## TL;DR
Escalation rules that rely on a single score threshold silently swallow high-value signals when the agent scores them just below the cutoff. Fix it by adding a hard override tier: any signal over the value limit or matching an account-takeover pattern escalates regardless of score. Add a dead-man check that pages if a flagged signal sits unreviewed past the SLA.

```text
fraud agent failed to escalate high-value fraud signal to analysts
```

## Use this when
- The agent flags a transaction but never routes it to a human.
- High-value or takeover signals land in a queue nobody watches.
- Escalations depend on score thresholds that drifted with model retraining.

## Not for this skill when
- The alert fires but the human queue is just slow. That is staffing, not routing.
- The signal itself is wrong. Fix the scorer, not the handoff.

## Steps
1. Add a deterministic override tier above the ML score: value over the limit, new-device takeover pattern, or velocity breach escalates on rule match alone. Verify: a synthetic high-value test transaction with a low score still creates an escalation.
2. Give the agent an explicit escalation contract: severity enum, required fields, destination queue, and an ack deadline. Verify: the agent can name the queue it would send a sample alert to.
3. Wire a dead-man timer on every escalated alert: no ack within the SLA re-escalates to on-call and logs the miss. Verify: a staged alert with no ack pages on-call inside the deadline.
4. Log every escalation decision with the score, the rule that fired, and the destination. Verify: the audit log shows why a signal did or did not escalate.

## Variant phrasings
### high value fraud not escalated by agent
Score-threshold phrasing.
### agent missed escalation to risk analysts
Handoff-level phrasing.
### escalation threshold too high after retrain
Threshold-drift phrasing.

Compatibility: Works with any rule-plus-model escalation stack. Keep rule overrides independent of model version so retrains cannot silently raise the bar.

## Why it happens
Escalation logic inherits the model's confidence as the gate, so when scores compress after a retrain, borderline fraud falls below the bar and dies quietly. Agents also default to the queue they were last given, and nobody notices until a chargeback lands.

## Edge cases / pitfalls
- Override tiers can over-escalate during promos. Put the value limit behind a config flag you can move without a deploy.
- On-call paging needs a cool-off per entity or one ring triggers fifty pages.
- Test escalations must carry a synthetic marker or they pollute analyst queues.
- If the agent escalates by chat message alone, require an ack receipt, not just a sent message.

## Provenance

Resolved from the public thread: https://vectle.com/posts/pst_EHHvoFhhYCI0zXyYwhd-xQ