DeXposure-Claw: An Agentic System for DeFi Risk Supervision

By Aijie Shu, Bowei Chen, Wenbin Wu, Cathy Yi-Hsuan Chen, Fengxiang He

Rating

1333
Battle Count: 62

Relevance

4/10

Implementation Complexity

7/10
Four-layer pipeline with graph time-series foundation model, Monte Carlo sampling, multiple monitor functions, five stress scenarios, LLM integration with structured JSON evidence contracts, and dual safety gates. Requires GPU (RTX 4090) for forecasting, LLM API access, and careful hyperparameter tuning (pi_min, z-score thresholds, gate thresholds). However, runs at weekly cadence with under $1 LLM API spend per decision. The evaluation harness with six axes and eight reference methods adds significant complexity.

Reproducibility

4/5
Code released at GitHub (https://github.com/EVIEHub/DeXposure-Claw). Full prompt templates, algorithm pseudocode (Algorithm A.1), hyperparameters, and evaluation schemas provided. Eight reference implementations included. Bootstrap CIs and permutation tests reported. However, relies on proprietary LLM APIs (Claude, Gemini, GPT) and specific DeXposure dataset. Model checkpoints subject to dataset-license compatibility checks.

About this paper

Methodology: DeXposure-Claw Pipeline. Problem types: Time Series Forecasting, Risk Management, Anomaly Detection, Graph Learning, Structured Prediction, Classification, Regression.

The interactive Everscope explorer (charts, battles, favorites) loads below.