Research
Bridging the Opacity: Evidence-Backed Cross-Chain Transaction Correspondence Reconstruction Across Heterogeneous Blockchains
Cross-chain bridges move assets between blockchains, but in doing so they break the transaction trail: the mapping from a source transaction to its destination transaction usually lives inside the…

- arXiv
- 2609.18158
- Published
- 2026-09-16
- Authors
- Dan Lin, Huan Xiao, Ziwei Li, Xiapu Luo, Jiachi Chen, Jiajing Wu, Zibin Zheng
AI summary
Overview
Research area: Blockchain security and cross-chain forensics (arXiv category: Cryptography and Security, cs.CR) — specifically, reconstructing the link between a source-chain transaction and its destination-chain transaction across bridges and heterogeneous ledgers.
Technical level: Advanced. The paper combines a formal problem formulation, LLM-assisted protocol-specification synthesis, byte-level parsing across EVM, Bitcoin and Solana, and a quantitative evaluation on 54,578 transaction pairs.
Scope (one sentence): The paper presents XSplicer, a system that reconstructs cross-chain transaction correspondence from public protocol documentation and on-chain artifacts alone, without privileged access to bridge-internal mappings.
What This Paper Is About
Cross-chain bridges move assets between blockchains, but in doing so they break the transaction trail: the mapping from a source transaction to its destination transaction usually lives inside the bridge operator's backend. The authors survey 131 bridge protocols and find that only 16.79% expose a complete public mapping from a source transaction hash to its destination transaction hash, which turns bridges into blind spots for investigators tracing illicit funds. The paper's goal is to recover those source-to-destination links using only public evidence — protocol documentation, public example transaction pairs, and on-chain data — and to attach explicit evidence to each recovered link.
Key Contributions
- Problem formulation. The authors formalize cross-chain transaction correspondence reconstruction (xTCR) for third-party investigators as a two-part task: candidate search bounded by destination-side constraints, followed by verification using protocol evidence shared between the two ledgers.
- System design. They design and implement XSplicer, a three-stage pipeline that compiles public protocol artifacts into a unified semantic specification and then into two executable rule operators — an Analyzer (extracts destination-side constraints) and a Verifier (checks cross-chain shared evidence) — for heterogeneous ledgers.
- Evaluation. They evaluate XSplicer on 54,578 transaction pairs across seven mainstream bridges spanning EVM, Bitcoin and Solana (Allbridge, deBridge, Wormhole, THORChain, WanBridge, Orbiter, Symbiosis), including adversarial-noise experiments and two real-world case studies.
- Forensic findings. They show that official bridge interfaces can miss valid correspondences, and that protocol-level hard evidence resolves ambiguities that time- and amount-based matching leaves behind.
Main Findings
- A measured observability crisis: Of 131 mainstream cross-chain protocols sourced from the Chainspot aggregator, only 16.79% provide a verifiable public mapping from a source transaction hash to its destination transaction hash.
- Overall recovery performance: XSplicer achieves a 92.5% global recovery rate, with per-protocol accuracy ranging from 87.92% to 98.61%. The best individual protocol result is Wormhole at 98.61% conditional accuracy.
- Coverage is high for most protocols: Candidate coverage exceeds 93% for most protocols, meaning the true destination transaction is usually inside the constructed candidate set. THORChain's candidate sets average only 2.79 transactions; Allbridge's average 9.16.
- The main coverage failure is routing complexity: Symbiosis has the lowest coverage at 87.92%, which the authors attribute to destination settlement being obscured by intermediate contract executions and routing steps.
- Hard evidence survives adversarial noise: Under adversarial noise, the hard-evidence verifier retains the correct match in 100% of tested cases, whereas soft-clue matching degrades as ambiguity increases.
- Soft clues are genuinely ambiguous, but hard evidence compensates: Wormhole has a soft-clue ambiguity ratio of 56.56% (candidates within a ±5% amount margin of the source amount) yet still reaches 98.61% conditional accuracy because the hard-evidence-first strategy filters amount- and time-only matches.
