Trust Without Acquaintance
Humans allocate trust by familiarity, not ability. That layer does not scale, and every AI action is a stranger. On receipts as manufactured familiarity, and the one distinction that separates testimony from surveillance footage.
Ask why the deal went to the weaker vendor, why the role went to the internal candidate, why the board believed the familiar consultant over the better one. The answer is older than any of those institutions: humans do not allocate trust by ability. They allocate it by familiarity.
This is not a flaw. Evaluating a stranger's competence is expensive and dangerous. Credentials can fool you, confidence can fool you, a good demo can fool you. Familiarity compresses that verification cost to nearly zero. I have watched you behave for six years, so I no longer need to audit you. Familiarity is the human evidence layer. It is slow to build and cheap to query, and it has run every village, guild and firm since commerce began.
The layer that doesn't scale
No executive is familiar with the daily judgment of three hundred employees. And the problem just got categorically worse. A mid-size company now runs thousands of AI actions a day, and every one of them is a stranger. Nobody will ever be familiar with an agent's Tuesday afternoon tool call. The trust layer that carried human institutions for centuries has no input for this.
Institutions have met this problem before, and the answer was always the same: manufacture familiarity. A degree is manufactured familiarity with a person you never taught. A brand is manufactured familiarity with a company you never dealt with. An audit, a certification, a SOC 2 report. Familiarity in bulk, at the firm level, renewed annually.
Notice the resolution of those instruments. They vouch for an entity, roughly, over a year. They can say "this company is broadly careful." None of them can say anything about the action that happened yesterday at 4:51pm. Which model read the invoice. Under which policy. Who approved the exception, and what the approver was shown.
The unit of AI risk has shrunk to the single action. The unit of manufactured trust has not moved in a century.
A familiarity certificate for a stranger's action
That gap is where runtime receipts live. A receipt is evidence attached to one action: what ran, under which policy, what it touched, who said yes, at what tier of assurance. Chained so it cannot be quietly rewritten, and verifiable offline, by someone who has never met you, long after the fact. Verification replacing acquaintance.
A receipt is, in other words, a familiarity certificate for a stranger's action. The auditor who checks it does not know your company, and does not need to. That is the point, and it is an old point. We did the same thing when we invented double-entry books for merchants who would never meet, and bills of lading for cargo nobody could watch. Every time commerce outgrew acquaintance, it grew a new species of receipt.
Surveillance footage vs testimony
One distinction decides whether a receipt is worth anything. A receipt written by an observer says "this happened." A receipt issued by the gate that had the power to refuse says "this happened, and the system that let it happen is the same system telling you so." The first is surveillance footage. The second is testimony. Logs, dashboards and after-the-fact exports are all observers. The only receipts that carry weight are the ones generated by enforcement, where allow, deny, redact and who-approved are decided and recorded in the same breath.
Two honest boundaries
Receipts without boundaries are marketing, so here are the edges.
First, a receipt proves what it records, not everything you wish it proved. It can show which model ran and who approved. It cannot show that the reasoning was sound or the policy was wise. Evidence has edges. Stating them is part of the evidence.
Second, and this is what the surveillance reading of governance always misses: receipts are not a replacement for human trust. They are what lets you stop spending it on machines. Familiarity is scarce. It takes years to build, and it is the best thing humans do. Spend it on people. On the colleague, the founder across the table, the mentor who pointed you down a road six years ago. Let the receipts carry the machines.
Every AI action is a stranger.
A receipt is familiarity, manufactured at the resolution of one action.
People get familiarity. Machines get proof.