RERUN

  1. A claim goes in.

    Someone picks a published result, like “93% accurate”. They point to the exact code and data, and set how close the Lab's number must get.

  2. People weigh in.

    You forecast YES if you think it will hold up, NO if you don’t. Each forecast tips the scale, and that sets the chance everyone can see. It has no effect on the verdict.

  3. The Lab seals itself in.

    The internet is cut and the code runs again. An AI may fix setup problems, like a wrong file path. It can’t touch the science, and plain code reads the final number.

  4. A receipt comes out.

    The result goes on the blockchain, where nobody can change it. Anyone who disagrees then has 24 hours to challenge it, and a different AI runs it again.

  5. Right forecasts get paid.

    When the time is up, right forecasts pay out and wrong ones pay nothing. If the test couldn’t be run fairly, everyone gets half back.

One claim goes in. A receipt comes out.

Many published results don't hold up, but checking takes months and nobody is paid to do it. So we built a machine that retraces them on secure hardware, while people forecast the result. Scroll, and follow one claim all the way through.

Three ways it can end

  • YES wins✓ NUMBERS CHECK OUT

    The code ran and got a number inside the claim's range. That doesn't prove the study's method is right.

  • NO wins✗ DOESN'T HOLD UP

    The code ran, but the number came out outside the range, or it ran past its time limit.

  • Half back⚠ INCONCLUSIVE

    The software wouldn't run, the data was missing, or an independent rebuild disagreed. None of that counts against the paper, and every share pays back half.

Where claims come from

  • Famous papers we picked

    Well-known results, retraced from public data, like the minimum wage study and the debt paper that shaped budget cuts.

  • The scout

    An AI agent reads new research twice a day, checks that it can be retraced, and drafts a claim. It finds the paper's number, and a second AI must find the same one.

  • Anyone

    Suggest a paper, or fill in a claim yourself with the code, the data and the number.

Every claim points to one exact version of the code and fingerprints its data, so nobody can swap them later. See what the scout found or suggest a paper.

Who does what

The AI is never the judge. Each job goes to whoever can't cheat at it.

  • People

    • Pick and approve claims
    • Set the odds by forecasting
    • Challenge a verdict they doubt

    Can't change a verdict

  • The AI

    • Fix setup problems when code won't run
    • Write the short report
    • Run the extra checks on each receipt

    Can't touch data or model code, and never decides

  • Plain code

    • Read the number from the output
    • Compare it with the allowed margin
    • Pay out on the blockchain

    Does the same thing every time

If you think a verdict is wrong

  1. The verdict posts

    Nothing pays out yet.

  2. The window opens

    On claims open for forecasts, anyone has 24 hours. Say what looks wrong and put down a deposit.

  3. A second run

    The Lab runs it again with a different AI model.

Both runs agree

The verdict stands, and the challenger loses the deposit.

See it happen on a demo claim

The runs disagree

There's no verdict, every share pays back half, and the challenger gets the deposit back plus a reward.

If nobody challenges, anyone can make the verdict final once the time is up, and winners collect. The demo claim's window was only a minute long, so we could test it quickly. Real claims get 24 hours.

You don't have to take our word for it

  • Open any receipt

    Your browser re-checks it there and then: the write-up against its fingerprint, the verdict against the numbers, and the AI's proof of secure hardware.

  • Read the blockchain

    Every claim, forecast and verdict is public on NEAR. This site only reads it.

  • Run it yourself

    Each claim lists the exact code version, commands and data fingerprints, so anyone can repeat the run.

Where it stands today: this site runs on NEAR mainnet, where money claims use real USDC (at most 5 a forecast and 50 a claim) and points are free. The AI runs on secure hardware, and since 11 November so does the Lab, on Phala Cloud. Receipts from before then came from an ordinary computer and say so.

What a verdict tells you

Whether the code and data named in the claim produce the published number. The label on each claim says whose code it is. Each receipt also shows extra evidence: a stress test for luck, and an experimental rebuild from the paper's description.

What it can't tell you

Whether the science is true. A study can reproduce and still reach a wrong conclusion. It also can't repeat a physical experiment like a drug trial, though it can re-check the analysis when the data and code are available.

Only have the paper? Check its printed numbers with no code at all.

Words you'll see

Claim

A research finding being tested, such as "raising the minimum wage didn't cost jobs".