- Hard-evidence availability varies widely by protocol: The proportion of queries for which protocol-level deterministic evidence can be reconstructed from public artifacts ranges from 66.30% (Orbiter) to 92.65% (THORChain). Wormhole is 77.09%, Allbridge 91.85%, WanBridge 89.45%, Symbiosis 87.79%, deBridge 74.67%.
- Conditional accuracy per protocol: Allbridge 96.80%, THORChain 93.17%, Orbiter 91.97%, WanBridge 90.47%, deBridge 88.12%, Symbiosis 87.92%, Wormhole 98.61%. Allbridge and THORChain exceed 93% once the ground truth is in the candidate set.
- Latency: Average online latency per source transaction ranges from 46.41 seconds (THORChain) to 146.34 seconds (Wormhole); Allbridge 123.48 s, deBridge 120.42 s, WanBridge 120.54 s, Orbiter 112.11 s, Symbiosis 98.68 s.
- Official APIs are not reliable correspondence oracles: The paper documents that LayerZero's Value Transfer API requires an application for a production API key, Across's documentation advises developers against caching response data, Wormholescan offers metadata browsing rather than a verifiable forensic index, Multichain's APIs and explorer became inaccessible after it ceased operations in 2023, and Stargate's legacy API was deprecated.
- Representation gaps defeat naive matching: In a deBridge transfer from Ethereum (0x93b1...11d5) to Solana (4V7f...1QMb), the shared identifier orderId is a hex bytes32 on Ethereum but a [u8;32] array on Solana, and receiver fields are bytes32 values that must be decoded into valid Solana public keys before comparison.
- Post-shutdown recovery: After Multichain ceased operations, XSplicer recovered more than 1,900 historical transaction pairs.
- Bybit laundering case: XSplicer identified 754 illicit cross-chain transfers worth $105.6 million in the Bybit laundering incident, tracing THORChain-mediated flows.
- Decision states prevent forced answers: The system outputs Confirmed (unique hard-evidence match), Supported (soft clues only, no competitive ambiguity) or Unresolved (broken evidence chain or irreconcilable ambiguity) rather than always producing a single ranked answer.
- Scale context: DeFiLlama is cited as reporting roughly $380M in 24-hour bridge volume and $15B over one month, and the Elliptic 2025 report is cited for over $21.8 billion in high-risk on-chain assets transferred via cross-chain mechanisms.
- Not reported in the available content: The baseline comparison against existing state-of-the-art cross-chain tracing tools (RQ2), the detailed component-ablation results (RQ3), and the chain-pair breakdown (Table 3) are referenced but their numbers are not included in the excerpt provided.
Methodology in Plain English
The authors start from a practical constraint: an outside investigator — a security firm or a regulator — has no access to the bridge operator's internal database, and often no working official API either. So the system must work from what anyone can see: the protocol's public documentation, a handful of publicly known example transfers, and the raw transactions on both blockchains.
XSplicer works in three stages.
Stage I — Learn what a bridge leaves behind. The system feeds protocol documentation and known example transaction pairs to a large language model, with preprocessing to discard irrelevant material (such as SDK installation guides) and to convert raw transaction traces into structured summaries. The LLM extracts two kinds of knowledge: destination-side constraints (which chain the funds go to, what the recipient address looks like, chain IDs and address-encoding variants) and cross-chain shared evidence (identifiers, nonces or hash references that appear on both sides). A conservative merge rule keeps a field only when multiple independent sources agree; conflicting items are marked pending.
Stage II — Turn that knowledge into code. LLMs are not used at runtime, because online processing is high-volume, latency-sensitive and needs deterministic byte-level operations. Instead, a task decomposer splits the specification into constraint-recovery tasks and evidence-verification tasks, and an interface synthesizer writes two small Python modules: an Analyzer that pulls destination constraints out of a decoded source transaction, and a Verifier that extracts, normalizes and compares shared evidence across the two ledgers. Generated code is automatically validated against public examples and repaired if it fails. For THORChain, for example, the Analyzer parses the memo field for a chain marker and a Bitcoin address, and the Verifier decodes the OP_RETURN payload on Bitcoin and compares it to the source transaction hash after stripping "0x" and normalizing case.
