How the AI decides

Published under Rule 11, and kept current with each version of the engine.

The code itself is not published while the challenge is live, because publishing it would let people copy it or work out how to game it. This is the account of what it does instead, and it is written to be checkable against what actually happens rather than to sound impressive.

How the AI decides

Required by Rule 11. This is the plain-English account of what the AI looks at, what it weighs, and what stops a trade. It is kept current with each version of the engine.

Engine version 1.0 · 14 September 2026

The short version

An offer arrives. The AI works out what the item actually is, looks for evidence of what things like it have genuinely sold for, decides what the challenge could realistically get for it, subtracts what the trade would cost to complete, and compares that with what the challenge is holding now. Several things can stop a trade outright regardless of the numbers. If the evidence isn't good enough, the answer is "not enough evidence" rather than a guess.

1. What is this thing, exactly?

Almost every expensive mistake starts here, so identification comes before valuation rather than alongside it.

A "Pink Floyd first pressing" and a 1970s repress can differ by a factor of ten, and they look the same in a photograph. So the AI tries to pin down the exact thing — model number, catalogue number, edition, year, variant — before it looks at a single price.

Where it can't, it says so, and the range it records gets wider rather than the figure getting confident.

It also scores how likely the item is to be different from how it's described. This turns out to be the most reliable signal in the whole system. Tested against deliberately planted false descriptions — a catalogue number that belongs to a different format, a variant that never existed, a release dated two years before the record came out — real items score around 0.04 and planted ones 0.75 to 0.88. There is no overlap. A high score stops the offer rather than reducing its value.

2. What have things like this actually sold for?

Only completed sales count, and only ones the AI has found for itself.

A price, link, listing or receipt supplied by the person offering carries no weight at all. It can prompt the AI to go looking. It is never itself evidence. This isn't a comment on anyone's honesty — it's that a valuation built from what the seller says it's worth isn't a valuation.

Three things the AI has to get right here, each of which it previously got wrong:

Individual sales, never averages. A site's "median sold price" mixes mint copies with damaged ones. Used as a starting point it drags every estimate towards the middle and away from the truth.

Match the condition. This was the single biggest error in the system. The AI was pricing well-used items against sales of pristine ones, and pricing ordinary items on intuition. Fixing it cut the error on the category it was measured on by roughly three quarters — the largest single improvement the engine has had.

Tell a sale from an asking price. For a great many ordinary second-hand things, nothing public records what they actually fetch. What the open web holds is what people are asking, which is a different and consistently higher number. An engine that cannot tell the two apart will be confidently wrong in a predictable direction, which is worse than being randomly wrong. How far wrong, and on what, is recorded but not published — see Rule 9.

A claim made earlier and since withdrawn: that the AI undervalued cheap things and overvalued expensive ones. That was measured on one category, against reference prices that later turned out to be wrong, and it does not survive a second category being tested. The bias follows the kind of thing and where it sells, not the price.

It also appeared to undervalue records, at every level. That part is now in doubt — the price archive those reference figures came from only lists auctions above a threshold, which would make its prices read high for cheap common records. When we checked three of them against every eBay sale in three months, the market said roughly half what our own reference set said. We are re-checking before claiming anything about records either way.

Read the description, not the headline. Sellers' titles overstate condition against their own description text, routinely. A listing titled "Excellent" described the record as played-tested and the sleeve as good. Reading the body changed the answer repeatedly.

3. Is the source real?

The open web lies confidently, and this turned out to be the biggest practical problem in the project.

Four separate sources have been caught returning a perfectly normal-looking page containing wrong information: a marketplace that silently dropped both the search term and the "sold" filter; a price archive that returned a different query's results; a gold calculator serving prices six years out of date with no warning; and AI-written "price guides" quoting thousands of dollars for a record that trades in the low hundreds — pages built to rank in search results, which means built to be found by exactly the step the AI depends on.

So every source is checked before anything on it is believed:

Does this page actually mention what we searched for?
Does the source return the same thing for a query that can't exist? If so it's broken right

now and everything it returned is worthless.

Does the data carry its own date? A price with no date is not evidence, and the date has to

come from the data, not from when the page was fetched.

Is this source wildly out of line with the others? A source claiming ten times what everything

else says is wrong, not insightful.

