Decisions & Thresholds

Configure how Sentra maps a 0-1000 risk score into PASS, REVIEW, HOLD, and BLOCK.

Sentra converts model probability into an integer risk score from 0 to 1000. That score is mapped to a decision using configurable thresholds.

Default thresholds

Decision	Default threshold	Meaning
`BLOCK`	`score >= 850`	Stop the transaction.
`HOLD`	`score >= 700`	Temporarily freeze or delay settlement.
`REVIEW`	`score >= 500`	Send to analyst queue.
`PASS`	`< 500`	Continue the flow.

The exact mapping in the backend is:

if score >= block_threshold:
    return "BLOCK"
if score >= hold_threshold:
    return "HOLD"
if score >= review_threshold:
    return "REVIEW"
return "PASS"

System config

Thresholds are stored in system_config under the risk_thresholds key and exposed through Admin APIs.

{
  "block_threshold": 850,
  "hold_threshold": 700,
  "review_threshold": 500,
  "prompt_version": "sentra-llm-v0.1"
}

Read config

curl "$SENTRA_API_BASE/v1/admin/config" \
  -H "Authorization: Bearer $ADMIN_TOKEN"

Required role: Admin.

Update config

curl "$SENTRA_API_BASE/v1/admin/config" \
  -X PUT \
  -H "Authorization: Bearer $ADMIN_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{
    "block_threshold": 880,
    "hold_threshold": 720,
    "review_threshold": 520,
    "prompt_version": "sentra-llm-v0.2"
  }'

Platform handling guidance

Decision	Customer-facing state	Internal handling
`PASS`	Continue normally.	Log score and request ID.
`REVIEW`	Show pending or processing.	Analyst reviews case and submits feedback.
`HOLD`	Delay settlement, payout, or fulfillment.	Require verification, AI enrichment, or senior review.
`BLOCK`	Stop the transaction.	Record final decision, reason, and support guidance.

Reasons and evidence

Sentra returns up to three reasons and up to four evidence items. Current deterministic reason sources include:

Large ticket amount.
Cross-border corridor.
Adverse transaction status such as failed or refunded.
KYC/KYB signals from AI enrichment when cached.
Score threshold breach from structured model features.

Use reasons and evidence for analyst context and internal support tooling. Avoid exposing raw KYC evidence directly to customers.