Start at the Lowest Line: We Signed Our First AI Deploy Today. The Meter Says I Approved It Too Fast.
This week a draft federal AI safety bill made the news for its threshold. As reported, the government could ask a court to halt a model only if it posed a foreseeable risk of more than 50 deaths or serious injuries, or more than $1 billion in damage. That is a halt threshold, not a statement that 49 is acceptable, and the company behind it says liability still applies below it. But it is the highest line there is. Almost every AI decision made today has nothing to do with human life. For all of those, there is a much lower line, it is cheap, and we started using it today: every consequential AI action carries the name of whoever approved exactly that change, on record.
Why the lowest line
A threshold answers one question: when may the state stop a lab. It says nothing about how the millions of ordinary AI decisions get made before anything goes wrong. A deploy, a purchase, a sent message, a published page. Those are made every day by agents and the people who run them, and when one goes wrong, the usual answer to "who decided this?" is a shrug.
There is more rigor on an expense report. An expense report has a named submitter, a named approver, the exact amount, a receipt and a date. Nobody thinks that is heavy. Most AI actions with real-world effects carry none of it.
So the idea is simple. Before an AI does something that matters, a named person approves exactly that thing, and the approval is kept where it can be checked. That is the same principle as a car that makes you keep your hands on the wheel, applied where it actually works: to decisions that can wait a few seconds for a human.
How it works
We built a small tool for this and call it quorum.
The AI proposes. It states the action, what changes for whom, and the exact files being changed. For our deploys, it also names the exact image that will ship.
A person reads and signs. One dialog shows the action and its consequence. The person types their word, which unlocks their signing key. The word never passes through the AI, so the AI cannot sign on their behalf.
The approval covers one exact change. If a file changes after the approval, the approval no longer counts, and the deploy refuses to run. There is no blanket "the AI may act for me."
Higher stakes, more signers. Our tiers go from "just record it" through "one named human" to "three signers, two of them human, one from outside, and any dissent stops it."
Dissent stays on the record. A no is recorded with its reason. It is never overwritten.
The record can be checked. Each entry is chained to the one before it, so an edit or a deletion shows. The key that anchors the whole thing is published in DNS on our own domain, where only the domain owner can change it.
The first record
Today Claude, our AI, proposed a deploy: adding this feature to the Enterprise column of our pricing page. I read it and signed it. The deploy script checked that the approval existed, that the files matched what I approved byte for byte, and that the image was the one named in the proposal. Then it shipped.
The tool also measures how long each signer takes to decide, against the size of the change. It is called the rubber-stamp meter, and its first reading was me: 19 seconds to approve a 3.5 KB change, flagged as fast. The change was about ten lines. Whether 19 seconds is enough to read ten lines properly is exactly the question the meter exists to ask, and it should not have to take my word for the answer. We publish the reading anyway, because a meter that only reports flattering numbers is not a meter. The meter's own first version was wrong in the other direction: it judged me against the whole 72 KB page instead of the ten lines I actually read. We fixed that before writing this.
Before that first signature, automated security reviews flagged 12 problems in the tool. The worst one was mine: the first version let whoever registered first under my name become the root human, which means the AI could have done it. We moved the root of trust to DNS on our domain, which the AI cannot edit. Two of the 12 findings were never described in detail. For those we fixed what we judged the likely problem.
What it does not solve
We would rather say this than have someone else say it for us.
A signature is not understanding. It records who approved, not whether they understood. Small changes, stated risks, independent signers and the meter help. Nothing removes the problem.
Consent can be abused to shift blame. "You clicked approve" must never move liability onto the person who knew least. Our answer is that an approval counts only for the exact change that person was shown, and the vendor's own decisions sit in the same record.
Bad actors will happily sign. A record names misuse. It does not prevent it.
Someone has to decide what counts as consequential. Our default is that actions with outside effects need a signature, never that the AI decides for itself which of its actions matter.
Small approved steps can add up to something nobody approved.
The record is tamper-evident, not tamper-proof. Someone with write access could rewrite it. Published checkpoints and our version history would show it.
None of this replaces a high threshold for catastrophic risk. It is the floor underneath it, and it is a floor anyone can build today for almost nothing.
Where this goes
This is day one of a two-person company using it on itself, so treat the numbers above as a demonstration, not data. Our own deploys go through it from now on, and we will publish what the record shows, including the readings that make us look careless.
It is offered on our Enterprise tier, as software you run yourself: the signing tool, the ledger and the policy file. Today the signing dialog runs on macOS. Planned next: an API for team applications, public verification of the record, and joining each approval to what the AI spent carrying it out. If the lowest line is where you want to start: Enterprise pricing.
How do AI models see YOUR brand?
AIPM has audited 250+ domains. 15 seconds. Free while still in beta.
Was this useful? Thirty seconds, no cookies, no tracking, no third parties, your address hashed and never stored. If the box below does not load, the same question lives at https://analytics.dugganusa.com/nps.html?post=start-at-the-lowest-line-we-signed-our-first-ai-deploy-today-the-meter-says-i-approved-it-too-fast
Comments