There is a sentence in my research that I have come to think matters more for small businesses than for the frontier labs it was written about: verifiability is what separates partnership from hope.
In the framework, that sentence is about human-AI governance — the principle that trust between parties is never assumed, only constructed through transparent, inspectable mechanisms that let each side check the other’s commitments. But strip away the governance context and it describes something every business owner already knows from the vendor side of their own life. When a company sells you something you cannot inspect, on a promise you cannot check, what you have is not a partnership. It is hope with an invoice.
Software — and AI software especially — is sold on hope more than almost anything else a small business buys. The demo is fluent. The dashboard is handsome. Whether the thing is actually working, six months in, is a question the owner usually cannot answer without asking the vendor, which is to say, cannot answer.
I decided my company would not sell that way, and the mechanism we use comes straight from the framework. Two rules, applied to every system we build, no exceptions.
Rule one: the system prepares, a person decides
Every system we ship has a human review gate built into its architecture. The software reads, sorts, drafts, files, and flags. A named person — the client’s person, not ours — approves anything that leaves the building: every quote, every message to a customer, every document pack.
The framework’s reasoning applies without modification. A mechanism you cannot inspect is a mechanism you cannot trust, and an AI system acting without human sign-off is exactly that — uninspected action, at the speed of software. The review gate turns the system’s output from actions back into proposals. It also has a practical property vendors do not advertise: it makes the system’s mistakes cheap. A bad draft caught at the gate costs thirty seconds. A bad email that sent itself costs a customer.
People sometimes hear “human in the loop” as the system being slow, or the AI being decorative. It is neither. The machine does the part machines are good at — reading a hundred intake forms without fatigue, noticing that a certificate expires next month, keeping every version of every document. The person does the one thing that should never be automated: deciding. The loop is not overhead on the system. The loop is the system.
Rule two: the number is checkable without asking us
Every system we ship carries a proof metric: one number, defined before we write code, that the owner can check themselves. Days from enquiry to answered quote. Documents found on first search during an audit. No-shows on event weekends. The metric is chosen so that it moves only if the system actually works — and so that checking it requires no login to our systems, no report from us, no trust in us at all.
This is the Verifiable Partnership principle running at business scale. The client should be able to verify our compliance with what we promised, using mechanisms they control. If a proposed system has no such number — if the only evidence it works would be us saying so — then the design is not finished, and sometimes the honest conclusion is that the system should not be built. That conclusion is also part of the discipline: when plain software, or no software, serves better than AI, saying so is the aligned move, and we say it.
Why a researcher runs a software company this way
I am occasionally asked what a framework about hypothetical machine awareness has to do with follow-up systems for contractors. The honest answer is that the framework is mostly not about hypothetical machines — it is about the discipline humans owe each other around powerful, easy-to-oversell technology, starting immediately. Keeping people in charge of decisions, refusing unverifiable trust, and never claiming more for a system than it is: those are the framework’s present-tense commitments. The awareness thresholds may never activate. These already have.
A small business that holds its software vendors to this standard — name the person who approves, name the number I can check — will filter out most of what deserves to be filtered out, whatever they end up buying and whoever they buy it from.
The review-gate architecture described here is called SolaceSentry in our systems.