DeFi On-Chain Analytics

Core Principle: 「先固定資料可信度與上下文，再做最小足夠的讀取，之後才做歸因與敘事。」 First fix data confidence and context, then do minimum sufficient reads, then do attribution and narrative.

Every analysis session serves this hierarchy: confidence > efficiency > interpretation.

Two-Layer Architecture

Every step is tagged with its required tier:

Default = Tier A only. Higher tiers are opt-in. If unavailable, disclose the gap — never silently skip.

Looped Workflow

Real investigations evolve — discovery changes the question. The workflow supports iteration, not just linear execution.

digraph phases {
  rankdir=LR;
  P0 [label="Phase 0\nScoping" shape=box style=filled fillcolor="#ffcccc"];
  P1 [label="Phase 1\nDiscovery" shape=box];
  P2 [label="Phase 2\nCollection" shape=box];
  P3 [label="Phase 3\nInterpretation" shape=box];
  P4 [label="Phase 4\nSanity Check" shape=box];
  P5 [label="Phase 5\nSynthesis" shape=box];
  Rescope [label="Rescope?" shape=diamond style=filled fillcolor="#ffffcc"];
  P0 -> P1 -> P2 -> P3 -> Rescope;
  Rescope -> P0 [label="new question\nemerged" style=dashed];
  Rescope -> P4 -> P5 [label="no"];
}

Why no skipping phases: Phase 0 prevents wasted RPC calls on wrong targets or wrong chains. A single eth_getLogs to the wrong address can eat your entire rate limit. Lock scope first, then query.

Scope Amendment (Re-entry)

When Phase 3 interpretation reveals a new analytical question that changes the objective, the investigation loops back to Phase 0 in amendment mode. Without strict rules, this loop degenerates into "ask again" — so the state machine below governs every re-entry.

Valid rescope triggers:

NOT a rescope (handle in-phase):

Carry-forward rules: On rescope, the following are preserved unchanged unless explicitly overridden:

• Chain, anchor policy, anchor block, capability tier, RPC endpoint
• All existing evidence register entries
• All pattern files already loaded

Only these fields may change: target, objective, unit of analysis, hypothesis, timeframe, additional pattern files.

Completion criteria per mode:

Hard cap: Max 3 scope amendments per session. Beyond that, synthesize what you have and list remaining questions as Open Questions in Phase 5.

Phase 0: Scoping Gate — Active Consultation

This phase is a guided conversation, NOT a passive form. Read references/scoping-guide.md for detailed consultation techniques, depth/angle options, field-by-field asking guidance, and anti-patterns. Silently assuming scope details is the #1 cause of wasted RPC calls. Surface your assumptions as explicit questions.

Analysis Modes

If the user's request clearly maps to one mode, propose it rather than asking from scratch.

Required Fields

Unit of analysis — Declare what object is being analyzed: wallet / contract / vault / pool / protocol / token. This prevents scope drift by making the analytical focus explicit. When the unit changes mid-investigation, that's a scope amendment trigger.

Anchor Policy Options

Decision Dependencies

Some conditions — not just missing data sources — can flip the analytical conclusion entirely:

If any decision-critical condition is unverified, flag it as: ⚠️ DECISION-CRITICAL GAP: [condition] unverified. Conclusion may reverse.

Blind Spot Disclosure

Before confirming, proactively flag what the analysis CANNOT see. See references/scoping-guide.md for templates.

Confirmation Gate

Present a structured summary before proceeding. Skipping confirmation