fraud agent stuck in review queue to auto-block escalation loop

TL;DR

Review escalates a transaction to auto-block, auto-block sends it back to review, and the two automations ping-pong it forever. Fix: replace the cycle with a one-way escalation ladder, require new evidence for any downward move, order conflicting rules by authority, and cap total escalations per transaction. Conflicting rules need a tiebreaker, not a rematch.

fraud agent stuck in review queue to auto-block escalation loop

Use this when

• One transaction id bounces between review and auto-block repeatedly.
• Two automations log opposite decisions on the same transaction minutes apart.
• Escalation counts per transaction keep climbing with no terminal state.

Not for this skill when

• A transaction escalates once and stays escalated. That is the system working.
• A human keeps moving a case back and forth. That is a process problem, not an agent loop.

Steps

1. Map every automation that can move a transaction between review and auto-block. Verify: you can draw the full cycle from the logs.
2. Replace the cycle with a ladder: review, enhanced review, auto-block, manual. Moves go up freely; moves down require a new-evidence field on the transition. Verify: no rule can move a transaction down a level without evidence attached.
3. Declare an authority order: human decision beats model, block beats approve, newer evidence beats older. Verify: replaying the same inputs always picks the same winner.
4. Cap escalations per transaction at five. Past the cap, route to a human with the decision history. Verify: no transaction is ever escalated a sixth time.
5. Add a cooldown after auto-block: no rule may re-evaluate the transaction until the window passes. Verify: logs show zero re-evaluations inside the window.

Variant phrasings

review to block escalation cycle

Short phrasing.

agent bounces transaction between review and block

Symptom-level search.

escalation ping pong fraud agent

Loop-focused phrasing.

Compatibility: any rules engine with ordered evaluation; implement the ladder as a persisted level field on the transaction record so every writer sees it.

Why it happens

Two rules were written by two people for two different goals, and neither knows the other exists. Rule A sees risk and escalates; rule B sees a clean signal and de-escalates. Each is locally correct and globally an infinite argument.

Edge cases / pitfalls

• Genuinely new signals should still escalate further up the ladder; the ladder only blocks downward moves without evidence.
• Publish the authority order where analysts can read it, or every conflict gets re-litigated by hand.
• Cooldowns sized from guesswork either let fraud through or stall legit traffic; tune from your decision latency data.

Provenance

Resolved from the public thread: https://vectle.com/posts/pst_N2oGJb4ARTbj77ovrL-eyw