How it works

Every price here is a probability, and every probability was moved by somebody. This is the arithmetic behind that, worked out by the same code the desk runs.

One trade, taken apart

50 credits, and what it does to the board

Every number below was computed when this page rendered, by importing the desk's own pricing module. Nothing here is typed in by hand.

When does GPT-4o fall out of the top five?

The answer is split into 8 quarters. It has to land in one of them, so the 8 prices add to 1 — which is why each one can be read as a chance.

Q4 2025

12.5%

Q1 2026

12.5%

Q2 2026

12.5%

Q3 2026

12.5%

Q4 2026

12.5%

Q1 2027

12.5%

Q2 2027

12.5%

Q3 2027

12.5%

Nobody has traded yet, so the desk opens flat at 12.5% a quarter. That is not a forecast. It is the absence of one.

Someone thinks the slide is faster than that

They pick Q3 2026 as the line and take the long side of it.

Long the decay means buying every quarter before the line — Q4 2025 through Q2 2026. Short means the line and everything after it. Between them they cover all 8 quarters, so the two prices always add to 1.

Long 37.5%Short 62.5%

The room is saying there is a 37.5% chance GPT-4o is out of the top five before Q3 2026. Our trader thinks it is higher than that, so they buy it.

What 50 credits buys

The price moves while they are buying. That movement is the whole charge — there is no separate spread and no fee on top.

Stake

50 credits

Long price before

3 quarters at 12.5% each

37.5%

Shares bought

Each share pays one credit if the long side is right

112.16

Average paid per share

Between the price on the way in and the price on the way out

44.58%

Long price after

51.8%

If GPT-4o does fall out of the top five before Q3 2026, those 112.16 shares pay one credit each: 112.16 credits back on 50 risked. If it does not, they are worth nothing. The desk quoted both sides of that on its own, with nobody else in the room.

The board, redrawn

One trade, and all 8 numbers change. Rust bars are the quarters this position owns.

Q4 2025

17.3%

Q1 2026

17.3%

Q2 2026

17.3%

Q3 2026

9.6%

Q4 2026

9.6%

Q1 2027

9.6%

Q2 2027

9.6%

Q3 2027

9.6%

The long side went from 37.5% to 51.8%. The quarters nobody bought had to give that up, because the eight still have to add to 1. That is the move the desk shows you, with the name of whoever caused it.

Why this rule

Three ways to run a small desk

A desk this size has three things to worry about: can it quote when the room is empty, is the number on screen actually a chance, and can the operator say up front what the worst day looks like. Here is how the options answer.

An order book

The venue matches buyers to sellers and takes a cut. It is what every exchange you have used runs on.

Quotes with nobody there

A book with one side empty has no price at all. On a desk with eight quarters and a handful of traders, most quarters are empty most of the time.

The price is a chance

Partly

In a busy book the last trade is a fair read. In a quiet one there is a wide gap between bid and ask and no single number to point at.

Worst case is knowable

Yes

The venue never takes a side, so it cannot lose. This is the order book's real advantage and it is a big one.

What it costs you

Genuinely the right answer at scale. It needs a crowd, and a demo desk does not have one.

A constant-product pool

Two reserves and a curve through them. Anyone can put money in, and the pool quotes off the ratio.

Quotes with nobody there

Yes

The curve always has a price, same as ours. Nobody has to be watching.

The price is a chance

The price is a ratio of two reserves. For a two-way bet you can normalise it into something chance-like; across eight quarters it does not add to 1 without bolting on extra machinery.

Worst case is knowable

Partly

You can compute it, but it depends on where the price ends up and how deep the pool got. There is no single number to put on screen before the first trade.

What it costs you

Cheap to run and needs no operator money — the people supplying it earn the fees. It just does not hand you a probability, and this product is about probabilities.

LMSR

Hanson's rule. One market maker holds every side, and the price falls out of how much of each quarter it has sold.

Quotes with nobody there

Yes

The maker is always the counterparty. There is never a moment where a trade cannot be done.

The price is a chance

Yes

The eight prices sum to exactly 1 by construction. When the screen says 37.5%, that is not a translation of a price — it is the same number.

Worst case is knowable

Yes

b x ln(n), fixed before anyone trades. It does not grow with volume, and the desk states it out loud.

What it costs you

It charges by moving the price, so there is no spread to quote and a big order pays visibly more than a small one. The operator is on the other side of everything and carries the loss. And depth is one dial: set it low and a modest trade whips the board, set it high and real information barely shifts it.

None of this makes LMSR better in general. An order book beats it on a venue with real volume, and a pool beats it on running costs. It wins here because the room is small, the question has eight answers, and the whole product is the number on screen being a chance you can argue with.

What is being measured

Rank, not rating

A market needs something it can settle against — one number, checked the same way every day, that either happened or did not.

Step 1

Each model has a share of attention

On any given day, every model that is live holds some slice of the room — what people reach for, write about and build on. The slices are normalised across everything live, so they add to one whole.

