Beamline publishes a signed random value about every ten minutes. It claims you cannot know that value before it is published. Say what the next one will be, and if you are right, a month of Claude Pro is yours.
One prize, to the first person who takes it. The offer closes at 2026-08-31T00:00:00Z: a prediction counts if GitHub stamped its issue before that instant, which is the same clock the scoring already runs on. Full terms.
The target is the full 512-bit output the beacon will
publish: 128 hex characters, exactly as it appears in every pulse. Your
prediction is lodged as a GitHub issue, which stamps it with a time we do not
control, and it is scored automatically against the next pulse to land.
Round — is the value you
would have had to name before this commit existed. Every field below is checkable:
the signature against the beacon’s public key, and
prev_output against the round before it.
—
The output is 512 bits of SHA-512. There are about 1.34 × 10154 values it could take, and the beacon draws from a quantum source, live space-weather readings, and the kernel CSPRNG, mixed and run through a NIST SP 800-90A DRBG.
For scale, the observable universe holds roughly 1080 atoms. One guess per atom would cover essentially none of the space.
So: nobody is going to win this by guessing. That is the design, not a loophole. A challenge you could win by luck would be a raffle, and it would tell you nothing about whether the beacon is sound.
Then why run it? Because the interesting attack is finding a flaw. A seeding bug, a predictable timestamp component, a DRBG misuse, a source that turns out to be replayable. Guessing is not the route in. If a flaw exists, the prediction becomes cheap, and this page is where you would prove it.
Losing attempts are not wasted. Every prediction is scored on how many leading bits it shared with the real output. Against an honest beacon that averages exactly 1.0 bit. The scoreboard below reports the running mean, so the failures accumulate into a public test for bias that anyone can recompute from the published pulses.
Every attempt, including the ones that fail. A challenge where the operator chooses which attempts to acknowledge is not a challenge.
| Round | Who | Predicted | Shared bits | Result |
|---|---|---|---|---|
| Loading… | ||||
The operator of this beacon is also the person holding the prize. That is a conflict, and pretending otherwise would be the least trustworthy thing on this page.
What protects you. We are not holding either clock. Our good intentions do not come into it. Your guess is an issue GitHub timestamps; the pulse is a commit from a public Actions run, signed with the same Ed25519 key the whole chain hangs from. Deciding a winner is string equality between two records we cannot backdate, run automatically, in the open. Nobody adjudicates anything, and if we tried, the evidence would already be in someone else’s hands.
What does not. Beamline can emit a pulse, dislike it, and publish the next one instead. It is a documented gap in the threat model, and it applies here: an operator seeing a winning prediction could in principle withhold that pulse. It would not be invisible. A withheld round means a scheduled run with no commit behind it, and both the schedule and the run log are public. It is also the reason the prize is a subscription rather than a house: the incentive to cheat should stay smaller than the cost of being caught. Watch the chain, not the promise.
On the deadline. The offer closes at 2026-08-31T00:00:00Z, and the cut is on your guess rather than on the pulse that settles it: a prediction lodged inside the window is still scored, and still wins, if the round it names lands afterwards. That is the only version of the rule a challenger can act on, because the issue timestamp is the one side of the comparison you control. The beacon does not stop at the deadline, and predictions go on being scored after it. The chain is the product; the prize is not.
Beamline documents four other ways it could be broken, covering forged chains, uncommitted draws, disagreeing verifiers and statistical bias. They carry no prize. They are written down because a claim nobody has stated plainly enough to attack is not much of a claim, and they are listed under Try to break it.