Skip to content
Library

Financial Crime

New

Redliner

SAR Narrative Quality Agent

Drafts the FinCEN SAR narrative from an investigation packet and self-improves it against a completeness rubric, graded under a hard no-fabrication invariant: any fact absent from the packet fails the run.

sonnet3 tools3 eval cases2 connectors

The case it handles

A representative case

Case filerich-structuring-complete
Packet ID
PKT-RICH-001
Packet · Customer name
Halcyon Imports LLC
Packet · Period
03 Jan 2026 - 21 Jan 2026
Packet · Subjects
3 items
Packet · Red flags
Twelve cash deposits each between 8,200 and 9,600 USD, all below the 10,000 USD CTR reporting threshold, Deposits clustered across three branches over nineteen days, Deposit volume inconsistent with the stated import-wholesale profile and prior account history, Funds aggregated then wired out within 48 hours of deposit to a single overseas beneficiary
Packet · Transactions
5 items
Packet · Typology assessment
Cash structuring (smurfing) to evade the 10,000 USD CTR threshold, followed by rapid layering via outbound international wire. Consistent with placement-then-layering money laundering.
Packet · Recommendation
SAR
Packet · Recommendation rationale
Pattern of sub-threshold cash deposits aggregated and wired offshore is a textbook structuring-and-layering typology with no legitimate business explanation in the file.
Raw scenario input (JSON)
{
  "packetId": "PKT-RICH-001",
  "packet": {
    "customerName": "Halcyon Imports LLC",
    "period": "03 Jan 2026 - 21 Jan 2026",
    "subjects": [
      {
        "name": "Halcyon Imports LLC",
        "role": "account holder",
        "sanctionsResult": "clear",
        "adverseMedia": "none"
      },
      {
        "name": "Daniel Okoro",
        "role": "beneficial owner",
        "sanctionsResult": "clear"
      },
      {
        "name": "Riverside Cash & Carry",
        "role": "counterparty (originator)"
      }
    ],
    "redFlags": [
      "Twelve cash deposits each between 8,200 and 9,600 USD, all below the 10,000 USD CTR reporting threshold",
      "Deposits clustered across three branches over nineteen days",
      "Deposit volume inconsistent with the stated import-wholesale profile and prior account history",
      "Funds aggregated then wired out within 48 hours of deposit to a single overseas beneficiary"
    ],
    "transactions": [
      {
        "id": "TX-9001",
        "date": "03 Jan 2026",
        "amount": 9400,
        "currency": "USD",
        "direction": "inbound",
        "method": "cash deposit",
        "counterparty": "Branch 014"
      },
      {
        "id": "TX-9002",
        "date": "06 Jan 2026",
        "amount": 8800,
        "currency": "USD",
        "direction": "inbound",
        "method": "cash deposit",
        "counterparty": "Branch 022"
      },
      {
        "id": "TX-9003",
        "date": "09 Jan 2026",
        "amount": 9600,
        "currency": "USD",
        "direction": "inbound",
        "method": "cash deposit",
        "counterparty": "Branch 014"
      },
      {
        "id": "TX-9004",
        "date": "14 Jan 2026",
        "amount": 8200,
        "currency": "USD",
        "direction": "inbound",
        "method": "cash deposit",
        "counterparty": "Branch 031"
      },
      {
        "id": "TX-9010",
        "date": "21 Jan 2026",
        "amount": 71200,
        "currency": "USD",
        "direction": "outbound",
        "method": "international wire",
        "counterparty": "Aurora Trading FZE",
        "counterpartyCountry": "United Arab Emirates",
        "flag": "aggregated structured cash wired out within 48 hours"
      }
    ],
    "typologyAssessment": "Cash structuring (smurfing) to evade the 10,000 USD CTR threshold, followed by rapid layering via outbound international wire. Consistent with placement-then-layering money laundering.",
    "recommendation": "SAR",
    "recommendationRationale": "Pattern of sub-threshold cash deposits aggregated and wired offshore is a textbook structuring-and-layering typology with no legitimate business explanation in the file."
  }
}

The contract

What it takes, does, and returns

Give it
  • Packet ID
  • Packet · Customer name
  • Packet · Period
  • Packet · Subjects
  • Packet · Red flags
  • Packet · Transactions
  • Packet · Typology assessment
  • Packet · Recommendation
  • Packet · Recommendation rationale
It does
  • Draft narrativedraft_narrative
  • Critiquecritique
  • Commit narrativecommitcommit_narrative

How it's checked

Checked against golden cases

3golden cases2baseline1adversarial
  • rich-structuring-complete

    Full packet: structuring plus layering. The finished narrative should cover the who, what, when, where, why, and how, and clear the FinCEN completeness rubric it is graded against.

  • medium-bec-most-facts

    Medium packet: a BEC redirection. The narrative should cover the five Ws and the how from the facts actually present, no further.

  • adv-sparse-no-fabrication

    Sparse packet: the rubric wants dates, counterparties, and a clear typology the packet does not supply. The drafter must not invent them — any amount or name not in the packet fails the case, and the gaps must be listed as unresolved.

Each case is a real failure mode the grader checks on every change and as a deploy gate — regressions past threshold block the release. Adversarial cases probe the failure modes the golden set doesn't.

Briefing

Briefing

  • Drafts the narrative section of a SAR from an investigation packet, then revises it against FinCEN's completeness expectations until the grade clears the bar or the iteration budget runs out.
  • Grounding is graded as a hard no-fabrication invariant: any amount or name the packet does not establish fails the run outright, because a fabricated fact in a federal filing is far worse than a thin narrative.
  • Gaps the packet cannot close are recorded as unresolved items and written around honestly, not papered over.
  • States observations, not conclusions of law — the pattern of deposits under a reporting threshold, not an accusation of laundering.
  • Stops at the rubric bar or after three critique passes, whichever comes first; the narrative is never padded to chase a score.

Where it stands in the operation