About
Methodology, the maths behind the jar, and the data to check it yourself.
What Wassily is not
- It does not predict price. It estimates whether a token that already reached $10K goes on to reach $30K.
- It has no edge that guarantees returns. A handful of features cannot forecast a market.
- The jar is not decoration. If the model is weak, the jar stays empty and the site names the gate that blocks it.
1. Hoeffding’s coins
Wassily is named after Wassily Hoeffding (1914–1991), born in what was then the Grand Duchy of Finland, trained in Berlin, and later a professor at Chapel Hill. In 1948 he described U-statistics, the family AUC belongs to. In 1963 he proved how fast an average of bounded random variables settles near its expectation: the chance that it strays by t shrinks like exp(−2nt²).
2. Why it fits Robinhood Chain
Every launch on Robinhood Chain is a coin of unknown bias. One flip says nothing; thousands say a lot, and Hoeffding tells you exactly how much. Wassily keeps flipping, writes every result down, and refuses to believe the average until it has nowhere left to hide.
3. What Wassily measures
Every token on Robinhood Chain DEX pools that reaches $10,000 peak market cap joins the study. Wassily records whether it reaches $30,000 peak market cap, using four feature families: launch hour (sine and cosine), day of week, a holder count sampled once at 48 hours, and the lore text (length, a missing flag, words in the name, and a 15-bucket hashed bag of words). That makes d = 28. On live data the lore columns stay zero: a token's description cannot be observed for past launches, so no token trains on it. Nothing derived from price, volume or liquidity is ever a feature.
4. Why the jar fills slowly
Hoeffding bound for U-statistics · Wassily Hoeffding, 1963
A model can look good by luck on a small sample, so the jar is driven by a proven floor, the cross-validated AUC minus a penalty ε, never by the raw score.
AUC is a two-sample U-statistic, so Hoeffding’s inequality applies with the smaller class as the effective sample size.
Jar = clamp((floor − 0.5) / (0.6 − 0.5), 0, 1), capped at 95% until all four gates pass: n ≥ 2,000, n₊ ≥ 200, fold σ < 0.05, and a time-split gap ≤ 0.04. For the VC bound, d = 28 is a chosen capacity parameter (the feature count), not a derived VC dimension.
5. Check the work
Checking which data source this server uses
Download every labelled token, rerun the model, and check whether the jar is honest.