AI Model Cards

Intended use: Assist reviewers by clustering likely matches across sanctions sources

Inputs: Visitor name and context signals (e.g., aliases, locale)

Outputs: Match candidates with similarity signals, not determinations

Human oversight: Human-in-the-loop adjudication required

Known limits: Ambiguous names; transliteration variance

Evaluation signals: Threshold tuning, false positive/negative tracking

Intended use: Summarize relevant risk context for reviewer attention

Inputs: Candidate matches and associated metadata

Outputs: Short risk summaries highlighting key factors

Human oversight: Human approval required; summaries are advisory only

Known limits: May omit non-critical details; relies on source quality

Evaluation signals: Reviewer feedback, time-to-decision metrics

Intended use: Extract text from ID images when enabled by org policy

Inputs: ID photo captured with consent

Outputs: Parsed text fields for faster form fill

Human oversight: Human validation; disabled by default

Known limits: Image quality, glare, occlusion

Evaluation signals: Field accuracy, rejection/override rates