ONCHAIN REPUTATION IS MORE THAN A SCORE.
@zerufinance zScore research highlights a key distinction: reproducing a formula accurately doesn’t prove it predicts real-world behavior.
The system analyzes 11,722 protocol-chain pairs, 30+ categories, and 5.2M wallets. Its production scorer uses nine scorers, 52 components, and ~100 calibrated constants.
Three generations:
🔹 February 2025: Wallet clustering and neural networks.
🔹 Uniswap research: 91.79% of liquidity-provider and 90.83% of swap predictions fell within 50 points of targets.
🔹 September 2026: zScore-N distills the production formula to approximate scores from 0–1,000.
Key results: ▪️0.58 RMSE, R² 0.99997. ▪️ Maximum error: 16.4 points across 13,184 wallets. ▪️ With 10% missing features, output drift fell from 51.4 to 17.9 points, though clean-data error was higher.
Yet replication isn’t predictive proof.
An Aave analysis found 75%+ of liquidations among 5,000+ wallets involved scores below 300. ZAPs simulations reported 30–90% less adversarial reward capture.
These signals need independent validation, transparent features, and reproducible testing.
MY TAKE
zScore-N shows strong formula-replication accuracy. The real test is whether zScore reliably predicts repayment, Sybil activity, and malicious behavior on independent data.
A model can replicate a formula perfectly and still inherit its weaknesses.
Better decisions—not just accurate scores—are the real benchmark.
2590221760



