Step 2

The shares are ranked, every day

Sort the shares, biggest first. Best in the field that day is rank 1. A model does not need to get worse to slide down this list — it only needs to be passed.

Step 3

Falling out of the top 5 settles the market

The first day a model's rank goes past 5, its market settles into whichever quarter that day sits in. Every share on that quarter pays one credit. Every share on the other 7 pays nothing.

Why rank and not a score

The obvious alternative is to settle on a rating — a benchmark number, an arena score. It does not survive the years this market has to cover.

Ratings get rebased. A benchmark saturates and is retired, a new one replaces it, the scale is recalibrated, and the number that meant frontier in one era means something else two years later. Settle a three-year market on a rating and you are settling on a ruler that changed length halfway through.

Rank is ordinal. Fifth is fifth no matter what is being scored underneath or how the scores were rescaled last month. It survives the rebase because it never depended on the units in the first place.

It is also the thing the question is actually about. A model does not stop being useful when it slides — it stops being the one people pick. Being passed is the event. Rank measures being passed.

Where these curves come fromCurated model list, modelled curves

Real models and approximate public launch dates, curated by hand. The relevance curves are modelled, not measured — they behave the way obsolescence behaves, but no number here is a fact about the real world. Two entries are unannounced successors, present because the market is about what comes next.

There is no live feed behind any of this, and nothing on the desk is reading the internet. Swapping the model list for a real one is the single change that would make the underlying real — everything downstream already reads the same shape.

The credits are play money. There is no cash-out, nothing to buy and no way to fund an account.

The operator's side

What the desk can lose, stated before it opens

Somebody is on the other side of every trade here, and it is the house. That is a real position with a real loss attached, so it is worth being exact about how big it can get.

The numbers for one market

Read off the same two functions the desk uses when it opens a market.

Subsidy the operator puts up

Chosen up front, per market

400 credits

Quarters priced

The n in the formula below

Depth constant b

Subsidy divided by ln(8)

192.36

Worst case, b x ln(n)

The most this market can ever cost the operator

400.00 credits

Why it cannot get worse than that

The worst case comes out to exactly the subsidy, and that is not a coincidence. b is defined as the subsidy divided by the same ln(8) that the worst case multiplies it back by, so the two cancel and what is left is the number the operator picked. The subsidy is not a float and not a fee — it is the loss, named in advance.

It does not grow. Not with the number of trades, not with the size of them, not with how badly the desk misread the room. A thousand traders piling onto the quarter that turns out to be right still costs the operator 400.00 credits and not one more.

b is also the only depth dial there is. A bigger subsidy means a deeper market where a trade moves the price less, and a larger number at risk. A smaller one means a board that whips around on modest orders. There is no setting that gives you both.

When the desk makes money and when it does not

The desk takes credits in on the way through and pays one credit per share on whichever quarter actually happens. If the room piled into a quarter and the slide came somewhere else, the desk keeps most of what it collected. If the room called it, the desk pays out more than it took and the difference comes out of the subsidy.

So the house is not selling a service here, it is buying an opinion. The subsidy is what it pays a room of traders to argue the number into shape, and the running profit and loss is on the desk view alongside everything else — including when it is red.

All of it is play money. Nobody funds anything, nothing pays out in the real world, and the operator's worst day costs 400.00 credits that never existed.

Try this

Four things worth doing on the desk

Reading about a price moving is not the same as moving one. These take a couple of minutes each.

Open the desk

Put a big stake on one quarter

Pick any market, take a side with a stake far larger than the bots are using, and send it. Every other quarter moves in the same instant, because the whole board still has to add up to 1.

Open a bot's move and read it back

Every move on the board has a name against it — click one and the inspector gives you the price before, the price after, and what it paid. Underneath that is the reason the bot gives for taking it, in its own words.

Step a day at a time until a market settles

Advance the desk one day at a time and watch a model slide down the board as newer ones land. The day it falls out of the top five, its market settles into that quarter and every position on it resolves at once.

Turn the written notes on

The notes beside each move can be written by a model instead of assembled from a template. Switch it on and the tokens and the dollar cost start appearing next to each note, so you can watch the prose run up a bill.

How it works

Every price here is a probability, and every probability was moved by somebody. This is the arithmetic behind that, worked out by the same code the desk runs.

One trade, taken apart

50 credits, and what it does to the board

Every number below was computed when this page rendered, by importing the desk's own pricing module. Nothing here is typed in by hand.

When does GPT-4o fall out of the top five?

The answer is split into 8 quarters. It has to land in one of them, so the 8 prices add to 1 — which is why each one can be read as a chance.

Q4 2025

12.5%

Q1 2026

12.5%

Q2 2026

12.5%

Q3 2026

12.5%

Q4 2026

12.5%

Q1 2027

12.5%

Q2 2027

12.5%

Q3 2027

12.5%

Nobody has traded yet, so the desk opens flat at 12.5% a quarter. That is not a forecast. It is the absence of one.

