S010 Stagegate Runner

Topic TypeTCAML Framework ReferencePrimary UseAnalyst-supervised local AML agent architecture and RAG useScopeJurisdiction-neutral public referenceLast Reviewed2026

The S010 stagegate runner is a deterministic local execution path for verification and evidence-controlled AML operations.

It turns structured operation input into a controlled workflow with gate status, event log, reports and case folder output. The runner does not perform live source screening by itself. It controls process discipline around content that has already been prepared or loaded.

Important: demo auto-confirm mode is for deterministic tests and examples only. LLM-generated drafts must be run through analyst-interactive gate confirmation.

What the runner controls#

The runner controls the progression of a review through S010 gates:

case setup
→ declared-data inventory
→ subject classification
→ verification plan
→ source checks
→ evidence capture
→ discrepancy handling
→ risk trigger mapping
→ case file assembly
→ routing / handover

Each gate has an exact keyword. Generic phrases such as ok, yes, continue or go ahead do not advance the workflow.

Why deterministic gates matter#

An LLM may produce useful content, but it should not decide whether evidence is sufficient or whether an escalation can be released. A deterministic gate machine creates a predictable control boundary:

  • each gate has a known state;
  • invalid keywords are rejected and logged;
  • gates cannot be confirmed out of order;
  • already confirmed gates cannot be double-confirmed silently;
  • reopened gates must be confirmed through the same controlled path;
  • escalation override can block release and require escalation routing.

Interactive and demo modes#

The runner supports two different use cases:

Mode Use Boundary
Interactive Real review of a draft or case input. AML analyst enters exact gate keywords.
Demo / auto-confirm Deterministic examples and repository tests. Not for LLM-generated drafts or real reviews.

The important rule is that an LLM draft must never be used to self-confirm the workflow. If the input came from an LLM, a AML analyst must run the validated file interactively and type the gate keywords.

Case folder output#

A controlled run should produce a local case folder structure similar to:

00_case_metadata/
01_declared_data/
02_verification_plan/
03_source_checks/
04_evidence/
05_discrepancies_and_risk/
06_reports/
07_audit_trail/
08_handover/

This structure supports review, escalation, audit reconstruction and later migration into a client’s internal case management system.

Audit trail#

The runner should preserve:

  • gate status;
  • gate event log;
  • rejected keyword attempts;
  • reopening reasons;
  • escalation override events;
  • source check references;
  • evidence references;
  • handover status.

The audit trail is not a final decision. It is the record of how the case moved through the controlled workflow.

Relationship to S010 output#

The runner uses S010 structured output to keep the review machine-readable. S010 output preserves the decision boundary by requiring Pending analyst review as the agent-produced decision state. The reviewer decision occurs outside the agent output.

References#

Tom Custos AML Knowledge BaseLearning & OperationsDownload source .md