zScore-N: A Neural Network for On-Chain Wallet Reputation Scoring

Girish G N; Ashutosh Sahoo; Akshay SP; Gurukiran S; Dhanashekar Kandaswamy

Source-linked AI summary

zScore-N: A Neural Network for On-Chain Wallet Reputation Scoring

Girish G N, Ashutosh Sahoo, Akshay SP, Gurukiran S, Dhanashekar Kandaswamy

arXiv:2609.08247v1Sep 8, 2026cs.AI

TL;DR

Hand-authored wallet scoring is a large, piecewise system that must first interpret untyped transaction history, limiting differentiability and reliable treatment of incomplete records. zScore-N distills this deterministic scorer into a smooth neural network, reproducing its outputs closely while reducing systematic missing-data error and serving at production scale.

Problem
Hand-authored wallet scoring combines many bounded nonlinear components into a piecewise function, while raw transaction history requires substantial protocol and sector classification before scoring.
Method
zScore-N uses a smooth, end-to-end differentiable architecture trained by distillation from a deterministic scorer that can label wallets on demand without label noise.
Results
0.58 points: the network reproduces the scorer more closely than gradient-boosted trees and linear regression, while mask-augmented training halves missing-data error and removes systematic downward bias.
Takeaways & Limitations
The network replaces the formula in production, serving inline at a median of 142 milliseconds per wallet across live token-distribution and lending deployments.
Takeaways & Limitations
The augmented model cannot recover information that was never captured, and robustness has a clean-record fidelity cost of 7.96 points versus 0.86 for the plain model.

Abstract

from arXiv · show

Wallet reputation scores decide who receives an airdrop, who can borrow, and who enters an allowlist across decentralised finance. They almost always begin as hand-written formulas: compositions of clamped logarithmic, linear and square-root transforms over behavioural features, with every threshold and point award set by hand. Such a formula is readable and deterministic, but it is piecewise and non-differentiable, it cannot improve as data accumulates, and it cannot distinguish a feature that is genuinely zero from one its pipeline failed to capture. We present zScore-N, the neural network that replaced ours in production. The formula served as its teacher: calibrated against 5,208,952 wallets sampled across 2019-2024 and verified to reproduce production output to within 2.3e-13, it supplies unlimited labelled training data at zero label noise. The trained network reproduces it to 0.58 points RMSE on the 1000-point scale (R^2 = 0.99997), against 2.25 for gradient-boosted trees and 28.04 for linear regression on identical features and splits. Trained with missing-value masks against uncorrupted targets, it halves the error that incomplete data introduces: at 10% feature-level missingness the formula drifts 51.4 points from its own complete-data output with a systematic -12.5 point bias, while the network drifts 17.9. The network carries the score at production scale, across a population of millions of wallets spanning six orders of magnitude in size and activity.

1 Introduction

Wallet reputation scoring must operate over a large, skewed population and reconstruct sparse on-chain histories, while hand-authored formulas are difficult to reason about and conflate missing values with genuine zeros. zScore-N replaces that formula with a differentiable neural scorer distilled from it, achieving high fidelity and improved missing-data robustness.

Motivation: 5,208,952 wallets spanning 2019–2024 calibrated the production scoring function across a population whose median and 99th-percentile wallets differ by six orders of magnitude in value and activity.The active-address population includes 9,266,185 Ethereum addresses in a three-month window, and one production export scores 311,815 wallets end to end.
Motivation: 1,066,071 transfer events across 55 chains for 157 wallets illustrate the sparse, multi-chain reconstruction required before behavioural features can be scored.Raw contract calls must be resolved into protocols and actions before a feature exists.
Motivation: A large calibrated formula combines bounded component awards and nonlinear transforms into a piecewise output with no derivative at its boundaries.Its line-by-line readability does not make aggregate behaviour easy to reason about.
Motivation: At 10% feature-level missingness, the formula is 51.4 points RMSE from its complete-data output and carries a −12.5-point systematic bias because missing values are encoded as zero.In the production export, 65.2% of wallets have an unrecorded value in one input field and 23.5% in another.
Approach: zScore-N uses the formula as a deterministic teacher, adds explicit missingness indicators, and is trained against uncorrupted targets.The teacher supplies unlimited labelled examples at zero label noise after production output was reproduced within 2.3 × 10−13.
Results: 0.58 points RMSE and R^2 = 0.99997 show that zScore-N reproduces the formula more closely than gradient-boosted trees at 2.25 and linear regression at 28.04.All methods use identical features and splits.
Results: Mask-augmented training halves the error introduced by missing features at every tested rate from 5% to 40% and removes the formula’s systematic downward bias.The model also provides a defined gradient everywhere, making score sensitivity to behaviour computable.

