Grant Application: Structural Detection of MEV-Related Behavior in CoW Auctions
Grant Title
Grant Application: Structural Detection of MEV-Related Behavior in CoW Auctions
Author
- Name: Uri Yacobi Keller / Simplex TDA
- Personal GitHub: @urihamster
- Simplex TDA GitHub: Simplex-TDA
Experiences and Qualifications
I am a mathematician and data scientist with 8+ years of industry data science and engineering experience, holding a Bachelor’s in Mathematics and a Master’s in Economics.
Through Simplex TDA, I have been researching Topological Data Analysis (TDA) applied to blockchain transaction structure. This proposal builds directly on our recent Ethereum MEV detection research, where we investigated whether TDA can detect sandwich attacks from transaction structure and how early such signals emerge during block construction.
This work is fully open source and was published as a Pre-FRP on the Flashbots forum. This proposal extends that work to CoW Protocol’s richer auction and solver-competition data, testing whether structural signals can identify MEV-related behavior before settlement.
Grant Category
Research / MEV / Analytics
Grant Description
CoW Protocol exposes unusually rich information about solver competition: batch auctions, user orders, submitted solver solutions, winning solutions, JIT liquidity orders, and on-chain settlements.
We propose to investigate whether this information can be represented structurally and used to identify MEV-related and anomalous execution behavior before settlement.
The initial focus will be sandwich-like behavior because it provides a concrete and independently labeled starting point. The broader objective is to determine whether structural representations of CoW auctions can reveal execution patterns that are difficult to capture with conventional transaction-level or tabular features.
What We Have Already Done
This proposal builds on an existing, validated Ethereum MEV research project rather than starting from an untested hypothesis.
- Baseline Ethereum MEV Research
Using EigenPhi’s DeFi-Behavior labels, we investigated whether Topological Data Analysis could identify sandwich attacks from transaction structure:
- On canonical Ethereum blocks, TDA alone achieved 0.943 ROC-AUC.
- Combining TDA with additional structural and tabular features reached 0.962 ROC-AUC.
- In an incremental evaluation, the model achieved 0.884 ROC-AUC using only the first 10% of block transactions, suggesting that structural signals can emerge early during block construction.
This is just a summary, you can find the full report and code at the repo on Github.
- CoW Protocol Validation Pilot
We have begun testing whether this approach transfers to CoW’s auction-based execution environment. In an initial study covering 3,824 CoWGPv2Settlementtransactions:
- A structural same-sender bracket heuristic identified 84 candidate settlements (2.2%).
- Rather than treating candidates as true labels, we reconstructed a stratified sample using CoW Orderbook data and Ethereum execution data, evaluating:
- Transaction ordering
- Asset overlap
- Pool/venue overlap
- Economic direction (detecting entry and exit round-trips)
- Among 31 sampled candidates, 2 were classified as confirmed sandwiches and 1 as probable. Among 30 structurally matched non-candidate controls, 0% landed on sandwich or probable.
- Correcting for sampling proportions, approximately 4.3% of bracket candidates were confirmed sandwiches (~7.9% confirmed or probable). The known EigenPhi-labeled sandwich in the sample was independently recovered by this pipeline.
This gives us both a reproducible candidate-generation and validation methodology, alongside evidence that genuine sandwich events can be identified in CoW’s settlement data at a workable scale (~100 confirmed events per month at current volume).
The Research Question
The central question is not simply whether we can identify sandwiches after the fact. Building on our Ethereum MEV research, it is:
Can structural (Topological) signals identified in our Ethereum MEV research be transferred to CoW auctions and solver activity, and can they identify MEV-related behavior early enough to be useful before settlement?
Our preliminary topological analysis is encouraging: while $n=3$ confirmed/probable cases is too small to train a classifier, those cases occupy unusually high percentiles of existing topological measures, including $H_1$ Betti numbers and persistence entropy.
The proposed grant will turn this initial result into a properly labeled research dataset and test whether structural signals emerge in CoW’s auction and solver structure early enough to be detected prior to settlement.
