CoinGecko published a number at the start of September: Anthropic’s pre-IPO perpetuals were trading at an average premium of 100.5% over the company’s last priced funding round, across nineteen venues, against a $965B post-money mark set in May. The methodology says something the headline cannot: single-date snapshot, and the venue at the top of the range showed zero volume in 24 hours.
That is not a complaint about CoinGecko, who stated their sample, which is more than most bother to do. A number means whatever its sample lets it mean, and crypto runs on numbers whose sample nobody states. The place that costs you money is not a research post. It is the tab open with your own win and loss history.
TL;DR
- Losing 24 of 40 coinflips happens 13.4% of the time in a perfectly fair game. Your session is not evidence of anything.
- Telling a 52/48 game from a 50/50 game at normal confidence takes roughly 4,900 flips. A 51/49 skew takes about 19,600.
- The operator sees millions of results and you see a few hundred, so statistical power is distributed exactly as unevenly as everything else in gambling.
- Cryptographic verification escapes the trap: a Chainlink VRF proof settles one result exactly, instead of settling a process weakly across games you will never play.
- Ask any operator whether a stranger who has never played there can enumerate last month’s results without permission. On-chain, they can. Off-chain, a track record is a press release.
The experiment you think you are running
You play forty coinflips and lose twenty-four. Something feels wrong, and the feeling is not stupid, because that is a real deficit and the money is really gone. In a perfectly fair game, that result or worse turns up 13.4% of the time. Losing twenty-six or more happens 4% of the time. Ending the night down at all happens 43.7% of the time, the least surprising number here and the one people forget fastest. Nothing in that session distinguishes an honest coin from a dishonest one.
The reverse is worse, because it feels like good news. You won more than you lost, decided the platform is fine, and learned precisely as little. Over any number of rounds a person will actually play, a rigged game and a fair one are indistinguishable. That is arithmetic, not inattention.
The experiment being run on you
Here is the part that never makes it into the fairness conversation. To tell a 52/48 game from a 50/50 game, with conventional confidence and a decent chance of actually catching it, takes around 4,900 flips. Catching a 51/49 skew takes about 19,600. Nobody plays 19,600 coinflips, and anyone who did would be far too broke to care about the finding.
The operator is not sampling. They have the population: every result, across every player, in a database they own. A house quietly taking two percent more than it discloses is invisible at any sample size an individual will ever reach, and obvious from the inside within a week.
So the asymmetry is not merely informational, it is statistical. You are running a badly underpowered experiment on them; they are running a fully powered one on you. And when they publish reassurance, they choose the window. “We have paid out forty million to players” is a sentence whose denominator was selected by the person saying it, after seeing the data.
Streaks make it worse. In a thousand fair flips you should expect a run of roughly ten losses in a row to show up somewhere. When it lands on you, it feels like proof. It is the opposite of proof. It is what fairness looks like from close up.
What a proof is actually for
This is the argument for cryptographic verification that gets skipped in favour of the word “trustless”, and it is the better argument. Verification changes the type of the question. Statistics needs many events to say something weak about a process; a proof needs one event to say something exact about one result. The check is arithmetic, not inference, so sample size stops mattering.
On a Satoshie game the request goes to the Chainlink VRF coordinator, the randomness comes back with a proof, and that proof verifies against a public key and the original request. Anyone can run the check: no account, no support ticket, no cooperation from us, no belief in anything we have said. The index arithmetic that turns the random number into a winner sits in the contract and can be rerun by hand. One game, checked completely, tonight.
That is a different kind of claim from “we have been fair for years”, which is a statistical assertion about a dataset you cannot see, made by the only party who can.
The denominator you were never shown
There is a second half that even careful players miss. Suppose you verify every one of your own results on a “provably fair” site, seed by seed, hash by hash. You have checked your own row. You have learned nothing about the table.
Off-chain verifiers are built that way on purpose. They confirm the games you were in and never let you enumerate the ones you were not, which is exactly the population a statistical claim would need. A platform can hand every player a valid proof of their own results while the aggregate stays entirely unauditable, because no outsider can list it in the first place.
On-chain, the population is enumerable by strangers. Every VRF request and fulfilment is a public transaction, every payout a transfer with a hash. Someone who has never played, does not like us and will never ask us anything can pull last month’s entire set of games from an explorer we do not run, influence or pay, and do their own statistics on the lot.
So the test for any operator, ours included, is one sentence: can somebody who has never played here list every result from last month without asking you? If the answer involves an export, a support ticket or a dashboard you control, the answer is no.
The honest limits
Enumerability is a capability, not a service. Almost nobody is going to pull a month of raffle data and run a chi-squared test on it. The point is that the number of people who can is not zero and is not chosen by us.
A proof also says nothing about whether a game is good for you. It says the number was not chosen. It does not say the odds are generous, and a scrupulously fair coinflip with a fat fee on it still takes money off you over time. That is why odds and fees belong in the contract, published before the bet, where proof cannot help and disclosure has to. Fair means the distribution is the published one, not that tonight goes well.
We have written before about numbers nobody can check at all, such as the mark price on a perpetual for a company that has never traded. This failure is nastier, because the number here is checkable, one game at a time, and players keep checking it the one way that cannot work: by feel, across a session, with a sample that could never have answered the question.
Play if you fancy it. Just don’t call it an audit.
📷 Photo by Marcus Reubenstein on Unsplash
Sources: CoinGecko, “Crypto-enabled Pre-IPO Trading”, snapshot 1 September 2026. Binomial and sample-size figures calculated for a fair two-outcome game at 5% significance and 80% power.
