Trust is upstream.

Humans trust ahead;
agents trust in reverse.

Trust isn't one thing. Humans extend it before the evidence arrives; agents accrue it backward, from proof. We've spent years trying to make AI "trustworthy" without noticing that, for a machine, trust runs the other way — and that the part no machine can verify is the part that was always the point.

By Chris Huber Reitz · a working thesis, held to its own standard

Two strangers shake hands on a deal before either has proof the other will deliver. Down the hall, two agents negotiate the same deal — and neither moves until the other's record is on the table. Same word, trust, doing two opposite jobs.

We talk about "trustworthy AI" as if trust were a single substance a machine earns the way one person earns another's — slowly, relationally, on the benefit of the doubt. So we pour the work into alignment and assume trust is whatever's left when the system behaves. But that gets the mechanism backward, and it mis-builds the place that matters most: the handoff between a human and a machine.

The one moveTrust has two grounds, not one

Underneath every act of trust is a question — because trust is what you do when you have to act before you can be sure. The thing nobody says out loud is that humans and agents, facing that same uncertainty, ask different questions to shrink it.

A human grants first and adjusts later. An agent grants nothing it can't substantiate, and revises as the proof arrives. One leans into the future; the other reads from the record. Compress it and you get the law:

The turnThe agent economy doesn't abolish human trust — it instruments it

Here's the part almost everyone misses. The machine layer doesn't delete the human kind of trust — it instruments it. Reputation scores, signed receipts, money staked and seized when someone cheats: none of these replace "I trust you'll be fair." They make that feeling provable and priced.

Calibrated authority is reciprocity rendered evidential — the felt thing converted into a number you can check. Not a smaller, colder trust. Trust with the gamble engineered out: fairness you no longer take on faith, because the system can verify it and the market can price it.

The residueWhat neither ground escapes

Now the floor drops out. Verification is not trust — verification is what you do instead of trusting. The whole point of trust, since Luhmann and Baier, is that it's the vulnerability you accept beyond the evidence. So when an agent verifies everything it can, the part left over — the part no receipt covers — is integrity: whether the counterparty will honor the spirit when the letter runs out.

That residue doesn't vanish in the machine economy. It can't be instrumented — the moment it could be, it would stop being integrity and become one more column in the audit trail. It's the irreducible thing both sides still gamble on.

Why it's upstreamGet the ground wrong and the rest can't save you

Every permission you hand a machine — and every permission it hands back — is a trust decision made on a ground you probably haven't named. Treat the two grounds as identical and you mis-design the handoff at its root; no amount of alignment work downstream repairs a join built on the wrong question.

That's the claim in one line: trust isn't what you bolt on after the system works. It's upstream of whether it works at all. It belongs in the architecture, at the altitude where permission moves between humans and machines — the altitude Organizational Intelligence Design is built to occupy.

One honest caveat, and it's the whole method: this is a working thesis, not a verdict. So I hold it to its own standard — calibrated authority, applied to my own claim. What would prove me wrong is below, beside the pre-registered pilot that tests it. Receipts, not assurances.