A source that fails any of these is discarded, not discounted. The number discarded is recorded with the valuation, because it says something about how much the answer should be trusted.

4. What's it actually worth to us?

Two different questions, and the AI keeps them apart:

Realistic sale value — what the challenge could actually get for the thing in a private sale.

Not the retail price. Not what it's listed for.

What the trade costs — travel, postage, any authentication, and anything the asset costs to

keep, like storage, insurance or fees, over the following year.

Everything is compared in pounds. An earlier version of the engine scored offers on nine weighted dimensions and produced a number out of ten, which made a strong score on one thing cancel out a fatal problem on another. Points hid that. Pounds don't.

Fewer than three usable comparable sales and the answer is "insufficient evidence." That isn't a failure state — it's the correct answer, and it will be a common one. Records are an unusually well-documented category. For a bicycle or a chest of drawers there is no equivalent archive, and the honest expectation is that the AI declines to value a good deal of what it's offered.

5. What stops a trade regardless of the numbers

Some things are gates, not factors. A good score cannot outweigh them:

the item is probably not as described
ownership, legality or the right to pass the asset on can't be established
the category is one where fakes are common and authentication hasn't happened
the evidence isn't good enough to value it at all
a fact the offer depended on turns out, at inspection, to be different or uncheckable — the item

is then valued as though that fact had never been claimed

the cost of completing the trade would exceed what the asset is worth

6. Value doesn't have to go up every time

The AI is allowed to take something that looks sideways or backwards if it judges the position improves — better liquidity, a more tradeable category, somewhere with more potential counterparties. Kyle MacDonald went paperclip, pen, doorknob, camp stove. Every one of those lost money once travel was counted. Every one was necessary.

7. Where a person comes in

The AI decides. The operator can stop a trade on specific grounds — safety, legality, fraud, privacy, platform rules, unreasonable cost, practicality, or his schedule — and can decline a recommended trade on judgement. Every one of those is published with its reason. A telephone call happens before any acceptance, and the item is inspected in person before any trade completes.

The operator isn't supposed to overrule the AI because he'd have traded differently, and when he does intervene it goes on the record.

How well does this work?

Not as well as the phrase "AI valuation engine" suggests.

The engine has been tested blind, in rounds, over weeks, against hundreds of verified dated sale prices. No accuracy figure is published, and Rule 9 sets out why: every figure this project has produced has had to be withdrawn, twice because the prices being measured against turned out to be wrong themselves. Tests against past sales are not a track record. When enough real valuations exist to support a figure, it will be published with its sample size and its range attached. Until then a percentage would be a claim, not a measurement.

What can be said without a number:

Valuing second-hand things is hard, and not only for machines. On one set the auction houses'

own pre-sale estimates — written by professionals who had the objects in their hands — were no more accurate than the AI's. On another they were clearly better. One George I walnut chair they estimated at £250–350 sold for £3,800.

The range the AI gives is too narrow. The "quick to optimistic" band is meant to contain the

true price about four times in five. It does not manage that.

It is at its most confident exactly where it is least reliable. On the items it judged well

evidenced, the band contained the true price less often than on the items it flagged as thin. That is the worst way round, it has now been seen on two separate test sets, and it is the single thing about the engine that most deserves distrust.

A good description is worth a great deal. The same objects described in one line rather than

properly produced roughly double the error. That is why the offer form asks for model numbers, serial numbers and dimensions rather than asking politely.

There are categories it is markedly worse at than others, and which ones is not published.

Rule 9 holds the per-category breakdown back until the end of the challenge, and holds back worked examples that would together amount to instructions, because knowing where the engine is weakest is a map for anyone who wants to exploit it. The breakdown is recorded from the start and will be published in full at the end.

It declines to value most of what it is offered. Three dated, condition-matched sales is a hard

bar for ordinary second-hand goods, and for a great many things no such record exists anywhere public. Expect "not enough evidence" to be a common answer. It is the correct one.

Everything above comes from small samples. None of it is a measurement of how the engine will perform on real offers, because no real offer has been valued yet.

What this account does not claim. That the engine is accurate. That an outsider can verify any of it — the valuations aren't published and the code isn't published while the challenge is live. That the AI is immune to being talked into a higher number. Those limits are set out in the rules, and this page doesn't quietly walk them back.