Mercury Method Lab：构建可审计的AI证据链与决策系统

Mercury Method Lab addresses the critical need for traceable, auditable AI decision-making in complex scenarios. As a method lab for the Mercury Agent ecosystem, it provides structured evidence management, audit tracking, and decision log systems—ensuring AI judgments are verifiable through strict state flows and constraints. This post breaks down its core components, processes, and value.

With AI Agents increasingly used in complex decisions, the lack of traceability and auditability has become a key issue. Mercury Method Lab is not a runtime component but a 'method, evidence, audit, and migration lab' compatible with Mercury Agent. It complements the upstream Mercury Agent (which handles runtime, interfaces, permissions, and schedulers) by focusing on method routing, evidence chain management, artifact tracking, memory candidate selection, decision logging, action plan generation, and audit reports. The division: Agent does 'what to do', Lab does 'how to record and verify what was done'.

The Lab defines a 7-step data flow: inbox → raw → segmented → cleaned → uncertain → memory_candidates → decision_logs/action_plans/audit_reports. Each stage has clear input/output rules:

• inbox: Pre-stored materials with potential value.
• raw: Immutable original inputs (no direct overwriting).
• segmented: Split complex materials into manageable units.
• cleaned: Structured data without noise.
• uncertain: Suspicious info pending verification.
• memory_candidates:筛选ed info for long-term retention.
• Final outputs: Decision logs, action plans, audit reports.

Data flow is managed via a state machine (config/state-machine.json), with permissions configured in config/permissions.json and method registration in config/methods.json.

Five Forbidden Rules:

1. No storing speculation as facts (distinguish observation vs inference).
2. No overwriting raw materials (immutable original data).
3. No putting everything into long-term memory (scarce resource, need screening).
4. No single agent handling取证, judgment, and audit (separation of powers).
5. No clearing unexplained anomalies (preserve异常 signals).

Six Specialized Agents:

• fact-cleaner: Data cleaning and structuring.
• constraint-checker: Verify compliance with real-world constraints.
• equilibrium-explainer: Explain decision trade-offs.
• action-translator: Convert decisions into executable actions.
• redteam-auditor: Independent audit to challenge decision rationality.

Three Access Entries:

• GitHub Issue: For non-Git users to submit viewpoints.
• submissions/viewpoints/*.md: For local users who can write Markdown.
• submissions/agent-queue/*.json: For agents like OpenClaw to read directly.

Volcano Ark Integration:

• Base URL: https://ark.cn-beijing.volces.com/api/coding/v3.
• API key: Only read via environment variable ARK_API_KEY (no hardcoding).
• Default model: doubao-seed-2.0-code.

The Lab is ideal for:

1. Decision systems requiring strict audit tracking (e.g., financial risk control, medical diagnosis assistance).
2. Multi-source information fusion for knowledge management (building trusted knowledge bases).
3. Long-running AI projects (maintaining decision history for review and iteration).
4. Compliance-sensitive applications (meeting regulatory requirements for AI explainability).

Its value lies in enabling traceable, verifiable AI decisions, fostering trust in AI systems.

Limitations:

• Early stage (v0.2.0), with most docs in Chinese.
• Tightly coupled with Mercury Agent ecosystem (may be challenging for non-Mercury users).
• Strict process constraints increase complexity (not suitable for quick prototypes).

Conclusion: As AI auditability becomes a necessity, Mercury Method Lab provides a valuable framework for building trusted AI systems. Its structured flows, constraints, and role division offer actionable insights for scenarios needing 'trust but verify' mechanisms.