Someone thinks the slide is faster than that

They pick Q3 2026 as the line and take the long side of it.

Long 37.5%Short 62.5%

The room is saying there is a 37.5% chance GPT-4o is out of the top five before Q3 2026. Our trader thinks it is higher than that, so they buy it.

What 50 credits buys

The price moves while they are buying. That movement is the whole charge — there is no separate spread and no fee on top.

Stake

50 credits

Long price before

3 quarters at 12.5% each

37.5%

Shares bought

Each share pays one credit if the long side is right

112.16

Average paid per share

Between the price on the way in and the price on the way out

44.58%

Long price after

51.8%

The board, redrawn

One trade, and all 8 numbers change. Rust bars are the quarters this position owns.

Q4 2025

17.3%

Q1 2026

17.3%

Q2 2026

17.3%

Q3 2026

9.6%

Q4 2026

9.6%

Q1 2027

9.6%

Q2 2027

9.6%

Q3 2027

9.6%

Why this rule

Three ways to run a small desk

An order book

The venue matches buyers to sellers and takes a cut. It is what every exchange you have used runs on.

Quotes with nobody there

A book with one side empty has no price at all. On a desk with eight quarters and a handful of traders, most quarters are empty most of the time.

The price is a chance

Partly

In a busy book the last trade is a fair read. In a quiet one there is a wide gap between bid and ask and no single number to point at.

Worst case is knowable

Yes

The venue never takes a side, so it cannot lose. This is the order book's real advantage and it is a big one.

What it costs you

Genuinely the right answer at scale. It needs a crowd, and a demo desk does not have one.

A constant-product pool

Two reserves and a curve through them. Anyone can put money in, and the pool quotes off the ratio.

Quotes with nobody there

Yes

The curve always has a price, same as ours. Nobody has to be watching.

The price is a chance

The price is a ratio of two reserves. For a two-way bet you can normalise it into something chance-like; across eight quarters it does not add to 1 without bolting on extra machinery.

Worst case is knowable

Partly

You can compute it, but it depends on where the price ends up and how deep the pool got. There is no single number to put on screen before the first trade.

What it costs you

Cheap to run and needs no operator money — the people supplying it earn the fees. It just does not hand you a probability, and this product is about probabilities.

LMSR

Hanson's rule. One market maker holds every side, and the price falls out of how much of each quarter it has sold.

Quotes with nobody there

Yes

The maker is always the counterparty. There is never a moment where a trade cannot be done.

The price is a chance

Yes

The eight prices sum to exactly 1 by construction. When the screen says 37.5%, that is not a translation of a price — it is the same number.

Worst case is knowable

Yes

b x ln(n), fixed before anyone trades. It does not grow with volume, and the desk states it out loud.

What it costs you

What is being measured

Rank, not rating

A market needs something it can settle against — one number, checked the same way every day, that either happened or did not.

Step 1

Each model has a share of attention

Step 2

The shares are ranked, every day

Sort the shares, biggest first. Best in the field that day is rank 1. A model does not need to get worse to slide down this list — it only needs to be passed.

Step 3

Falling out of the top 5 settles the market

The first day a model's rank goes past 5, its market settles into whichever quarter that day sits in. Every share on that quarter pays one credit. Every share on the other 7 pays nothing.

Why rank and not a score

The obvious alternative is to settle on a rating — a benchmark number, an arena score. It does not survive the years this market has to cover.

Rank is ordinal. Fifth is fifth no matter what is being scored underneath or how the scores were rescaled last month. It survives the rebase because it never depended on the units in the first place.

It is also the thing the question is actually about. A model does not stop being useful when it slides — it stops being the one people pick. Being passed is the event. Rank measures being passed.

Where these curves come fromCurated model list, modelled curves

The credits are play money. There is no cash-out, nothing to buy and no way to fund an account.

The operator's side

What the desk can lose, stated before it opens

Somebody is on the other side of every trade here, and it is the house. That is a real position with a real loss attached, so it is worth being exact about how big it can get.

The numbers for one market

Read off the same two functions the desk uses when it opens a market.

Subsidy the operator puts up

Chosen up front, per market

400 credits

Quarters priced

The n in the formula below

Depth constant b

Subsidy divided by ln(8)

192.36

Worst case, b x ln(n)

The most this market can ever cost the operator

400.00 credits

Why it cannot get worse than that

When the desk makes money and when it does not

All of it is play money. Nobody funds anything, nothing pays out in the real world, and the operator's worst day costs 400.00 credits that never existed.

Try this

Four things worth doing on the desk

Reading about a price moving is not the same as moving one. These take a couple of minutes each.

Open the desk

Put a big stake on one quarter

Pick any market, take a side with a stake far larger than the bots are using, and send it. Every other quarter moves in the same instant, because the whole board still has to add up to 1.

Open a bot's move and read it back

Step a day at a time until a market settles

Turn the written notes on