Reproducibility: Why Every Lab Figure Is Built to Be Recomputed

A measured figure is only as trustworthy as your ability to check it. Every figure the lab publishes names the methodology version used, and the raw rounds behind it are sealed every day with a public SHA-256 fingerprint. Anyone given a sealed day's records can confirm they are unaltered, recount the rounds and should get the same number.

What does "reproducible" mean here?

It means that a sceptic does not have to trust us. A journalist, a researcher or a regulator can take a published figure, obtain the sealed rounds behind it, check them against their published fingerprint, apply the stated method and arrive at the same result.

Three things make that possible:

• The rounds. Kept as per-round records, one row per round, with the result and the time it was recorded. Each UTC day's records are sealed shortly after midnight. Raw rows are released on request to logged-in users through the data page (Telegram login), up to 5 requests in any 24 hours.
• A fingerprint for each day. A SHA-256 fingerprint of the day's records, with its row count, published on the data page when the day is sealed and never changed afterwards. If a single character in the file changes, the fingerprint changes, so you can confirm you hold exactly the records we counted.
• A named method. Each figure states the methodology version that produced it. The current version is lab-v2.0. Earlier versions stay readable in the methodology archive.

Why does this matter for crash games?

Because the alternative is "trust us", and that is exactly the weakness we are trying to remove from the category.

Players are surrounded by claims about crash games: stated RTPs, screenshots of winning streaks, predictor apps with "95% accuracy", videos about hot hours. Almost none of them can be checked. A figure that can be recounted from sealed records and a public method is a different kind of statement: it can be shown to be wrong.

It also keeps us honest. We are a small team and we will make mistakes. Sealing the rounds and releasing them for checking means our mistakes can be found by people other than us, and the fingerprints mean we cannot quietly change the records afterwards.

How do you recompute a figure?

Once a game has a published measurement, the steps are:

1. Open the game page from the games list and find its data and citation section. It lists the sealed days and their fingerprints.
2. Get a sealed day. Raw rows are released on request to logged-in users through the data page (Telegram login), up to 5 requests in any 24 hours. Compute the SHA-256 of the file you receive. On macOS or Linux, shasum -a 256 followed by the file name prints it. Compare it with the published fingerprint for that day.
3. Read the method on the methodology page for the version named beside the figure.
4. Recount. For the return at a 2x cash-out, for example: count the rounds whose result reached 2x, multiply by 2, and divide by the number of rounds. For the 95% interval, use the standard deviation of the per-round returns divided by the square root of the number of rounds, times 1.96.
5. Compare your result with the published one. They should agree to the published precision.

Games that are still being counted have no figure to recompute yet. Their pages say how many rounds are still needed rather than showing a provisional number.

What does a recount not prove?

A recount proves that the published figure follows from the sealed rounds and the published method. It is a check on our arithmetic and our honesty. It does not, on its own, prove that the rounds we recorded are every round the game produced everywhere; we can only record what a provider makes public through demos and live result streams. Three Sources of Information explains those limits.

It also proves nothing about the next round. A long, correct count describes how a game behaves. It does not predict the next round.

What happens if a recount fails?

If your recount disagrees with ours:

1. Check the fingerprint. A different file gives a different answer.
2. Check the method version. A figure counted under one version will not match a recount under another.
3. Tell us. If the mismatch remains, it is a finding about our work.

We then investigate, publish what we found, and correct the figure as a new, dated publication. We do not silently edit a published number.

How is this different from other ways of establishing trust?

Two other models are common around casino games:

• Certification by a testing lab. A lab tests a random number generator and issues a certificate. This is valuable, but the data and working are usually not public, so a player cannot check it. Games such as Spaceman rely on this model.
• Provably fair seeds. A game commits to a secret seed before a round and reveals it after, so anyone can confirm a round was not changed. Aviator works this way, and our hash commitment check does the comparison in your browser. This proves a round was not altered after the fact; it does not prove the game's overall return.

Recountable measurement adds a third model: a public count of many rounds, with the records sealed and available for checking, that tests the stated return directly. It is the only independent check available for games that cannot be verified round by round, and a useful complement for those that can.

What does this mean for the myths?

The same standard applies to the claims players act on. When we say a predictor bot cannot work, or that hot hours do not exist, the test is meant to be one anyone can repeat on fresh rounds. See the checks for predictor bots, patterns and hot hours and is it rigged?.