JunpengLiang/Onchain-EA-Template

CAIA On-Chain Investigation Benchmark

CAIA evaluates whether tool-using agents can investigate Ethereum activity from raw JSON-RPC evidence. The public viewer separates objectively scored questions from open-ended investigation questions.

Configurations

Configuration	Public input	Output style	Primary capability
objective_questions	Transaction or block anchor	A–E choice plus evidence	Classification and exact facts
subjective_questions	Transaction, block, or range	Auditable report	Forensics and open discovery
diagnostic_questions	Confirmed event or transaction	A–E choice or exact fill-in plus evidence	Failure localization

The test rows intentionally contain no answer keys. Gold labels, accepted evidence, scoring atoms, and adjudication notes are held outside this public repository.

Core objective-question design

The initial objective set is deliberately stratified along only two dimensions:

1. input_transparency: how much case-specific information is given to the agent before it starts investigating;
2. temporal_scope: how many blocks the agent must search or use to reconstruct context.

Current input_transparency levels are:

Value	Information disclosed
prepackaged_evidence	Anchor transaction plus packaged raw transaction, block, receipt, and log records
transaction_hash_only	One transaction hash; all context must be retrieved by the agent
address_and_block	One address and one block, without a target transaction
block_only	One block number, without a target address or transaction
transaction_hash_and_block_range	One transaction hash plus a bounded multi-block window
address_and_block_range	One address plus a bounded multi-block window

Current temporal_scope levels are single_block and long_range_64_blocks. span_blocks stores the exact inclusive block count.

The first release contains the following controlled comparisons:

Case family	Input transparency	Temporal scope	Purpose
FAMILY-SANDWICH-22029771	Packaged evidence vs. transaction hash only	Single block	Measure the cost of retrieving and organizing context
FAMILY-BLOCK-18563314	Address + block vs. block only	Single block	Measure the value of an entity-level search hint
FAMILY-RANGE-25360664-25360727	Transaction + range vs. address + range	64 blocks	Compare precise and broad anchors under a long search window

These rows are paired within a case_family_id; they must remain in the same split to prevent case leakage. This is an initial coverage matrix, not yet a complete Cartesian product of all transparency and span levels.

Evaluation protocol

Agents may use only the tools and sources declared in each row. A fresh session must be used for every question. Questions sharing a case_family_id must stay in the same data split, and a diagnostic version of a case must never be shown before its open-discovery version.

Objective questions

Each row visibly contains a question and option_a through option_e, similar to conventional multiple-choice benchmark datasets. The agent may still need to query raw chain data before choosing an answer. Evidence is required so that a lucky guess is distinguishable from a grounded answer.

Objective rows additionally expose input_transparency, temporal_scope, block_from, block_to, and span_blocks. If provided_files is non-empty, every referenced evidence file is part of the public dataset. File names and questions avoid revealing the hidden answer or transaction roles.

Recommended metrics:

• category accuracy;
• related-transaction set F1;
• entity-role F1;
• evidence validity;
• calibration score.

Subjective questions

Subjective rows have no answer options. They range from anchored transaction forensics to open block and block-range investigation. Hard open-world cases do not disclose a target transaction, address, asset, protocol, category, or expected number of findings.

Diagnostic questions

Diagnostic rows intentionally disclose more context and test one downstream skill at a time. The initial rows test transaction-role assignment, localization of a core anomalous transaction, identification of the directly affected funds contract, and identification of the final on-chain asset receiver. They are reported separately and do not count as open discovery.

Diagnostic questions may be multiple-choice or exact fill-in. Address answers are normalized as case-insensitive 20-byte hexadecimal values; transaction hashes are normalized as lowercase 32-byte hexadecimal values. Evidence remains mandatory, so an exact string match without a valid chain reference does not receive the evidence point. A receiver-role label describes observable fund flow and does not by itself establish real-world identity or intent.

Recommended metrics:

• reference-event recall and rank;
• adjudicated precision;
• entity and relation accuracy;
• mechanism and impact reconstruction;
• evidence, provenance, coverage, and calibration.

The open track may use reference_event_partial ground truth. An unlabelled agent finding is not automatically a false positive; novel findings require blind adjudication.

Public row schema

Common fields include:

• question_id, case_id, and case_family_id;
• chain and anchor information;
• the exact user-visible question;
• answer format and, for objective rows, option_a through option_e;
• allowed tools and forbidden sources;
• disclosure level, difficulty, and ground-truth mode;
• optional task-specification and output-schema references.

Reusable JSON Schemas for agent submissions are stored in schemas/.

Leakage controls

• No gold_answer, correct_option, reference event, victim, attacker, profit, or accepted-evidence field is present in this repository.
• Hidden fields must not be committed merely because they are omitted from the Dataset Viewer; every file in a public repository is downloadable.
• Case-level rather than question-level splitting is required.
• Evaluation should run in fresh isolated workspaces without prior case history.

Data and ethical limitations

On-chain evidence establishes transaction behavior, not necessarily the real-world identity or intent of an address controller. Labels should describe observable mechanisms and should distinguish direct observations, calculations, inferences, and unresolved claims.

Version

This is a publication template for benchmark version 0.1.0. The included rows are initial evaluation cases and do not yet constitute a statistically complete benchmark release.