How Phoenix AI Explains Itself

We don't believe in black boxes

Every Phoenix decision ships with:

• Top contributing signals ranked by influence
• Counterfactual — what would have flipped the verdict
• Confidence score with calibration
• Audit trail of model versions, feature values, and policy decisions

Example output

> alert: wire-req-9182
  verdict: HOLD (confidence 0.96)
  top signals:
    - sender_domain_age: 4 days  (+0.41)
    - new_beneficiary_country: 1st time  (+0.27)
    - language_bec_similarity: 0.91  (+0.19)
    - amount_vs_history: 12.4σ above mean  (+0.13)
  counterfactual:
    - if beneficiary established >90 days → SCORE 0.34 (clear)
  policy:
    - dual_control required (rule R-027)

Why it matters

Analysts work *faster* when they trust the model. Regulators *audit* explanations, not black-box logits. Customers deserve clarity when a payment is held.