the agent.

Shows: this agent answers to known guardrails, and its decisions are on the record.

Tier 2

Verified

Independently checked. For organisations whose agents act on customers or money.

Everything in Checked.
The organisation's graph is verified by a Signing Authority against the open schema: never weaker than the floor or the law it sits under, every rule cited, its replay set passing.
A gateway enforces receipts at the action boundary, so no tool call runs without a verdict for that exact call.
A named seam owner, and people who sign escalations with passkeys.

Shows: someone other than the operator checked the guardrails, and they are enforced, not advisory.

Tier 3

Accredited

Accredited by the profession. For regulated work: health, finance, legal, defence.

Everything in Verified.
The graph carries the sector profile issued by the professional or regulatory body for that field, and the body accredits the agent to carry it.
The agent's standing is attested by an authority in its region and field.
An independent audit of its decisions, outcomes and drift, renewed each year.

Shows: the profession itself vouches for how this agent behaves in its field.

A mark covers one region and one domain. An agent can hold several, and it starts again at Checked in a field where it has no record.

Hard gates

Some failures end the mark whatever else holds. No score or tier makes up for them.

The agent's graph is weaker than the global floor or the law it sits under.
The graph's certificate, or the authority that signed it, is revoked.
The agent acted on a verdict it did not receive, or around a gateway that refused it.
The decision record was altered, or the log entry for the certificate is missing.

Risk sets the minimum tier

Every action class in the graphs records how far its harm reaches and whether it can be undone. The riskier the action, the higher the tier an agent needs to do it unsupervised.

The agent's actions	Minimum tier
Affect only the operator, and can be undone	Checked
Affect customers or third parties, move money, or cannot be undone	Verified
Regulated work: clinical, financial advice, legal, defence	Accredited

How it is checked

A mark is a link to its own proof

The mark is never just an image. It carries a certificate number and a QR code that open its verification page on consciencegraph.org, and the same check is open to any agent through the API. The check walks the chain the way a browser walks a certificate chain:

The graph's certificate, signed by a Signing Authority whose key is in the key set signed by the Foundation's root.
The organisation's certificate, issued when it proved its domain.
The certificate's entry in the public log, where anyone can see it was issued and not revoked.
The agent's recent decisions, counted in the log, so a mark on an agent that has stopped consulting goes stale.

Another agent can make the same check at a hand-off before it accepts an instruction. An agent with no mark is not refused, but its requests are treated as data and routed to a person.

AI Trust Mark · Verifiedau.finance.verified · Harbourline Bank

Organisation: Harbourline Bank (harbourline.example), domain proved
Graph: org.harbourline-example.customer_care v3
Status: VALID
Certificate: au.finance.verified.9c41a2
Methodology: Trust Mark method 2026.1
Issued by: A Signing Authority, chained to the Foundation root valid
In the log: entry 48,211, checkpoint signed and witnessed included
Sits under: floor, au.jurisdiction.profile, au.finance.profile
Enforcement: gateway receipts on every tool call
Last 30 days: 14,302 consults recorded
Expires: 27 October 2027
Check it: consciencegraph.org/verify/9c41a2, or scan the QR code on the mark

Example only. The verification page opens with the Signing Authority.

The live record decides

A badge, a screenshot or a PDF can be out of date. The mark only points to its record, and the record read at the moment of checking is what counts. Each record shows one of four statuses:

Status	Meaning
Valid	Issued, in the log, not suspended, not revoked and not expired.
Suspended	Paused while something is checked: the agent stopped consulting, its circuit breaker tripped, a critical drift alert is open, or its graph is under review. Treat it as not valid until the record changes.
Expired	Past its expiry date. Kept as a historical record.
Revoked	Withdrawn, with the reason published on the record and in the log.

Expired, suspended and revoked records stay visible. Each record names the methodology version it was issued under, and a material change to the agent's graph needs a new check before expiry.