2 The Scoring Formula

The production scorer is a large, proprietary system built over reconstructed wallet histories and calibrated to a highly skewed population. Its determinism and calibration make it a strong, zero-noise teacher for the neural replacement.

2 The Scoring Formula: Feature definitions, component structure, thresholds, point awards, and transform mappings are proprietary and withheld.The paper describes the scorer’s scale and character rather than its full specification.
2.1 What has to happen before a wallet can be scored: Raw wallet histories require resolving untyped contract calls into protocols and economic sectors across dozens of chains before scoring is possible.This mapping is continuously maintained against a long tail of contracts lacking public classification.
2.1 What has to happen before a wallet can be scored: Temporal wallet behaviour requires reconstructing timelines and matching events across each wallet’s entire life, because aggregates alone omit holding periods, obligations, exits, and recent failures.The reconstruction is performed per economic sector.
2.2 The scale of the function: The scorer comprises nine scorers, 52 component scores, five nonlinear-transform families, and roughly a hundred calibrated constants.Each sector has its own feature vocabulary and contributes to a single bounded output.
2.3 Calibration: 5,208,952 wallets sampled across twelve months from 2019–2024 supplied percentile-based calibration for the scorer’s saturation points.The calibration reflects a distribution whose median and 99th-percentile wallets differ by six orders of magnitude.
2.4 Why it makes a strong teacher: Determinism lets the formula label any wallet repeatedly at zero label noise while preserving distributional structure learned from more than five million wallets.This makes it a strong supervision source despite the difficulty of reasoning about the formula in aggregate.

3 The Network

zScore-N is designed to replace the production scorer while preserving bounded outputs, differentiability, and fidelity. It uses explicit missingness indicators and deterministic-teacher training to learn from incomplete records without corrupting the target.

3.1 What the model had to do: The design requirements were scorer-level fidelity, differentiability everywhere, explicit separation of absent and zero features, and preserved score bounds.These requirements rule out large score shifts, nondifferentiable alternatives, arithmetic-only missingness handling, and out-of-range outputs.
3.2 Input representation: Explicit binary missingness indicators let the network distinguish genuine inactivity from fields the pipeline failed to capture.The behavioural features and indicators enter as separate inputs.
3.2 Input representation: Answer leakage is prevented by excluding the final score, component scores, and assigned tag from the model inputs.These values appear only as targets or stratification variables.
3.2 Input representation: Heavy-tailed quantities are compressed before training, while normalisation statistics are fitted on the training partition and reused unchanged elsewhere.This structurally prevents full-dataset scaling leakage.
3.3 Shape: The architecture widens a conditioned input into a high-dimensional representation, holds that width to form interactions, then tapers to one bounded score.The taper compresses learned structure into factors that move the score, with depth and width selected by systematic sweep.
3.3 Shape: Smooth, infinitely differentiable activations make ∂score/∂feature continuous across the input space, while an unsquashed linear output preserves extreme-score resolution.Bounds are imposed explicitly rather than produced by saturation.
3.4 Training against a deterministic teacher: The deterministic teacher supplies unlimited labels at zero noise, so fidelity is principally bounded by model capacity and optimisation.Evaluation is stratified by difficult regions to expose failures hidden by aggregate agreement.
3.4 Training against a deterministic teacher: Mask augmentation withholds features while retaining complete-record targets, and two variants separately optimise clean-record fidelity and production incomplete-record accuracy.A robust regression loss is quadratic near zero and linear in the tail, limiting the influence of extreme wallets.

4 Evaluation Setup

The evaluation fixes the data, splits, inputs, metrics, and corruption procedure before comparing zScore-N with linear regression, boosted trees, and the production formula. It measures fidelity, regional failures, run stability, and missing-data bias under identical conditions.

Evaluation design: Every metric, baseline, split, and stratum was fixed before results were computed.This declares the scoreboard independently of the outcome.
Evaluation design: 65,919 distinct scored wallets form the reproducible export, with raw behavioural features and deployed-formula scores and no duplicate or error rows.The export is drawn from the broader production population.
Evaluation design: The production scorer was reimplemented before training and verified against every stored score, attributing later disagreement to the network rather than the target implementation.This verification licenses the comparison.
Evaluation design: A single wallet-level train/validation/test partition was reused verbatim across the network, baselines, mask variant, and ablations.Identical partitions prevent comparison results from being confounded by partition luck.
Evaluation design: The inputs exclude answer variables and fit normalisation constants on training data alone, preventing answer and statistic leakage.Both rules are enforced in code rather than by convention.
Baselines: Linear regression measures reachable linear behaviour, while gradient-boosted trees provide the strong tabular baseline and a nondifferentiable comparison.Both alternatives use identical inputs and splits.
Robustness: Missingness is injected as zero independently into corruptible features at rates from 5% to 40%, comparing the formula, plain network, and mask-augmented network against uncorrupted formula outputs.This isolates the value of explicitly representing unknown while holding architecture, data, and split constant.
Metrics: Fidelity is reported in 0–1000 score points using RMSE, MAE, R^2, and maximum absolute error, with separate strata for floors, ceilings, sparse bands, and zero-encoded fields.Stratification is intended to reveal local failures hidden by aggregate fidelity.