Proposed Work
Milestone 1 — Build and Validate a Larger Labeled Dataset
- Duration: 4 weeks
- Scope: Extend the current dataset to a substantially larger historical sample of CoW settlements. The candidate-generation heuristic will remain frozen. Candidate cases will be reconstructed using CoW Orderbook and Ethereum execution data.
- Validation Categories: Confirmed Sandwich, Probable MEV, Not Sandwich, Ambiguous.
- Deliverables:
- Reproducible candidate-generation pipeline
- Larger labeled candidate/control dataset
- Validation methodology & documentation
- Public research report
- Success Criteria: A sufficiently large and reproducible labeled dataset to support out-of-sample statistical evaluation of structural features.
Milestone 2 — CoW-Specific Structural Representations
- Duration: 4 weeks
- Scope: Adapt the structural and TDA approach developed in our Ethereum research to CoW’s auction environment, incorporating CoW Orderbook data alongside on-chain execution data.
- Graph & Topological Primitives: Map relationships between orders, solvers, submitted solutions, prices, executions, JIT orders, and settlement outcomes. Compare TDA-derived features against conventional tabular and graph baselines.
- Deliverables:
- Open research implementation of the CoW structural representation pipeline
- Evaluated set of structural primitives
- Baseline comparison against conventional features
- Out-of-sample evaluation on the validated dataset
- Success Criteria: Demonstrate whether structural representations provide measurable predictive information beyond conventional transaction and auction features.
Duration: 4 weeks
- Scope: Reconstruct progressively earlier states of the auction:
Auction Opens → Orders → Early Solutions → Late Solutions → Close → Winner → Settlement - Evaluation: Evaluate models using only information available at each specific stage to analyze how early structural signals emerge.
- Deliverables:
- Pre-settlement evaluation & timing analysis report
- Final research report
- Reproducibility materials and research code
- Success Criteria: Determine whether CoW auction structure contains predictive information about subsequent MEV-related or anomalous execution behavior before settlement.
Timeline & Budget
- Total Duration: 12 weeks from commencement
- Total Funding Requested: $15,000 USD (or equivalent in xDAI / USDC)
| Milestone | Description | Duration | Funding |
|---|---|---|---|
| Milestone 1 | Build and validate a larger labeled dataset | 4 weeks | $5,000 |
| Milestone 2 | CoW-specific structural representations | 4 weeks | $5,000 |
| Milestone 3 | Pre-settlement detection | 4 weeks | $5,000 |
| Total | 12 weeks | $15,000 |
- Commencement: Upon grant approval by the CoW Grants Committee and assigned Steward onboarding.
- Budget Allocations: Data acquisition/processing, auction reconstruction, graph/TDA representation development, statistical evaluation, and open-source documentation. No COW token allocation is requested.
Expected Impact
CoW Protocol provides deep transparency into batch auctions and solver competition. The proposed research investigates whether that auction data can be mapped into structural representations to detect MEV-related and anomalous execution patterns before settlement.
Key Potential Ecosystem Benefits:
- Pre-Settlement MEV Signal: If successful, structural signals could provide an early indication of MEV-related conditions to solvers, block builders, and execution infrastructure before the full transaction history is observable.
- Reproducible Labeled Benchmark: Produces a validated dataset of CoW MEV-related instances to support further research into intent-based MEV.
- Solver Quality & Execution Assurance: Provides structural monitoring primitives to help the CoW ecosystem analyze solver behavior and identify potentially anomalous execution patterns.
- Open Research Infrastructure: Makes research methodology, reproducibility materials, and selected tooling available to the broader Ethereum research community.
Additional Information
The project is intended to produce reproducible research and tooling that can be used by the CoW and broader Ethereum research communities. Downstream production systems and commercial applications may be developed separately.
Gnosis Chain / EVM Address
0x4933FbB6943f39CE47FEb48A5606597689Dd2503
Terms and Conditions
By submitting this grant application, I acknowledge and agree to be bound by the CoW DAO Participation Agreement and the CoW DAO Grant Agreement Terms.
Community Feedback & Technical Details
We welcome questions, feedback, and discussion from the CoW Grants Committee and the broader community. If you would like to review the specific mathematical setup, preliminary data sampling methods, or topological feature extraction code used in our initial evaluation, please reply below or reach out directly—we are happy to share additional technical details!