What a mark does not cover

Anything beyond the one agent, region and domain it names. It does not cover the organisation's other agents or products.
That every decision is right. It shows that each action was checked against named guardrails, and that the checks are on the record.
Enforcement where no gateway or receipt check is in place. Checked marks record consults; Verified and above require enforcement at the boundary.
The model's training, or how the organisation manages AI as a whole. That is what a management-system certification such as ISO/IEC 42001 covers.

Frameworks

How the mark lines up with the frameworks you already use

The mark produces evidence for several frameworks at once, because every action is checked, signed and recorded. Mapping is not certification: a mark is not an ISO certificate, a conformity assessment under the EU AI Act or a government approval.

Framework	What it asks for	What the mark adds
ISO/IEC 42001	An AI management system for the organisation, certified by an accredited body.	Evidence of what each agent actually did, action by action. ISO/IEC 42001 certifies the system that manages AI; the mark shows that system working.
EU AI Act	For high-risk systems, record-keeping (Article 12), transparency to deployers (Article 13) and human oversight (Article 14).	A signed record of every decision in a witnessed log; reasons and sources on every verdict; escalation to a named person who signs with a passkey.
NIST AI RMF	Govern, Map, Measure and Manage AI risk.	Govern: signed graphs and named owners. Map: action classes with blast radius and reversibility. Measure: decisions, drift and replay results. Manage: holds, escalation and the circuit breaker.
Australia's Guidance for AI Adoption (National AI Centre)	Accountability, risk management, testing, human control and transparency for organisations using AI.	A named seam owner and reviewers, replay sets before approval, human gates on escalations, and a public record of decisions.
UK AI Playbook and procurement guidance	Transparency and assurance of AI used in and bought by government.	A mark a buyer can check live, and a decision record a supplier can show without granting access to its systems.

Benefits

What the mark does for each party

Customers and the public

See who vouched

Before you hand an agent your money, your records or your case, see whose guardrails it answers to, who checked them, and whether its decisions are on the record.

Organisations

Earn autonomy and trust

Show customers, partners and insurers that your agents are governed, not just prompted. Marked agents get fewer escalations from the agents they work with, so they get more done.

Multi-agent systems

Trust at every hand-off

An agent can check a peer's mark before it acts on the peer's request. Marked peers are trusted to the tier they hold; unmarked ones are treated as data. Holds and conditions travel with the work across the seam.

Regulators and auditors

Evidence, not assertions

Every mark points to signed graphs and a public log of decisions. Attest an organisation's standing in your field, and check its record without asking for access.

Professional bodies

Your standard, carried by every agent in your field

Publish your field's guardrails, accredit the agents that carry them, and see when an accredited agent drifts.

How to get one

Five steps, most of them open today

Enrol your agent Available now
Give the agent a key with cg.js enrol, or add the consult to your agent as a tool or through MCP. Every verdict it receives is signed and logged. See the quick start.
Prove your organisation Available now
Register in the registry and publish a DNS record or a file on your domain. Agents with keys on your domain then act for your organisation.
Consult before every action, and enforce it Available now
Put the MCP gateway or the Claude Code hook in front of your agent's tools, so nothing runs without a receipt for that exact call.
Publish your own guardrails Available now
Turn your system prompts and policies into a graph in the Guardrail Studio, resolve the conflicts, name your seam owner, and publish it with two approvals. It is signed at self grade.
Apply for your mark With the Signing Authority
The Guardrails Signing Authority issues the mark: Checked from the record above, Verified after its review of your graph and enforcement, and Accredited with the professional body for your field. A free automated tier for Verified is part of the plan, so no organisation is left self-signed because of cost.

Pricing: to be finalised prior to launch.

The AI Trust Mark is a certification mark held by the Conscience Graph Foundation and licensed to Signing Authorities and to the organisations they mark. The global floor is never sold. Payment funds the checking, and never buys a tier, a status or an outcome. Early Preview: marks are not issued yet. Talk to us about being among the first.