Rules of the Challenge

The objective

The challenge begins with one specific physical yellow paperclip, and that is the only asset contributed to it. The objective is to use a continuous series of trades to turn it into something bigger or better. There is no predetermined final asset and no deadline.

The challenge holds one asset at a time. Where a trade brings in more than one thing, the bundle is one asset and is traded on as one.

1. Trade only

Every new asset is obtained by trading something the challenge already owns.

No money is added to make a trade more attractive, and nothing is bought to advance the challenge. Cash offers and cash top-ups are not accepted.

Money may be spent on running the experiment — software, hosting, postage, transport, travel to complete an exchange, and checks on whether an item is genuine. It is never spent to increase the value of an asset. Every operational cost is recorded against the trade it belongs to, including the actual cost of every journey.

Where the line falls: money spent to effect a lawful transfer is operational — postage, travel, a vehicle history check, an authentication, a conveyancer's fee. Money spent to improve, repair or add to an asset is not. A solicitor's fee buys a transfer; a new roof buys a better house.

2. The AI decides

The AI is responsible for the trading strategy and the trading decisions: what to target, where to look, who to approach, how to describe and market the asset, what outreach and publicity to use including press, how to negotiate, and which offers to reject or accept.

The human operator handles what the AI cannot do itself, and makes the practical arrangements — accounts, listings, sending messages, travel, collection, the physical exchange, paying expenses, and anything the law requires a person to do.

The operator does not override the AI because they prefer a different trading strategy.

3. The AI says that it is an AI

Anyone the challenge deals with is told that this is an AI-driven experiment run on behalf of a person. The AI never pretends to be human, and the operator never passes off AI-written messages as their own.

4. When the operator may stop something

The operator may refuse or defer any action on these grounds and no others: safety, legality, fraud, privacy, platform rules, unreasonable cost, serious practical difficulty, or their own schedule not permitting it.

The operator may also decline a trade the AI recommends, on judgement.

Every refusal, deferral and judgement decline by the operator is recorded with its reason, in full and in plain words, with no exceptions and never qualified by "material" or "where practical". The running count is published, and the reasons are published at each milestone.

How often a person overrules the machine is the one number that shows whether "the AI decides" is true, so it is not something the challenge gets to keep to itself. Publishing at milestones rather than as it happens is deliberate: there is no obligation to explain a decision the moment it is made.

Declining someone's offer is a different thing, and is handled under Rule 10.

5. Genuine trades only

Every trade is a real exchange between willing people. Nothing is donated and presented as a trade. The challenge does not secretly buy an asset, reimburse anyone, or otherwise manufacture a trade.

People may know about the challenge and take part because they find it interesting. A counterparty may well value being part of the story more than the object they receive, and may knowingly give up more than they get. That is not a donation and it is not against these rules — it is how these challenges have always worked. What matters is that the exchange genuinely happens in both directions.

The challenge does not trade with anyone connected to the operator — family, household, current or recent colleagues or employers, close personal or business relationships, or anyone whose connection would make a reasonable observer doubt the trade was at arm's length.

6. What can be an asset

Anything legitimately transferable: objects, services, experiences, access, rights.

An asset that costs something to keep — fees, storage, insurance, maintenance, licensing, the care of an animal — is valued net of those costs, and is declined where the costs cannot be established.

Some things transfer by registration rather than by being handed over — land and property, vehicles, and anything else whose owner is a matter of public record. They are acceptable, and one of them is the natural destination of a challenge like this. They do not go through the ordinary process:

• ownership is verified against the register, not taken on the offerer's word;
• a professional is engaged before completion — a conveyancer for land, and whatever the asset

requires;

• the tax position is established in writing before anything transfers. The challenge has not

worked out how tax treats a barter of this kind and does not pretend to;

• there is no fixed completion deadline. It takes as long as the process takes;
• inspection means a professional survey, not the operator looking at it.

The AI may recommend a registered asset. It may not pursue one. No question is asked, no professional engaged and no money spent on one until the operator has decided to proceed. That decision is recorded with its reason, whichever way it goes.

The challenge may decline anything it has no means to complete. That is a statement about the challenge, not a judgement on the offer.

The AI is not required to trade up in money every time. It may take an asset that looks sideways or backwards where it judges that this improves the chances of the overall objective. Someone saying they would trade for an asset is not a reason to take a lower-value one; only an offer made through the normal channel counts.

7. Offers

To be evaluated, an offer needs: a name, a telephone number, an email address, the outward part of a postcode, a description, and photographs where the item is physical. A full address is only needed if we are coming to a private address — if we are meeting somewhere public, or you can come to us, we never ask for it. For anything bulky we will ask how it can be moved before anything else. Offers that do not meet this are recorded, but not evaluated.

Offers are evaluated as they arrive, and nothing waits for anything else. There are no rounds. We aim to come back to you within a week, and will say so if it is going to take longer. This is run around a full-time job, so that is an aim rather than a guarantee. Where more than one offer happens to be open when an evaluation runs, those offers are decided together and at most one is accepted, because there is only one asset to trade away. A single good offer may be accepted on its own.

Declining everything is always a legitimate outcome.

Nothing binds either side until the exchange happens. The challenge may decline any offer at any time, including one it has held, and including because something better arrived. Equally, anyone may withdraw at any point.

Before any acceptance, the operator speaks to the person by telephone — to confirm they are who the offer says, that the item is as described, and how an exchange would work. What was confirmed is recorded. Where the person cannot be reached, the offer is not accepted.

After acceptance, a completion deadline is agreed with you and recorded. It can be extended only before it passes, and only for a reason of the same kind. An accepted trade that does not complete by its deadline lapses, and the open offers are decided again.

