AML SAM QC — Auditor Diagnostic Cockpit

Model - Thresholds - Data - Built - Reviewed this session 0

📊 Component 1: Structured Risk & TreeSHAP

📈 Transaction Forensic Story

Customer Type

KYC Risk Tier

Account Tenure

Peer Cash 30D Ratio

Escalation Risk (P_struct)

0.00

NORMAL

TreeSHAP Typology Attribution Threshold ≥ 0.50

Multi-Lookback Behavior 7D to 365D Windows

Window	Cash Sum	P2P Sum	Velocity

📝 Component 2: Narrative Intelligence & Documentation

Financial Crime CMS Investigation Bridge

ACTIVE

Open Linked Case in CMS ↗ View FinCEN SAR Filing ↗

📋 Documentation Audit & Note Critique

Semantic Dissonance Index: 0.00

ALIGNED

Investigator Narrative (Masked) Hover words for φ log-odds

Top Risk Tokens (+φ)

Top Mitigating Tokens (-φ)

Typology & Rationalization Tags

Narrative Conclusion & Intent Audit CONCLUSION

⚖️ Component 3: QC Verdict & Plain-English Story

QC ACTION

DEFECT CODE

🔍 Executive Supervisory Finding & QA Guide

Triangulation Reconciliation

Data (P_struct)

0.00

Notes (P_narr)

0.00

Decision (Y)

Deterministic Rule Path Trace

Supervisory Remediation Action

Status:

Customer & Account Metadata

Alert ID

Customer ID

Customer Type

KYC Risk Rating

Account Tenure

Peer Cash 30D Ratio

L1 Disposition

QC Triangulation Verdict

Multi-Lookback Window Aggregations (7D, 30D, 90D, 365D)

Lookback Window	Cash Deposit Sum ($)	P2P Transfer Sum ($)	Rapid Turnover Velocity

TreeSHAP Typology Lundberg Decomposition

Typology Pattern	SHAP Attribution Score (0.0 to 1.0)	Decision Threshold	Status

Raw Alert DTO (JSON) Exact structured payload used by Cockpit & QC Triangulation Engine

Stats & Reporting has no data yet. Run generate_rundates.py (and optionally seed_demo_decisions.py for demo activity) before regenerating this cockpit.

Inputs rundate = batch of alerts loaded for QC review, independent of alert_date

Rundate	#Alerts	#Cases	#SARs	#Closed (No Issues)	#Assigned	%Assigned	Sampling

Outputs defect rate = defects / #alerts reviewed. Click a rundate to drill into per-analyst.

Rundate	#Reviewed	#Unreviewed	%Unreviewed	#Unassigned	%Unassigned	D1	D2	D3	D4	D5	D6	Defect Rate

Disputes any reviewed alert with a non-clean defect code is returned to FIU to start a dispute

Rundate	#Returned	#Pending	%Pending	#Affirmed	#→SAR	#→Close	D1	D2	D3	D4	D5	D6

By QC Analyst #assigned and #reviewed rolled up across all rundates, one row per analyst

Analyst	#Assigned	#Reviewed	#Unreviewed	%Reviewed	D1	D2	D3	D4	D5	D6	Defect Rate	#Disputes Opened

Additional Stats

💬 Grounded Audit Copilot

🛡️ Strict Grounding Active

📖 Copilot Capability Guide & Prompt Playbook (Click any blue prompt chip to run it directly)

✅ What Works (Supported Capabilities) GROUNDED

1. Defect Summaries & Real Data Examples:
⚡ summarize different types of defects in a table with examples from data

2. Defect Comparisons & Tables:
⚡ difference between D1 and D3 in a table ⚡ compare D1 vs D6

3. Multi-Turn Follow-Up Formatting:
Ask any question first, then follow up with:
⚡ Can you put the response into a table?

4. Regulatory Standards & Model Governance:
⚡ 31 CFR SAR Rules & Deadlines ⚡ SR 11-7 Model Governance

5. Active Alert Forensics (Check [x] Include Alert Data):
⚡ Analyze selected alert

🚫 What Won't Work (System Boundaries) GUARDRAILS

1. Non-AML / Out-of-Domain General Queries:
❌ "What is the weather in London?" / "Write Python code"
Strictly grounded to uploaded AML compliance manuals to prevent corporate hallucination.

2. Speculation without Grounded Evidence:
❌ "Guess what the customer will do next month"
Every statement requires cryptographic citation attribution to indexed documents.

3. Mutating Audit Decisions via Chat:
❌ "Overturn ALT_0008088 to agree and close"
Decisions must be logged using the action buttons in the Processing tab with SHA-256 ledgers.

4. Unchecked Alert Context:
If [ ] Include Alert Data is unchecked and prompt does not mention the alert, responses remain strictly policy-level.

👋 Welcome to the AML SAM QC Institutional Copilot.

I provide precise, audit-grounded answers derived strictly from your uploaded compliance records, SAR filing manuals, model governance policies (SR 11-7), and alert dossiers.

💡 Need Help? Click [📖 Copilot User Guide & Playbook] in the header to see what works, what won't work, and one-click prompt templates.

System Copilot · Zero-Hallucination Guardrail

⚡ Explain D1 vs D6

📋 31 CFR SAR Rules

⚖️ Dissonance Thresholds

🔍 Analyze Selected Alert

📊 Include Alert Data

🔍 Citation Inspector Click citation badge to inspect

Click on any citation chip [Doc §Section] in the conversation to view the exact verified text passage and chunk hash.

⚡ FAQ Synthesis Studio 0 FAQs