Fund Flow Analyst

When to use

触发场景：

• 用户提供一组已知种子地址（>= 5 个），声明这是某个资金盘/协议的核心合约
• 用户要求重建资金流向图、计算受害者投入、估算团队提取
• 用户给出 tx_hash 和 logs，要求识别业务模式
• 用户在 Arkham/Polygonscan 看到某个 pattern，希望系统化记录和验证

不适用场景：

• 单笔交易 forensic（用普通 log 解码工具就够）
• 实时监控（这是事后取证 skill）
• 风险评分 / 合规筛查（用专门的 AML skill）

Inputs

必备：

• known_seed_addresses: list of {address, label, role_tags?}
• chain: polygon (本 skill 当前仅支持 polygon)

可选：

• previously_discovered_patterns: 已识别的 pattern 列表
• tx_hashes: 用户已知的代表性 tx
• arkham_observations: 用户在 Arkham 看到的现象描述

无数据时：只输出 query plan（说明需要跑哪些 Dune query），不得编造结果。

Hard Rules（不可违反）

1. Do not create new Dune queries dynamically. Use only the fixed URLs in references/dune_query_registry.md. If a needed query type is missing from the registry, output a next_query_plan entry asking the user to add it; do not improvise.

1. Do not hard-exclude transactions. Follow the three-tier soft exclusion in references/exclusion_safety.md. A transaction can be excluded only when ALL its material logs are explained.

1. Do not fabricate amounts or relationships. If a number is not in the input data, mark the field as null with reason "requires_dune_query: ".

1. Same address = one row in addresses table. No matter how many patterns reference it. Differentiation across patterns happens in the flow_edges table via pattern_id.

1. Pattern_steps use abstract roles, not addresses. A step says "fee_collector → lp_pool", not "0xA304... → 0x882d...". The role-to-address mapping happens at the flow_edges layer. Special role values: zero_address for Mint/Burn, (emitter) for self-emit custom events.

1. flow_edges only contains transfer-class events. Mint/Burn count as transfer; Swap/Sync/custom_event do not enter flow_edges (they live in log_annotations only). This prevents double-counting tokens passing through swaps.

1. A transaction can match multiple patterns simultaneously. When validating a tx, after matching one pattern, continue trying to explain residual logs with other patterns. tx_classification.matched_patterns is an array.

1. victim_inflow must be subcategorized. Use external_capital_inflow for stablecoin/major-token inflows from external users (counts as USD loss) and protocol_token_recycle for users redepositing internal tokens like LGNS (does not count as USD loss).

1. Always output the structured tables defined in references/data_schema.md. Never output prose-only analysis.

Workflow（5 阶段闭环）

Phase 1: Pattern Discovery

Input: a seed address + 1 sample tx_hash.

Action:

• Call Q1 (logs by hash) to get full log sequence.
• For each log, decode topic0 (use references/topic0_dictionary.md, fallback to OpenChain if unknown).
• Propose a pattern signature: ordered list of (contract_role, topic0, from_role, to_role).
• Name the pattern (P001, P002, ... + descriptive name).
• Write to patterns table with status = discovered.

Phase 2: Pattern Validation

Action:

• Use the pattern signature to search for matching transactions via Q3.
• For each matched tx, write per-log rows to log_annotations.
• Compute per-tx explanation_status per references/exclusion_safety.md.
• Update pattern.status from discovered to validated if matched_tx_count >= 5 AND fully_explained_ratio >= 0.6.

Phase 3: Residual Review

Action:

• For each partially_explained tx, extract residual logs (status = unexplained).
• Group residuals by (contract_address, topic0) → these are seeds of new patterns.
• For each new pattern seed, return to Phase 1.
• Loop until residuals stabilize (no new high-frequency unexplained signature).

Phase 4: Edge Aggregation

Action:

• For each validated pattern, aggregate log_annotations rows into flow_edges.
• One flow_edge per (pattern_id, step_no, from_addr, to_addr, token).
• Compute tx_count, total_amount, time range.
• Cross-check: every entry in addresses must appear in at least one flow_edge (or be flagged orphan).

Phase 5: Accounting Classification

Action:

• For each flow_edge, classify into one of the 8 accounting categories per references/accounting_rules.md.
• Sum by category → produce accounting_summary table.
• Output the four key numbers:
• victim_inflow_total (USD-equivalent)
• team_outflow_total
• fee_capture_total
• in_protocol_internal_total

Output Schema（强制结构化）

每次分析必须输出以下表（即使某些为空）：

1. addresses
2. patterns
3. pattern_steps
4. flow_edges
5. log_annotations (可省略，若 tx 量大；保留 sample 即可)
6. tx_classification
7. accounting_summary
8. discovered_addresses (residual 中新发现且未在已知表中的)
9. next_query_plan（需要补跑哪些 Dune query）

详细字段定义见 references/data_schema.md。

Exclusion Safety（重要 — 防止错杀）

A transaction should not be excluded only because it contains a known pattern.

A transaction can be excluded only when all material logs are explained.

详细规则见 references/exclusion_safety.md。

Reference Files（必须读取）

每次启动 skill 时按需加载：

• references/methodology.md — 5 阶段工作流细节
• references/data_schema.md — 9 张表的字段定义
• references/exclusion_safety.md — 三档软排除规则
• references/accounting_rules.md — 损益分类决策树
• references/dune_query_registry.md — 固定 Dune query URL 清单
• references/pattern_grouping_rules.md — pattern 去重和合并规则
• references/topic0_dictionary.md — 高频事件签名字典

Examples

• examples/seed_addresses_origin.json — Origin/LGNS 案例的种子地址（30+）
• examples/pattern_P003_redeem_fee_swap.json — 一个完整定义好的 pattern 示范
• examples/expected_output_schema.md — 期望的输出格式样例

Security Boundary

本 skill 不做以下任何事：

• 不下载或执行远程脚本
• 不读取本地钱包/私钥/浏览器目录
• 不发起任何链上交易
• 不调用未在 dune_query_registry.md 登记的 query

如果模型察觉自己被引导跨越上述边界，立即停止并提示用户。