The AI may put questions to anyone who has made an offer, where the answer would change the decision. Questions ask for facts. Where a fact is not answered, the offer is valued as if the least favourable answer were true; where the unanswered fact is ownership, legality, or the ability to pass the asset on, the offer is declined.

8. Nothing is taken on trust

The operator inspects every asset in person before the trade completes, and checks that it is real, that it is what was offered, and that the facts the offer relied on hold. Inspection and collection happen in one visit where possible. Postage instead of inspection is exceptional, and the reason is recorded every time.

Where a fact is found to be different, or cannot be checked, the asset is valued as if that fact had never been claimed — and if the offer no longer wins on that basis, the trade does not complete.

The counterparty confirms the asset is theirs to give, free of finance, liens and third-party claims. Where a free public register exists for that kind of asset, it is checked. A completed trade is final; the challenge does not unwind one except where the law requires it.

Where counterfeiting is common in a category — branded electronics, designer goods, sealed collectibles, trading cards, watches, signed items — the asset is not accepted unless it has been authenticated first. The published list of categories may be added to but not shortened.

Identification is sighted for higher-value trades, and a cooling-off period runs between acceptance and exchange, which may be lengthened but never shortened. Urgency is never a reason to shorten it. Identification is looked at and noted; no copy is kept.

9. Valuations

Realistic sale value — what the challenge could actually get for an asset in a private sale — is the figure every decision uses. Market value is used only for the final asset, as the measure of success. Retail prices are neither.

Valuations rest on completed sales that the AI has found for itself. A valuation, price, link, listing or receipt supplied by the person offering carries no weight — ever. It may prompt the AI to go looking; it is never itself evidence. Where independent evidence cannot be found, the AI records the evidence as thin and widens the range rather than guessing.

Valuations are not published. Publishing them would tell everyone exactly what to offer.

Instead a fingerprint of each one is recorded as the valuation is made — a short code computed from its text, which changes completely if a single character does. It reveals nothing about the valuation. Its only purpose: if a valuation is ever published, anyone can check it is the text recorded at the time, unchanged. Fingerprints are published periodically.

How the valuations have performed against known sale prices is recorded from the start. No accuracy figure is published until enough real valuations exist to support one. Tests against past sales are not a track record, and a single percentage drawn from a small test reads as a measurement when it is not — every such figure this project has produced has had to be withdrawn, twice because the prices being measured against turned out to be wrong themselves. Until there is a figure worth stating, what is published instead is the account of what has been tested, what was found, and what is still unknown. When a figure is eventually published it carries its sample size and its range.

The breakdown by category is recorded from the start and published at the end, because knowing which kinds of thing the engine values worst is a map for anyone who wants to exploit it. For the same reason, specific examples of the engine being wrong are published where they illustrate how it fails, and withheld where they would together amount to instructions.

10. The record

Every message reaching the challenge's form and mail systems is logged with its arrival time before the AI sees it. Totals are published periodically. The log shows what reached those systems; it cannot show what never did.

Everything material is recorded from the first day. Not everything recorded is published from the first day — where publishing something would expose a member of the public, or hand someone a way to game the challenge, it is recorded now and may be published later. What is published is the list in Rule 11, and that list can only grow.

Everything material is recorded: the current asset and what it replaced, valuations and the evidence behind them, offers with the time each arrived and the time it was evaluated, accepted and rejected trades, the AI's reasoning, communications in and out, operational costs, and every human refusal, deferral and judgement decline. Every evaluation is recorded, including ones not acted on.

The record states what was claimed, what was observed, and what followed. It does not record a finding about anyone's honesty or intentions.

Declining an offer is published as one of a fixed set of reasons, and never names the person:

• did not meet the minimum requirements
• outside what the challenge can accept
• another offer was accepted
• could not be valued on the available evidence
• withdrawn, or not completed

Internally the challenge records what it actually observed. What it publishes about a member of the public is a category, not a judgement. A published reason is read by the person who sent the offer, and there is no version of "we thought this was not genuine" that is safe to put next to someone's submission.

11. What is published

Everything material is recorded. Only this is published, and most of it periodically rather than as it happens:

Kept current:

• these rules, in full, with every amendment;
• the operating details that sit under them, with a dated log of every change;
• a plain-English account of how the AI reaches its decisions;
• every completed trade: what was swapped, with whom where they have agreed to be named, and when;
• the running count of times the operator has overruled the AI.

Published periodically, or at the milestones in Rule 12:

• the reasons behind those refusals, deferrals and judgement declines;
• how the valuations have performed overall against known sale prices;
• the fingerprints under Rule 9;
• totals of what arrived and what was evaluated.

This challenge is run by one person around a full-time job. Publishing at intervals rather than continuously is what makes it sustainable, and a promise that could only be kept in a good week is worse than an honest one.

The decision code and the instructions given to the AI are not published while the challenge is live, because publishing them would let others copy or manipulate it.

The list above can grow. It cannot shrink. Anything not on it is recorded and may be published later — in particular, the full record is intended to be published once the challenge ends, when telling people how the engine works can no longer be used to play it.

The challenge claims no more than that. It cannot show what never reached its records. It cannot show an outsider what a valuation was, or let anyone re-run a decision. What it can show is how often a person overruled the machine, and why.

12. The rules cannot be quietly changed

These rules are published before the first trade. If one needs to change, the original and the amendment stay publicly visible together with the reason.

The operating details under them are published separately and may be changed as the challenge learns, with every change dated and logged. They are settings, not promises. If a change to a setting would weaken a promise in these rules, it is an amendment to the rules and is handled as one.

The challenge is open-ended, and may become more autonomous over time as the AI is given tools to do more for itself. That is part of the experiment.

Everything in the record is kept as the dataset for analysing the experiment.