5 Results

zScore-N closely reproduces the deployed scorer while exposing where smooth approximation struggles and improving robustness to zero-encoded missingness. Its missingness-aware variant reduces corruption error, but incurs a clean-data fidelity cost and cannot recover information absent from the surviving features.

5.2 Fidelity: 2.3 × 10−13 maximum absolute error verifies the reimplementation against all 65,919 stored production scores before network comparison.Every wallet agrees within 10−6, and four of five components agree bitwise.
5.2 Fidelity: 0.58 points RMSE makes the network 3.9× closer than gradient-boosted trees and 48× closer than linear regression on identical features and splits.The comparison uses the held-out test partition of 13,184 wallets and score points on the 0–1000 scale.
5.2 Fidelity: The smooth activation retains accuracy: the network wins on RMSE, while its worst-case error is less than half that of the piecewise-linear alternative.This tests the architectural choice to make the model differentiable everywhere.
5.3 Where fidelity fails: Error roughly triples at a component floor and doubles at a ceiling, with the sparse bottom score band worst; the zero-history floor is reproduced exactly.These regional failures were predicted in advance, and the reported deviation is one to two points.
5.4 Stability: Three independent initialisations span 0.157 points of test RMSE, indicating operationally equivalent minima despite non-convex training.The reported RMSEs are 0.775, 0.932 and 0.790 on the 1000-point scale.
5.5 Robustness to missing data: The augmented model cannot recover information that was never captured, so performance is bounded by what remains statistically inferable from surviving features.Where a wallet record is too thin for informative features to survive, the data—not the architecture—sets the ceiling.

6 Running in Production

The production system places the neural scorer inside a four-stage pipeline and preserves deterministic, partner-facing behavior. Measured improvements come from efficient processing and deployment design, while data movement—not scoring—dominates end-to-end latency.

Pipeline placement: The model is the third stage in a four-stage wallet-scoring pipeline.Raw transactions are read, resolved into economic categories, transformed into live sector features, and then scored.
Runtime cost: More than 99% of request time is data movement rather than scoring.In ten production runs reading 10,000 transactions per request, reading took 327 ms, feature processing 29 ms, and scoring 24 ms against a 1,044 ms median response.
Runtime cost: 608× faster scoring reduced a 10,000-transaction wallet from 23.1 seconds to 0.038 seconds.At 25,000 transactions, the improvement reached 1,767×, from 58.3 seconds to 0.033 seconds, because batch logic was removed from single-wallet scoring.
Runtime cost: 142 ms was the median full-wallet score time across 1,981 production wallets.The P95 was 1,011 ms, the worst case was 2,326 ms, and one batch run scored 2,822,413 transactions in 139.68 seconds.
Capacity: The deployed architecture was sized from measured concurrency limits rather than vendor specifications.Throughput peaked at 20 concurrent large queries, degraded at 50, and collapsed at 200 concurrent 100,000-row queries.
Determinism and monitoring: Deterministic serving preserves identical outputs across implementation changes and production runs.Migration between languages required matching every numeric field within 1 × 10−9 absolute tolerance, including summation order and rounding conventions.
Operational interface: The unchanged API contract lets partners use the score without knowing whether the formula or network produced it.The score range, tag vocabulary, endpoint set, API-key guard, quota, and rate-limit metering remain unchanged.
Determinism and monitoring: Offline versioned retraining and continuous monitoring keep model updates separate from live traffic.The system watches score distributions and empty-feature rates, while mask-augmented scoring degrades gracefully when capture rates fall.

7 Conclusion

The paper replaces a deployed wallet-reputation formula with a neural network trained on that formula’s verified outputs. The network closely reproduces the score, improves missing-data behavior, remains stable across retraining, and runs behind the unchanged production API.

Main result: The deployed neural scorer reproduces the formula to 0.58 points on a 1000-point scale.This is 3.9 times closer than gradient-boosted trees and 48 times closer than linear regression on identical data.
Main result: The network halves incomplete-record error while removing the systematic downward bias of the formula.It holds to within a sixth of a point across independent retrainings and serves at a median of 142 milliseconds per wallet.
Production scope: The system scores at population scale across Ethereum, Base, Arbitrum, and Hyperliquid EVM.It supports protocols running live token distributions and lending markets behind an unchanged API.
Evidence standard: The paper’s evidence standard includes target verification, fixed evaluation criteria, strong tabular baselines, and spread reporting.The target was verified to 2.27 × 10−13 before training, with metrics, baselines, and failure strata fixed before reporting.