To Figure Out How Humans and AI Should Work Together, We Had to Ask a Simpler Question First
Everyone is asking how humans and AI should work together. We got stuck on that question until we backed up to the two questions a contract lawyer would ask first: who exactly are the parties, and what exactly is the deliverable?
Take the parties first. The usual framing types them as different things — the human, the machine — and then argues about supervision: who watches whom, and when the watching can stop. We made a different opening move. For this problem, treat both parties as the same kind of thing: cognition — things that read, reason, and produce claims. That's a modeling choice, not a discovery; we adopt it because of what it lets you ask. The moment both parties are the same type, "how should they work together" stops being a philosophy debate about humans supervising machines and becomes an engineering question: which jobs in the production of a trustworthy claim are interchangeable between any two reasoners — and which jobs provably are not?
Now the deliverable. The thing the parties produce together is not "output." A work product you can act on has to be two things at once: it has to hang together on its own terms, and it has to be anchored to something outside whoever wrote it. A beautifully argued claim with no anchor is a liability with good grammar. An anchored claim that contradicts itself is noise with citations. Whatever process gets you both — at a cost someone can see and approve — that process is the real deliverable of "working together."
Here is the uncomfortable part. Almost every organization already has an answer to this, and the answer is "reviews." Trust, but verify. The problem is that nearly no one can say what a unit of verify is. Two review processes that look identical on paper can buy you wildly different assurance — and if you can't tell them apart, you are pricing trust by the pound.
That's what this short series is for. Three pieces, in order, each one building on the last:
The first piece shows that "verification" is quietly two different purchases — machine certainty about what a claim's checkable core says, and independent judgment about what it means — and shows, on a claim of our own that we caught mid-draft, why filing those two as one verdict is exactly how confident-wrong output ships.
The second piece is the surprising one. For one worked case, we can prove — machine-checked, with no gaps — that a system's own machinery cannot close every question about its own work. Something external is structurally required. And read that exactly: the proof says nothing about that something being a person. We choose to staff that slot with the human — deliberately, and we keep that choice labelled as a choice, because the proof earns the slot, not the staffing.
The third piece turns all of it into something you can sign: a four-line contract for buying verification honestly — what counts, how often, and what each line costs. If you've ever red-lined an SLA, that piece is for you.
Taken together, the three pieces are the opening move of a discipline we're proposing for managing human-and-AI cognition the way we already manage every other costed service. We call it Cognitive Service Management — and the third piece earns that name properly, lineage and all, back through fifty years of safety engineering that already learned to price reviewer independence the hard way.
Two honesty notes before you start, because the series lives or dies on them. First: what's machine-proved in these pieces is proved for a specific worked case; the claim that the same structure holds everywhere is labelled as exactly that — the part we're asking you to test, not the part we've settled. Second: none of this makes an AI honest or a result true. It prices and provisions the checking. The discipline tells you what you bought; it does not promise the work is good — no honest verification regime ever does.
Read with the labels on. That's the whole method.
Field Effect Institute maps structures that recur across independent domains, tests where they hold and where they break, and verifies what survives with machine-checked proofs. A lens, not a system. Every claim carries its verification status.
Proofs for this series: github.com/field-effect-institute/witness-independence-proofs
Series introduction | Three pieces follow
Next: “Witness Independence: The Two Axes”