Open AML Agent Framework

This page explains how the public AML knowledge base relates to the TCAML open AML agent framework.

The framework is designed for local, human-supervised AML/KYC/KYB workflows. Its purpose is not to make an AI approve or reject a customer. Its purpose is to help an AML agent prepare structured, evidence-first case work that a responsible reviewer can inspect, challenge and decide.

Framework model#

TCAML separates the AML agent system into layers:

knowledge layer      = AML concepts, red flags, source logic and workflow explanations
skill layer          = reusable AML skills and operation procedures
schema layer         = machine-readable output contracts
control layer        = S010 verification, evidence and stagegate discipline
runner layer         = deterministic local workflow execution
LLM bridge layer     = draft parsing, schema validation and safety checks
orchestration layer  = provider-agnostic LLM content drafting
analyst review layer   = gate confirmation, escalation and final decision

This separation keeps the public framework useful across jurisdictions and implementations. Local laws, internal policies, source credentials and decision authority stay outside the public baseline.

Why this matters#

AML work fails when facts, sources, evidence and decisions are blended into one narrative. The framework separates them.

Human-supervised architecture#

The framework is built around a strict distinction:

LLM / agent = drafts case content
Parser      = validates content and blocks safety violations
Runner      = controls stagegates and audit events
Reviewer    = enters exact gate keywords and decides next action
Output      = remains Pending analyst review until a reviewer decision is made outside the agent

This prevents a model from quietly approving its own work. A model-generated draft can be useful, but it must be controlled before it becomes review-ready case work.

Public framework components#

The TCAML Framework repository includes public components that can be loaded by an LLM, RAG system, coding assistant or local agent runtime:

• AML skills and operation workflows;
• source workflow guidance;
• evidence and workpaper templates;
• S010 verification and evidence control;
• stagegate rules and gate keyword registry;
• structured output schemas;
• local runner examples;
• LLM-to-S010 parser and safety boundary;
• mock and OpenAI-compatible orchestration provider interface;
• generic public implementation profile;
• embedded AML knowledge base.

What remains private#

An adopter must configure private implementation details separately:

• internal risk appetite;
• local regulatory forms and deadlines;
• domestic legal thresholds;
• source credentials and subscriptions;
• commercial database terms;
• internal escalation rules;
• client evidence retention rules;
• local model profiles;
• production monitoring and access controls.

Relation to this wiki#

This wiki is the public knowledge layer. It explains AML concepts and operational controls in reader-friendly form. The TCAML Framework repository provides the machine-readable framework artifacts, schemas, runner and orchestration interface.