Overview

Wassily reads every Robinhood Chain token that clears $10K and only trusts what a provable floor allows.

Phase 0 · warming up
Jar level
target AUC 0.60
ROC-AUC · 5-fold
connecting…
Proven floor
Hoeffding
Above $10K
0
labelled tokens
Reached $30K
0
0.0% · 0 stalled
Median holders
0
sampled at 48h
Connecting to Wassily feed
CA not deployed
Wassily never clocks out. Only the jar decides when a launch is allowed.

The jar

Filled by the Hoeffding bound for U-statistics

Connecting
0.0%

The jar stays empty until real tokens are labelled, 48h after launch. No numbers are shown before they exist.

Explore the proof

Live ingest

0 pulled · 0 priced · 0 holder counts · 0/min

Waiting for the first labelled token.

Validation gates

The jar is capped at 95% until all four pass

0/4
  • Sample size ≥ 2,000waiting
  • Survivors ≥ 200waiting
  • Fold σ < 0.05waiting
  • Time-split gap ≤ 0.04waiting

What the model leans on

Top feature families by coefficient weight, reached $30K vs stalled

Weights appear once real tokens are labelled and the model has something to learn from.

Proof & capacity

Hoeffding bound for U-statistics · Wassily Hoeffding, 1963

What-if · example values

No tokens are labelled yet, so these sliders start from example values to show how the bound behaves. The real jar is the one above.

Measured AUC
0.580
Penalty ε
0.046
Proven floor
0.534
2,346
0.580
28
30.8%
n₊ 723 · n₋ 1,623

Measured 0.580, ε removes 0.046, so the floor is 0.534.

Why the jar fills slowly

On a small sample, luck looks like skill. With n tokens and n₊ survivors, the true AUC sits at most ε below the measured one, with 95% confidence.

AUC is a two-sample U-statistic, so Hoeffding’s inequality applies with the smaller class as the effective sample size.

The jar fills with AUC − ε, never the raw score. A strong model on a thin sample leaves it empty, by design.