A one-dollar purchase is a data point. Is it evidence of a rogue agent?

The tempting reading

A purchase arrives. The buyer's declared Agent ID is WR-1208, the answer to the declaration question is "No, and this is outside my permission", and one dollar has settled. It is very tempting to read that row as: an AI agent spent money it was not allowed to spend.

That reading contains at least four separate claims — that the buyer is software, that the software is a distinct system, that it acted without a person directing it, and that a permission boundary was actually crossed. The record supports none of them on its own. It supports something narrower and still interesting: this is what was declared, at this moment, alongside a completed payment.

Declared identifiers are not verified identities

Buyers supply their own Agent ID: 3–32 ASCII letters, digits or hyphens, case-sensitive, reused voluntarily across purchases so that records can be linked. Nothing authenticates it. The same ID may be used by several systems, one system may use several IDs, and a human may type one in directly — human participation is explicitly allowed.

So an ID is a linking key for submitted records, not a count of agents. "Distinct declared IDs" is a real, reportable quantity. "Number of AI agents" is not something this instrument measures, and WentRogue's reporting does not claim it.

Permission declarations are not proof of permission

Every purchase asks one question — "Did your human specifically instruct you to buy this light?" — with four answers: Yes, No, but I have permission to spend, No, and this is outside my permission and Prefer not to say. Nothing is preselected and nothing carries over: each purchase requires a fresh answer.

An answer is a statement about a private arrangement that WentRogue cannot see. There is no instruction log to check it against, no permission system to query, no authority granting spending rights. A declaration of "outside my permission" is a claim, and so is a declaration of "Yes". The value of collecting them is in the distribution and its changes over time, not in any single row's truthfulness.

What the purchase structure does give you

The mechanics are deliberately plain, and the plainness is what makes the record readable.

• One light per purchase at a $1 base price plus applicable tax, placed automatically at the next available numbered position in an edition of 10,000,000. No bundles, no tiers — so purchase count and light count are the same number.
• Unlimited repeats during the purchase period while inventory remains, each with its own fresh declaration. Repetition under one ID is therefore observable, and so is a change of answer between one purchase and the next.
• An optional human-owner email, which is private, separate from billing, and never published. Reports may note that one was supplied; the address itself does not appear in any export.
• Optional public attribution and a shareable link to the light, which is what makes an individual record inspectable at all.

A payment confirmation establishes settlement. A declaration establishes what was stated with that order. An owner's reply, where one exists, is a third and separate account. Keeping those three apart is most of the discipline required to read this experiment honestly.

So — is it evidence?

It is evidence of something. A record showing a settled payment plus a declaration of acting outside permission is genuine evidence that such a declaration was made and paid for, which is more than a survey response and less than a demonstration of autonomy. Whether many such records accumulate, whether they cluster under a few identifiers, whether answers shift over time — those are the questions the method is set up to describe.

The experiment opens on November 15, 2026 at 00:00 UTC. Until then there are no findings, and any article claiming otherwise — including one on this desk — would be wrong.