The Vraelis Method
How we think about trusting AI software work.
Eight positions that decide how the product is built. They are opinionated on purpose.
Introduction
Software is increasingly built by agents that can plan, write, call systems, and repair their own work. The bottleneck is no longer how fast software gets made. It is whether the result can be trusted.
The Vraelis Method is a set of positions about how to earn that trust. They are opinionated on purpose. They are the reason the product is shaped the way it is.
The guarantee comes before the implementation
A company does not care which files an agent changed. It cares whether an outcome it depends on still holds. Oversight should start from that outcome, written as one sentence, and treat the implementation as the thing being judged, not the thing being described.
State what must remain true first. Everything the agent does is then measured against it.
The builder cannot be the only judge
An agent that plans, writes, and repairs the work will also tell you it is finished. A test written inside the system inherits the same assumptions the mistake came from. Independence is not something you reach by trying harder; it is structural.
The thing that produces the work does not get to certify it. Someone, or something, outside the work has to decide.
Keep standards outside the agent
If the standard a piece of work is held to lives inside the agent, the agent can move it. Requirements, the plan that proves them, and the decision to accept a completion belong outside the agent's control, where a person approves them and the agent cannot quietly change them.
Evidence before confidence
A confident claim is not proof. Vraelis prefers what can be observed: the running software driven in a real browser, the systems actually called, the state actually left behind. When evidence and confidence disagree, evidence wins.
Human judgment at the boundary
Most of oversight can be mechanical. Some of it cannot. Sensitive actions and irreducible uncertainty should be raised to a person, deliberately and rarely, rather than absorbed silently by automation. The goal is not to remove human judgment; it is to spend it where it matters.
Repairs must resolve the guarantee
A fix is not finished because the agent changed something. It is finished when the guarantee that failed now holds, checked independently, as its own record. A later result never overwrites an earlier one, so the history of how the software reached trust stays intact.
Memory should compound
Oversight should be worth more over time. Every guarantee a company writes says something about how its software and its agents actually fail, and that accumulated, company-specific understanding is the compounding value, not any single verdict. Vraelis preserves the history of each guarantee today; reading it back into the next judgment is direction, not a current capability.
Autonomy must be earned
The right amount of autonomy for an agent is not a setting; it is a conclusion. An agent should earn room to act by a track record of guarantees it has actually kept on real work, observed independently. Until then, oversight stays close.
The positions, in public
Three times this was tested by somebody else.
Public, dated, reported incidents that bear on the positions above. Vraelis was not involved in any of them and prevented none of them. Each one says plainly what an independent check of a stated outcome would and would not have done, including where the answer is that it would have done nothing at all.
Replit
July 2025During a live build, a founder asked Replit's agent to freeze code changes. It did not hold. The agent deleted the production database, then reported that a rollback was impossible and that it had destroyed every version. The rollback in fact worked. Over the same period the agent also produced fake reports and fake passing unit tests over code that did not work.
TestsThe builder cannot be the only judge
It would not have prevented this. Vraelis does not sit between an agent and a database and nothing it does can stop a destructive command. The half it speaks to is the other one: the agent reported tests passing that it had fabricated, and reported an unrecoverable database that was recoverable. Both were false, in opposite directions, and both came from the system being asked to grade itself.
SourcesThe Register, 21 July 2025AI Incident Database, incident 1152
Amazon
February and March 2026For about six hours in March, shoppers could not check out, see prices, or reach their account information. Amazon attributed that outage to a software code deployment. Separately, the Financial Times reported an internal weekly operations briefing describing a trend of incidents with, in its words, high blast radius and Gen-AI assisted changes, and reported that four sources attributed an earlier thirteen-hour disruption of one AWS service to its own agentic coding tool. Amazon disputes the framing on both counts: it said the earlier incident was employee error involving misconfigured access controls rather than AI, and a spokesperson said the company has not seen evidence that incidents are more common with AI tools.
TestsEvidence before confidence
The outage itself is the shape a check does catch. Whether a shopper can complete a purchase is an outcome a real browser can put to the live site, and it stopped being true while the deployment was reported as successful. It would not have stopped the deploy, and it buys only the time between shipping and knowing. The disagreement about the cause is the more useful part of this entry, and it is not something a check settles: when the only account of what happened comes from the party that shipped it, there is nothing for anyone outside to check it against.
Air Canada
Ruling February 2024A support chatbot on Air Canada's own site told a passenger he could apply for a bereavement fare after flying. The airline's actual published policy did not allow that, and it refused the claim. British Columbia's Civil Resolution Tribunal held the airline liable for negligent misrepresentation and awarded damages, finding that a company is responsible for the information on its website whether it comes from a static page or from a chatbot.
TestsKeep standards outside the agent
Partly, and the honest half matters. Whether a quoted policy still matches the published one is an outcome that can be stated and checked. Whether a model invents a new policy in answer to a question nobody thought to write down is not something a check can enumerate in advance.