evm-analyst

Analyze Polygon EVM fund-flow patterns from seed addresses, tx hashes, and Dune log outputs. Use fixed registered Dune queries only.

Fund Flow Analyst

Use this skill when the user provides Polygon seed addresses, transaction hashes, decoded logs, raw polygon.logs rows, or Dune query results and asks for fund-flow reconstruction, pattern annotation, or accounting classification.

Hard rules

1. Use only registered Dune queries in references/dune_query_registry.md. Do not invent query IDs or URLs.
2. Do not fabricate amounts, addresses, relationships, labels, or transaction counts.
3. Do not hard-exclude a transaction merely because it contains a known pattern. Exclude only when all material logs are explained.
4. Same address appears once in addresses; pattern differences belong in flow_edges and log_annotations.
5. flow_edges contains only transfer-class events, including mint/burn. Swap, Sync, Approval, and custom events stay in log_annotations unless they directly transfer value.
6. One transaction may match multiple patterns.
7. Separate external_capital_inflow from protocol_token_recycle.
8. If data is insufficient, output next_query_plan instead of guessing.
9. For topic0 decoding, use references/topic0_dictionary.md when available. Unknown topic0 should be marked unknown_event_signature.
10. Preserve raw evidence: include sample tx hash, log index, contract address, topic0, token address, and decoded from/to/amount where available.

Required output

Keep results concise and structured. Prefer these tables when relevant:

• addresses: address, role_guess, evidence, confidence
• patterns: pattern_id, pattern_name, description, confidence
• pattern_steps: pattern_id, step_no, event_type, contract_address, topic0, meaning
• flow_edges: tx_hash, log_index, from_address, to_address, token_contract, amount_raw, amount_normalized, edge_type
• log_annotations: tx_hash, log_index, contract_address, topic0, decoded_signature, interpretation
• tx_classification: tx_hash, classification, reason, unresolved_items
• accounting_summary: bucket, token_contract, gross_in, gross_out, net, notes
• discovered_addresses: address, first_seen_tx, reason_to_follow
• next_query_plan: query_code, purpose, parameters_needed

Do not output prose-only analysis.