Build a Multi-Reviewer Watchlist Disposition Discrepancy Queue — implementation

## Build
Ingest screening adjudications from multiple compliance analysts to identify systematic review discrepancies. Flag out-of-policy determinations and highlight missing source evidence for supervisory audit.

## User
Sanctions Audit Supervisor

## Workflow
1. Collect multi-analyst disposition records linked to FollowTheMoney entity screening candidates.
2. Compare adjudications against pre-defined screening policy rules to detect unauthorized false-positive clearances.
3. Verify if screening reviewers consulted the primary source lineage URL before recording dispositions.
4. Produce an aggregated discrepancy review queue highlighting rogue analyst patterns and unsubstantiated decisions.

## Interface
Supervisor review cockpit displaying analyst disagreement heatmaps, policy deviation lists, and case escalation controls.

## Data
Synthesize a dataset of multi-analyst review logs referencing FollowTheMoney entity schemas.
Starter examples below are synthetic. They are not downloaded records.

## Source
[OpenSanctions data toolkit](https://raw.githubusercontent.com/opensanctions/yente/main/README.md)
Adopt FollowTheMoney entity model properties and yente matching endpoint output structures.

## Tools
crewai, pydantic, fastapi.
crewAI coordinates independent auditor and policy-checker agent roles to cross-examine disposition logs.
Hosted or local model capable of multi-agent debate and JSON structured extraction.

## Deliverables
Submit runnable source, pinned dependencies, setup commands, a working interface, example outputs, and a recorded end-to-end demo. Include tests and an action log.

## Acceptance
- Identifies 100% of cases where an analyst cleared a high-score candidate without providing mandatory comments.
- Flags disposition variance where two analysts reached opposite conclusions on identical entity attributes.
- Generates a structured supervisor remediation task for each flagged variance.
- Produces valid discrepancy queue objects adhering to the audit schema.

## Starter Examples
```json
[
  {
    "input": {
      "review_event_id": "REV-3001",
      "entity": {
        "id": "ftm-66",
        "schema": "Vessel",
        "properties": {
          "name": [
            "Oceanic Star"
          ],
          "flag": [
            "pa"
          ]
        }
      },
      "adjudications": [
        {
          "analyst_id": "usr_9",
          "decision": "clear",
          "evidence_verified": false,
          "notes": "No match on IMO."
        },
        {
          "analyst_id": "usr_4",
          "decision": "escalate",
          "evidence_verified": true,
          "notes": "IMO match verified in maritime appendix."
        }
      ]
    },
    "expected_output": {
      "review_event_id": "REV-3001",
      "policy_violation": true,
      "discrepancy_score": 0.85,
      "supervisor_action": "require_re_review",
      "findings": "Reviewer usr_9 cleared vessel without verifying primary source appendix; contradicts usr_4 escalation."
    }
  },
  {
    "input": {
      "review_event_id": "REV-3002",
      "entity": {
        "id": "ftm-88",
        "schema": "Person",
        "properties": {
          "name": [
            "Alice Walker"
          ],
          "birthDate": [
            "1975-04-12"
          ]
        }
      },
      "adjudications": [
        {
          "analyst_id": "usr_2",
          "decision": "clear",
          "evidence_verified": true,
          "notes": "DOB mismatch against subject record."
        },
        {
          "analyst_id": "usr_7",
          "decision": "clear",
          "evidence_verified": true,
          "notes": "Different middle name and confirmed DOB variance."
        }
      ]
    },
    "expected_output": {
      "review_event_id": "REV-3002",
      "policy_violation": false,
      "discrepancy_score": 0,
      "supervisor_action": "approve_closure",
      "findings": "Consensus clear disposition supported by verified DOB and name discrepancies."
    }
  }
]
```

## Safety
Demonstration uses mock sanctions cases and analyst usernames; no actual AML or sanctions filings occur.
Use mock external writes. Require human review before any real deployment.