Stage III — Recover links online. A decoder flattens raw transactions from EVM, Solana and UTXO chains into a uniform view. The Analyzer's output prunes the entire destination ledger down to a small candidate set. The Verifier then applies hard evidence first; soft clues (amount, time, address) are used only as a fallback when structural evidence is absent. The output is a decision state with an attached evidence set, not a similarity score.
Evaluation uses labeled pairs from explorer/API records or a published dataset, with public construction examples kept disjoint from evaluation pairs. Metrics are candidate coverage, an oracle-assisted conditional accuracy (to isolate verifier precision from collection failures), hard-evidence availability ratio, soft-clue ambiguity ratio (share of candidate sets containing a non-target transaction within ±5% of the source amount), and latency.
Why This Matters
Research impact. The paper reframes cross-chain tracing from "find the most similar transaction" to "recover comparability, then verify protocol invariants." It provides a reusable formulation (xTCR), a documentation-to-executable-rules pipeline, and a metric suite that separates search-space failures from verification failures — a distinction prior heuristics-based work conflates.
Real-world applications:
- Post-mortem investigations after a bridge shuts down. The Multichain case shows that when an operator disappears and its APIs go dark, on-chain evidence plus documentation can still recover over 1,900 historical transfer pairs.
- Laundering and hack tracing. The Bybit case links 754 cross-chain transfers worth $105.6 million, giving investigators followable leads across chain boundaries.
- Independent auditing of bridge claims. When an official mapping does exist, XSplicer can cross-check it; the paper reports that official interfaces can miss valid correspondences.
- Exchange and custodian compliance screening. Deposit tracing that crosses chains can be reconstructed from public data rather than relying on a bridge vendor's cooperation.
Industry relevance. The 16.79% traceability figure quantifies a structural gap in DeFi infrastructure, and the documented API volatility (key-gated access, deprecation, advisories against caching) argues that forensic tooling should not be built on bridge-operated endpoints. The three-state output (Confirmed / Supported / Unresolved) also matches how compliance and legal workflows need to distinguish proof from lead.
Future Directions
- Extending beyond the seven evaluated bridges and the three execution models. Coverage of the wider 131-protocol landscape, and of ledgers beyond EVM, Bitcoin and Solana, is an open question.
- Improving candidate coverage for routing-heavy protocols. Symbiosis's 87.92% coverage, caused by intermediate contract executions and routing steps, points to a need for analyzers that follow multi-hop settlement paths.
- Raising hard-evidence availability. Orbiter's 66.30% hard-evidence availability ratio means a large share of queries fall back to soft clues or end as Unresolved; finding additional public invariants for such protocols is a direct next step.
- Making specification synthesis more robust and maintainable. The pipeline depends on LLM extraction validated by consensus and automatic repair; how well it tracks protocol upgrades, thin documentation, and adversarial documentation remains open, as does reducing residual dependence on soft clues in ambiguous candidate sets.
Target Audience
Blockchain security researchers and graduate students working on transaction tracing and interoperability; forensic analysts at security firms and analytics vendors; regulators and law-enforcement technical staff who need defensible, evidence-attached links rather than similarity scores; and bridge and exchange engineering teams interested in what their public interfaces do and do not reveal. Readers should be comfortable with how EVM, UTXO and Solana execution models differ, since the paper's core argument rests on that heterogeneity.
Authors’ abstract
Cross-chain bridges enable interoperability, but they also break the transaction trails needed to trace illicit funds. Third-party investigators typically cannot access the source-to-destination mappings maintained by bridge backends, and our survey of 131 bridges finds that only 16.79% provide complete public tracking. Existing approaches depend on official APIs, EVM-specific assumptions, or fragile temporal heuristics, limiting their ability to trace transfers across…
The full abstract on arXiv