# Escalate on Policy, Not Confidence

Route to a human on fixed policy rules, never on the model's self-reported confidence score.

## Why the CCAR-P exam tests this

Candidates must recognize that a model's stated confidence is not a calibrated probability and cannot be trusted as an escalation trigger.

## Definition

A model that says it is **0.92 confident** is not telling you there is a 92% chance it is correct. That number is a self-report generated by the same system that may be wrong — it is **not calibrated**, not a ground-truth probability, and is trivially high even when the model is hallucinating. Treating `confidence >= 0.9` as "safe to auto-approve" is a classic control failure.

The correct pattern is to **escalate on policy**: define fixed, deterministic rules that decide when a human must be in the loop — for example, any action touching money above a threshold, any irreversible operation, any request matching a sensitive category, or any low-agreement result from an ensemble check. These triggers live in **deterministic code**, not in the model's judgment.

Confidence signals can *inform* a policy (e.g. as one input among several) but must never *be* the policy. The governing question is "does this action class require review?" — a property of the request and its blast radius — not "how sure does the model feel?"

## Exam trap

A distractor will offer "auto-approve when the model's confidence exceeds 0.9." This is wrong because the confidence figure is uncalibrated self-report, not a real probability of correctness.

## Commonly confused with

- **fail-closed-vs-fail-open** — Fail-closed is what to do when a check errors; escalate-on-policy is when to route to a human by design.
- **alignment-boundary** — The boundary defines what your app must enforce; escalation policy is one such app-layer enforcement mechanism.

## Related

- fail-closed-vs-fail-open
- compliance-control-register
- guardrail-placement

---

Source: Claude Cert Prep — an independent, unofficial CCAR-P study resource (domain P5). Not affiliated with Anthropic.