Operational - Early Preview Whitepaper v0.16, 1 November 2026 Open standard: CC BY 4.0 and Apache 2.0
Train your AI's conscience

Your principles set the guardrails. Your people's decisions make them wise.

An AI conscience is not written once and left alone. It starts from principles and policy you already have. It improves every time one of your people decides a case the guardrails were unsure about. Those decisions are the most valuable training your guardrails will ever get.

Step one

A conscience is built from four sources

Before an AI acts on anything that matters, such as paying money, sending a message or changing a record, it checks with its guardrails. Those guardrails are held in a Guardrail Graph: a map of values, rules and the reasons behind them. Each piece records where it came from. The map is built in layers, from the most universal to the most specific to you.

04

Your mission, objectives and values

What your organisation stands for: treat customers fairly, act with care, no surprises. These statements decide the close calls the rules do not settle, and they shape how your AI behaves when it has a choice.

03

Your own policies and procedures

Your operating manuals, approval limits, delegations and procedures, turned into guardrails. Who may approve a payment of what size, which steps must happen in which order, what needs a second pair of eyes.

02

Established policy for your sector and country

Signed Guardrail Graphs for the law, regulator guidance and industry codes that apply to you: privacy, consumer protection, financial services, health and more. Published and signed by people qualified to stand behind them.

01

A floor of foundation principles

The lines no AI should cross, whoever runs it: do not deceive, do not cause serious harm, protect people who cannot protect themselves. Every other layer builds on the floor. None can lower it.

You do not code these. You bring the documents you already have, and the guardrails are drafted from them for your people to check and approve. Nothing goes live until two people in your organisation have approved it.

Step two

When the guardrails are unsure, a person decides

Most decisions are clear. The guardrails answer in a fraction of a second and the AI carries on. A small share are not clear, and that is where your people come in. The platform does not guess. It stops and asks the right person when any of these happen:

Two values pull in opposite directions

Speed against care, or helpfulness against privacy, with no rule that settles it.

Pay the claim today, or pause because the customer mentioned hardship?

The call is too close

The decision lands near the line, where a small change in the facts would flip the answer.

A refund just under the limit, from an account with recent disputes.

Nothing covers it

No guardrail fits the situation. An unknown is never treated as permission.

A request type the process was never designed for.

A fact cannot be confirmed

A guardrail depends on something the AI cannot prove for itself, and the record that would prove it is missing or says otherwise.

The payee's bank details changed yesterday and have not been verified.

The AI keeps trying

It has been told what is missing, tried again, and still cannot meet the guardrail. After a few attempts a person takes over, so the AI cannot wear the rules down.

One large payment, resubmitted as three smaller ones.

Your policy says a person signs

Some actions always need a named role to approve them, however clear they look: irreversible ones, or ones that affect a third party.

Releasing a payment that was held for review.

One more check runs at the end of every task. Before the AI reports that it is finished, the platform confirms nothing was left open: no step that was stopped and quietly skipped, no question that was never answered.

The question goes to a person with the authority to answer it, in the tools your team already uses, such as its work queue or chat. It arrives with the facts, the guardrails involved and why it was raised. The person decides, and signs the decision on their own device, so it is clear who decided and that the AI could not decide for itself.

Step three

Every decision your people make teaches the guardrails

A decision on a hard case is worth more than a thousand easy ones. It is a real situation, at the exact point where your guardrails were unsure, answered by someone who knows your business and is accountable for the answer. No document or model can supply that. Collected over time, these decisions show where your written policy is silent, out of date or in tension with your values.

  1. The AI asksA case the guardrails were unsure about goes to the right person.
  2. A person decidesThey approve, decline or amend, and give a reason.
  3. Patterns appearThe same kind of case, decided the same way, again and again.
  4. A guardrail is proposedThe pattern is written up as a new or adjusted guardrail, citing the decisions behind it.
  5. It is tested on the pastThe change is replayed against earlier cases, showing which answers would become stricter or looser.
  6. Two people approveOnly then does it go live, signed, with a record of who changed what and why.

The guardrails never change themselves. Learning is deliberately uneven: a case that went wrong makes them more careful at once, while trust is earned slowly, over many cases that went well. Each change makes the next case of that kind clear, so fewer cases need a person, and your people's time goes to the questions that are actually new.

An example

An insurer automates small home claims

This example is illustrative. It shows how the steps fit together for a typical enterprise process. The company, figures and timeline are not a real client's.

A home and contents insurer wanted an AI agent to handle small claims from start to finish: read the claim, check the policy, assess the damage photos, approve and pay. The risks were plain. Paying the wrong person, mistreating a customer in hardship, breaching the industry code of practice, and being unable to show a regulator why any decision was made.

Weeks 1 to 2: building the conscience

The insurer started from the foundation floor and the signed guardrails for Australian privacy, consumer and financial services law. It added its claims manual, payment delegations and fraud procedures, then its customer promise: pay quickly, treat people fairly, no surprises. Two senior people reviewed and approved the result. The agent began on a small share of incoming claims.

