zScore-N: A Neural Network for On-Chain Wallet Reputation Scoring
Girish G N, Ashutosh Sahoo, Akshay SP, Gurukiran S, Dhanashekar Kandaswamy
cs.AI
Quick verdict
A NN distills a proprietary DeFi wallet-reputation formula to 0.58 RMSE; rigorous protocol, but the score itself is unvalidated and nothing is reproducible.
Summary
The paper presents zScore-N, a neural network that has replaced Zeru AI's hand-written wallet-reputation formula in production. That formula — nine scorers, 52 component scores, roughly a hundred calibrated constants, all proprietary — serves as a deterministic teacher, verified to reproduce all 65,919 stored production scores to within 2.27e-13 before any training. Using smooth activations, explicit missingness indicators, and a robust regression loss, the network reaches 0.58 points RMSE on the 1000-point scale (R²=0.99997, max error 16.4) on a 13,184-wallet held-out split, versus 2.25 for gradient-boosted trees and 28.04 for linear regression on identical features and splits. A mask-augmented variant trained against uncorrupted targets roughly halves degradation under injected zero-encoded missingness (17.9 vs 51.4 points RMSE at 10% missingness) and removes the formula's systematic downward bias, at the cost of a 7.96-point floor on clean data. Three seeds span only 0.157 RMSE points. The limitations are structural: the target function and evaluation export are private, so no result can be externally verified; fidelity to a formula is not evidence that the score is valid, and no downstream outcome (repayment, sybil detection) is evaluated; injected missingness is per-feature independent rather than the correlated failures real pipelines produce; and no tree baseline with native missing-value handling appears in the robustness comparison. No code or data is released.
Review complete
Assessment at a glance
5
risk flags identified
● high 2
● medium 2
● low 1
5 strengths · 5 weaknesses · 6 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- A neural network distilled from the proprietary zScore formula reproduces it to 0.58 points RMSE (R²=0.99997) on 13,184 held-out wallets, versus 2.25 for gradient-boosted trees and 28.04 for linear regression on identical features and splits.
- The teacher was verified exactly before training — the reimplementation matches all 65,919 stored scores to 2.27e-13 max error — so all disagreement belongs to the student.
- Mask-augmented training with explicit missingness indicators roughly halves error under injected zero-encoded missingness at every rate from 5% to 40% and removes the formula's systematic downward bias, at the cost of a 7.96-point clean-data floor.
- The deployed system serves the score inline (24 ms scoring, 142 ms median per wallet) with unchanged partner API, deterministic output verified to 1e-9 across a language migration, and seed-stable retraining (0.157-point spread).
- Smooth activations beat ReLU and tanh even when approximating a clamped, piecewise teacher (Table 2: 0.763 vs 0.939 vs 1.158 RMSE under a fixed budget).
Strengths
Pre-declared evaluation protocol
Section 4 fixes every metric, baseline, split and failure stratum before results; one wallet-level split was materialised once and reused by all models, with scalers fitted on train only and answers excluded from inputs (Section 4.2).
Exact teacher verification
Section 5.1: the reimplemented scorer matches all 65,919 stored scores to 2.27e-13 max / 2.5e-16 mean error, and the paper explicitly notes this is equivalence of implementation, not evidence of accuracy — cleanly attributing all student error to the network.
Honest cost accounting
Stability and stratified reporting
Measured production grounding
Weaknesses
Total irreproducibility
Circular validity
Robustness evidence gap
Deployment-fidelity ambiguity
Thin uncertainty quantification
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
This is an industrial method paper from Zeru AI describing the replacement of its proprietary on-chain wallet-reputation formula (zScore) with a distilled neural network, zScore-N, now in production. The claims are precise and consistently scoped: the network reproduces the formula to 0.582 RMSE points on the 0–1000 scale (R² = 0.99997, max error 16.4) on a 13,184-wallet held-out partition, versus 2.245 for gradient-boosted trees and 28.042 for linear regression on identical features and splits (Table 1); mask-augmented training roughly halves error under injected zero-encoded missingness at every rate from 5% to 40% (Table 4: 17.9 vs 51.4 RMSE at 10%); the network is differentiable everywhere and distinguishes an unobserved feature from a true zero, which a clamped arithmetic sum cannot. The paper does not claim the score itself is valid — Section 5.1 explicitly says teacher verification 'is equivalence of implementation, not evidence of accuracy.' That honesty is welcome, but it concedes that every headline number measures agreement with an unvalidated target.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 6 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is zScore-N?
A neural network built by Zeru AI that reproduces its hand-written on-chain wallet reputation formula (zScore). It was trained on millions of the formula's own outputs and now serves the 0–1000 reputation score in production behind an unchanged partner API.
How accurate is zScore-N compared to the formula it replaced?
On complete data it matches the formula to 0.58 RMSE points on the 1000-point scale (R²=0.99997), beating gradient-boosted trees (2.25) and linear regression (28.04). The missingness-robust variant accepts a 7.96-point error floor on clean data in exchange for roughly halving the error when input features are missing.
Is the zScore-N code or dataset public?
No. The feature definitions, the formula's constants, the evaluation export of 65,919 scored wallets, and the model details are proprietary; the paper releases no code, weights, or data.
Why replace a hand-written formula with a neural network for wallet scoring?
The paper gives three reasons: the network is differentiable everywhere (so score sensitivity to behaviour becomes computable), it can distinguish a genuinely-zero feature from one the pipeline failed to capture via missingness indicators, and it can keep improving as more wallets are observed, unlike a fixed set of constants.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.