A man whose job was to read the president’s speech before the president did made more than $100,000 betting on which words the president would say. He did it on Kalshi, in public, on a regulated prediction market, for months. Nobody caught him. A news organisation did.
Gabriel Perez had operated the White House teleprompter since 2016. Kalshi runs “Mentions” markets, where you bet on whether a specific word or phrase will appear in a major speech. Perez, according to ABC News, cleaned up on them, including on the State of the Union. He was put on unpaid administrative leave after the reporting landed, Trump called it a disgrace, and as of this week he no longer works in the federal government. Kalshi’s response was to start asking users to disclose their employer.
TL;DR
- A White House teleprompter operator won over $100,000 on Kalshi betting on which words Trump would say in speeches he had already read, including the State of the Union.
- The platform did not detect it. Journalists did. Kalshi’s fix is an employer disclosure field, which is a policy patch on an architectural hole.
- This is the same defect as server-side RNG in crypto gaming: the outcome exists before you commit, and someone can see it.
- Chainlink VRF removes the window entirely. The random number does not exist until after your stake is locked on-chain, and the contract verifies the cryptographic proof before accepting it.
- Satoshie settles every coinflip and raffle this way on Base. There is no seed to leak, no teleprompter to read ahead, and no employer field to lie on.
The problem was never the gambling
Kalshi has spent 2026 insisting it is not a gambling venue. It has fought state regulators, lost a bid in New York to block gambling law from applying to it, and marketed itself as a truth machine that aggregates information better than polls do. Then it turned out one of its users was not aggregating information. He had it.
To be fair to prediction markets, informational advantage is the entire mechanism. Markets are supposed to reward the person who knows more. That is how prices get accurate. The line where that breaks is the line between knowing more and knowing the answer. Perez did not have a better model of what Trump might say. He had the script in front of him. At that point the market is not pricing uncertainty, it is processing a withdrawal.
What matters here is not that a person behaved badly. People always will. What matters is that the system had no way of noticing, and that its remedy arrived after the money left.
An employer field is not a security model
Kalshi now asks users where they work, and its terms prohibit trading on information obtained through your job. Consider what that actually is: a promise, checked by nobody, enforceable only after an investigation that in this case was conducted by ABC News rather than by the venue.
Every trust-based system ends up here. You cannot verify the thing that matters, so you write a rule about it, then you rely on journalists or an eventual blow-up to tell you the rule was broken. We watched the same pattern with exchanges that promised segregated funds and with stablecoins that promised no freeze functions.
Cryptography works in the opposite direction. It does not ask people to behave, it removes their ability to misbehave, and it does so before the money moves rather than after.
Crypto gaming has its own teleprompter
This story should be uncomfortable for most of the crypto gaming industry rather than a chance to point and laugh.
The standard “provably fair” setup at a crypto casino works like this. The operator generates a server seed in advance and shows you its hash. You supply a client seed. After the round, the operator reveals the server seed and you can check that it hashes to what was published. This is genuinely better than nothing, and it is also structurally identical to the teleprompter problem: the outcome exists before you bet, and exactly one party can see it.
Everything rests on the operator not reading the script. There is no cryptographic reason they cannot. They can pre-generate thousands of seeds, hash them all, and select which one to serve based on how large your bet is. The hash you were shown still verifies afterwards. The maths checks out and the game was still rigged. It is the same trick Perez ran, with the difference that his teleprompter was in the West Wing and theirs is in a database.
Ask any platform advertising provable fairness who generates the seed and when. If the answer is “we do, in advance,” you are trusting an employer disclosure field.
What VRF actually removes
Chainlink VRF was designed to kill this specific window. On Satoshie, your stake and the randomness request go on-chain in the same transaction. At that moment the number that decides your outcome does not exist anywhere, not in a database, not in a config file, not in the head of anyone at Satoshie. The oracle network then generates it, signs it with a key pair whose public half is registered on-chain, and returns it with a proof. Our contract verifies that proof before it will accept the value. A tampered number is not rejected by policy. It is rejected by the verifier, automatically, or the transaction reverts.
Then the resolution is arithmetic anyone can repeat. A coinflip is randomWords[0] % 2. A raffle winner is keccak256(VRF word + prior blockhash) % ticketsMinted. Both settle inside the same transaction that delivered the randomness, so there is no gap in which a human could look at the answer and decide what to do about it. Pull the transaction on BaseScan and check it yourself. You are not verifying that we behaved. You are verifying arithmetic, which is the only kind of verification that does not eventually depend on someone’s honesty.
That is the whole difference. Kalshi has to find out who its users work for. We do not have to find out anything about anyone, because there is no privileged position on the platform to occupy: no seed custodian, no admin key that reveals outcomes early, no operator who sees the result before the player does.
The industry keeps patching the wrong layer
Prediction markets got an AI referee to police $18 billion a month in outcomes. Crypto exchanges got compliance departments. AI agents are getting a court. Kalshi got a dropdown asking where you work. Every one of these is a human or an institution inserted to compensate for a design that lets someone know the answer first.
The alternative was available the whole time and it is unglamorous: build systems where the answer does not exist until it is too late for anyone to act on it. That is a smaller ambition than a metaverse and it is the only claim in this industry that survives contact with a determined adversary.
A guy with a teleprompter took $100,000 out of a market that calls itself a truth machine, and the truth machine found out from the news. Do not accept that architecture in your games.
Satoshie runs provably fair coinflip and raffle games on Base, settled with Chainlink VRF. Every outcome is verifiable on-chain, by you, without asking us anything.
📷 Photo by Thomas Fromont on Unsplash