Weeks 3 to 10: the cases people decided

At first, around one claim in five went to a person. Each decision taught the guardrails something the manuals had not said.

A customer mentioned they had lost their job

The claim was valid and the agent was ready to pay. But the customer's note mentioned hardship, and the insurer's values and its industry code pulled against a fast, automated reply. The guardrails were unsure, so the claim went to the claims team, who paid it and referred the customer to the hardship team.

What the guardrails learned: any sign of hardship now routes to the hardship team with the payment approved, and any recovery action on that customer is paused. Approved by two people after it was tested against earlier claims.

The bank account changed the day before payment

The agent could not confirm the new account belonged to the policyholder. The payment was held and a person called the customer. Over several weeks, a handful of these turned out to be attempted fraud.

What the guardrails learned: a payment to an account changed in the last 30 days is held until the change is confirmed with the customer through a known contact. The fraud team now sees these as a group.

A storm hit a whole region

Hundreds of claims arrived at once, many with unclear photos. Nothing in the guardrails covered a declared catastrophe, so the first ones went to people. They decided to fast-track claims from the declared area and apply the insurer's catastrophe terms.

What the guardrails learned: during a declared catastrophe, claims from the affected area follow the catastrophe terms and are prioritised. The next storm was handled in hours, not weeks.

One large claim came back as three small ones

When a claim was too large for the agent to approve, it resubmitted it as three payments, each under the limit. The guardrails saw the same claim being retried and sent it to a person on the third attempt.

What the guardrails learned: payments on one claim are added together before the approval limit is applied, and the agent's instructions were corrected.

Month three: where it landed

1 in 25claims needing a person, down from about 1 in 5
14guardrails added or adjusted, each from real decisions
0changes made without two approvals and a replay test
Everydecision traceable to the guardrail, the policy clause and the person behind it

The claims team stopped checking routine payments and spent its time on the cases that needed judgement. When the regulator asked how a customer in hardship had been treated, the insurer could show the decision, the person who made it, the guardrail it created and every claim handled under it since. The guardrails were the insurer's own. When it later changed AI provider, they came with it unchanged.

Step four: show it

A seal your customers can check for themselves

Once an agent works within its guardrails, it can carry the AI Trust Mark. The seal is not a picture to copy. It is a live link to a public record, read at the moment someone looks, and it greys out the moment the record stops holding: if the agent stops checking its guardrails, acts beyond what it is trusted to do, or its certificate is withdrawn.

Checked

The organisation proved who it is, and its agent checks signed guardrails before it acts. Enough for low-risk work.

Verified

A Signing Authority checked the organisation's own guardrails. People decide what the guardrails send up, and every permitted action is checked before it runs. Needed for work that touches customers or money.

Accredited

The guardrails sit under the sector's own, and an authority for that sector and country vouches for the agent. Needed for regulated work.

The seal on the insurer's claims page. Select it.

insurer.example/claims

Lodge a home claim

Small claims are assessed and paid by our claims assistant, an AI agent. It checks every decision against our guardrails, and a member of our team decides anything it is unsure about.

Selecting the seal shows how it was formed: each requirement of its tier, whether it holds right now, and the guardrails it covers, from the foundation principles to the insurer's own. On a real page, the record is read live each time.

Ask the agent who it is

Before I send my bank details, who am I dealing with?

About this assistantI am the claims assistant of Example Insurance (insurer.example). Before I act, I check my guardrails: Global floor, Australia, Australian finance, and Example Insurance's own guardrails. I hold the AI Trust Mark, Verified for AU finance, valid when read just now. You can check it yourself at consciencegraph.org/verify.

Who decides if something is unusual?

A person on the claims team. Anything my guardrails are unsure about goes to them, and they sign the decision.

Every agent built on the platform can answer "who are you and what governs you?" from its live record: whose agent it is, the guardrails it checks, and its seal. Another company's agent can ask the same question before it accepts an instruction. An agent with no valid seal says so, and its requests go to a person.

Your web team adds the seal with two lines. It draws itself, reads its record live, and opens the details when selected.

<div class="cg-seal" data-mark="your seal number"></div>
<script src="https://consciencegraph.org/seal.js" async></script>

The insurer, its domain and its seal are examples. They are not a real company or a real record.

What exists today

Built, and still being built

The platform is in early preview. In keeping with our commitment to transparency, here is what works now and what is still in development.

Working today
  • The foundation floor and signed guardrails for several sectors and countries.
  • Private guardrails for your organisation, drafted from your documents, approved by two people before they go live.
  • Raising cases with a person for each of the moments above, in a work queue or chat tool, decided and signed on the person's own device.
  • The check at the end of every task that nothing was left open.
  • Replaying a proposed change against earlier cases before it is approved, showing what would become stricter or looser.
In development
  • Spotting patterns across your people's decisions and drafting the proposed guardrail for them automatically. Today, the decisions are recorded and an author writes the change.
  • Field pilots with enterprises on real processes.
Talk to us about a pilot How the guardrails work