# Alignment Boundary

The line between what model training covers and what your application layer must enforce itself.

## Why the CCAR-P exam tests this

Candidates must not assume model alignment removes the need for app-layer controls; the boundary defines who owns which risk.

## Definition

The **alignment boundary** separates two kinds of safety: what the **model's training** provides (general helpful/harmless behavior, refusals for obviously harmful requests) and what **your application layer** must enforce (your specific policies, authorization rules, data-handling obligations, and business constraints).

The exam's core point: **training is not your control layer.** A well-aligned model still does not know your tenant boundaries, your spend limits, your regulatory obligations, or which actions are irreversible in *your* system. Those are **must-hold rules** that belong in **deterministic code** — authorization checks, redaction, escalation policy, fail-closed guardrails. Relying on the model to "just behave" for app-specific limits is placing a control on the wrong side of the boundary.

Use the boundary as a design question: for each risk, ask "does model training reliably cover this, or must my app enforce it?" Anything on the app side goes into your **control register** with an owner and evidence. Model alignment reduces baseline risk; it does not discharge your responsibility for the controls only you can define.

## Exam trap

"The model is aligned, so we don't need our own authorization checks" is the trap — training covers general behavior, not your app's specific must-hold rules.

## Commonly confused with

- **prompt-injection-defense** — Both stress app-layer enforcement; the boundary is the responsibility split, injection defense is one risk on the app side.
- **compliance-control-register** — The boundary decides which risks you must own; the register records how you own them.

## Related

- prompt-injection-defense
- compliance-control-register
- guardrail-placement
- escalate-on-policy-not-confidence

---

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