experiment 001pump.funevery 3 min

Every 3 min
Jev decides.burn, airdrop or roll?

$JEVLAB is run by Jev, a model that doesn't chat. Every 3 minutes it gets one question about the coin's fees. It answers in milliseconds. The engine obeys. No team cut, and every answer is public, misses included.

CAnot launched yet
next question in
—:——
trial #—
What happens to the pot?
🔥 Burn
—
🪂 Airdrop
—
🎲 Roll
—
ask Jevclaim feesone buyburn / airdrop
Waiting for the first trial.
—in the pot
—Jev's answer time
—dose of the pot
—"is it dumping?"
—trials run
—tokens burned
—tokens airdropped
—fees put to work

The chart, with every decision on it.

Market cap — · each dot is one answer from Jev.
The chart starts when the coin does.
burnairdroprollhollow = JEVLAB engine (on-chain seed)

Three moves. Which one works?

Three minutes after every decision, the price is written down next to it. That's the whole experiment: does a model that answers in 100 ms run a coin better than a fixed rule?

🔥

Burn

One market buy with the fees, then the tokens are burned. Supply goes down for good.

—times run
—tokens
—avg Δ 3 min
🪂

Airdrop

One market buy with the fees, then the tokens land in the top holders' wallets. Nothing to claim.

—times run
—tokens
—avg Δ 3 min
🎲

Roll

Nothing moves. The fees stay in the pot, so the next answer decides a bigger one.

—times run
—tokens
—avg Δ 3 min

Engine trials take their move from a public on-chain seed (see How it works), so each move gets a fair sample. Scored forward only: nothing is back-filled and nothing is deleted. Three minutes of a memecoin chart is mostly noise — if the model has no edge, this is where it shows.

How it works. The exact words.

Jev is TypeSafe AI's System One model. It doesn't write text: you send it a situation and typed questions, it sends back a decision with probabilities in 70–500 ms. On the right is the real request the lab sends, every trial.

1Read the coinMarket cap, price change over 3 / 15 / 60 min, the last buys and sells, the size of the pot, and how earlier trials turned out.
2Ask Jev onceOne call, four typed questions: the move, the dose (25–100% of the pot), "is it dumping?", "is there momentum?". No retries for a nicer answer.
3ObeyRoll: nothing moves. Otherwise claim 100% of the creator fees → ONE market buy → burn the tokens or send them to the top holders.
4Write it downWhat Jev saw, its probabilities, its speed, the transactions — and three minutes later, what the price did.
The JEVLAB engineUntil the model is plugged in, and whenever it cannot be reached, the JEVLAB engine takes the move from a public on-chain seed: the blockhash of the newest finalized Solana block when the clock hits zero — nobody knows it in advance. sha256(blockhash:trial#) mod 3 → burn / airdrop / roll, dose 100%. The slot and blockhash are saved with the trial; tap a row in the notebook and your browser recomputes it. These trials are tagged engine and counted separately from the model's, so the two can be compared.
POSThttps://api.typesafe.ai/v1/systemone
loading…

The notebook. Receipts included.

Every trial, newest first. Tap a row to see the state of the coin at that moment and how the move was chosen — engine seeds are re-checked in your browser.

#whenbymovedoseconf.mspotΔ 3 minwhat happenedtx
Empty. The first page gets written three minutes after the coin launches.

What the lab can't do.

The engine has three moves and nothing else.

It can't sell.Buy-and-burn, buy-and-airdrop, wait. There is no code path that sells the coin.
It can't keep anything.100% of the creator fees go into trials. No team cut. What it buys is burned or airdropped in the same trial.
It can't touch other funds.The lab keeps its own ledger: it only spends fees it claimed and only moves tokens it bought.
It can't re-ask.One call per trial. Whatever Jev answers is what happens. No overrides.
It can't edit the questions.If the protocol ever changes, it's experiment 002 with a fresh results table.
It can't hide a miss.Every trial stays in the notebook — skipped, failed and engine ones too.