Governance, Safety & Risk Management·Task 5.3·Bloom: analyse·Difficulty 3/5·10 min read·Updated 2026-07-14

Review Placement Trade-offs for the CCAR-P Exam

Apply human-in-the-loop validation strategies

SUBy Solomon UdohReviewed by Solomon UdohAI-assisted · human-reviewed
In short
Review placement is the choice of where a human sits relative to an action: pre-action approval blocks the action until a person signs off (safest, but adds latency and does not scale); post-action audit lets the action run immediately and reviews afterward (high throughput, but only fits reversible, lower-cost decisions); and sampled review checks a fraction of decisions to monitor quality without gating each one (monitors the system rather than guarding each outcome). The right placement follows from a decision's stakes, not from a single default applied everywhere.

Once a decision is routed, where does the human sit

The routing rule decides which decisions a person should weigh in on; this knowledge point decides where that person sits relative to the action. The CCAR-P exam treats it as an analyse-level skill because each of the three placements makes a different safety-versus-throughput trade, and choosing the wrong one either grinds a high-volume workflow to a halt or lets an irreversible mistake take effect before anyone looks. The governing idea is that involving a human earlier is safer and slower.

There are three placements, and they are not interchangeable defaults. Pre-action approval, post-action audit, and sampled review each buy you something and cost you something, and the right choice for a given decision is dictated by that decision's stakes - the same reversibility and cost that drove the routing rule. Applying one placement everywhere, regardless of stakes, is the error the analysis is meant to prevent.

Review placement trade-offs
The choice of where a human reviewer sits relative to an action. Pre-action approval blocks the action until sign-off (safest, slowest, does not scale). Post-action audit runs the action then reviews it (high throughput, only for reversible, lower-cost decisions). Sampled review checks a fraction to monitor quality (monitors the system, does not guard each outcome). Placement follows from the decision's stakes.

Pre-action approval: safest, slowest

Pre-action approval places the human before the action: the action cannot take effect until a person approves it. Its benefit is decisive - nothing irreversible happens unreviewed, so it is the safest placement and the right one for high-stakes, irreversible actions. Its cost is equally clear: it adds latency to every routed decision, and a person must be available to approve, so it does not scale to high volume. You pay in speed and staffing for the guarantee that no consequential action runs without a human first.

This is the placement the routing rule points to for its hardest case - a low-confidence, irreversible, high-cost decision. The whole reason to accept the latency is that the action cannot be undone, so catching the mistake afterward would be too late.

Post-action audit: high throughput, reversible only

Post-action audit places the human after the action: the action runs immediately, and a person reviews it afterward. Its benefit is throughput - nothing waits on a reviewer, so the workflow runs at full speed. Its cost is the flip side: by the time a wrong action is caught, it has already taken effect. That means post-action audit only fits reversible, lower-cost decisions, where a mistake found after the fact can be corrected without serious harm.

The exam's favourite failure is applying this placement to an irreversible action because it "worked fine" for a different, reversible workflow. Throughput is seductive, but for an action that cannot be undone, reviewing afterward reviews damage that is already done. Placement has to follow the stakes of this decision, not the convenience that suited another one.

Sampled review: monitors the system, not the outcome

Sampled review places the human over a fraction of decisions: it checks some proportion to monitor quality without gating each one. Its benefit is that it watches the system's overall health cheaply, without slowing the process. But its nature is crucial - it monitors aggregate quality rather than guarding individual outcomes. A bad decision can slip through simply by not being sampled, because sampled review never promised to look at any particular case.

This makes sampled review the wrong tool for protecting a specific flagged case. If a decision has been flagged as high-stakes, sampling gives it the same odds of being seen as any other item, which is not protection at all. Sampled review is for keeping an eye on the population; pre-action approval is for guarding the individual.

Three review placements along one action
Loading diagram...
Safety decreases and throughput increases from pre-action approval to post-action audit to sampled review; the decision's stakes pick the column.

What the CCAR-P exam trips candidates on

The first trap is applying post-action audit to an irreversible, high-cost action because it worked for a different, reversible workflow. The scenario transplants a throughput-friendly placement onto an action that cannot be undone; the credited reading is that the stakes changed, so the placement must too - an irreversible action needs pre-action approval, not an after-the-fact audit of damage already done.

The second trap is assuming sampled review provides the same protection as pre-action approval for a specific flagged case. A scenario flags a particular high-stakes decision and offers sampled review as the control; the credited reading is that sampling only monitors aggregate quality and may never see that case, so it cannot guard it. The exam rewards matching each placement to what it actually guarantees.

Worked example

A team runs a content-moderation workflow with post-action audit and it works well - flagged posts are removed automatically and a reviewer checks a log afterward. They now want to reuse the same post-action audit placement for a new workflow that issues irreversible account closures. Evaluate the plan.

The plan misapplies a placement that suited one stake profile to a very different one. Post-action audit worked for content moderation because a wrong removal is reversible and low-cost - you can restore a post, and the harm of a brief erroneous removal is small - so reviewing afterward is fine, and the throughput is valuable. The stakes there justified the placement.

Account closures are a different decision. Closing an account is irreversible and high-cost: by the time a post-action audit catches a wrong closure, the account is already gone and the damage done. Reusing post-action audit here reviews a mistake that can no longer be undone, which is exactly the trap of carrying a placement across a change in stakes. The fact that it "worked fine" for moderation is irrelevant, because the reversibility and cost that made it safe there are absent here.

The correct placement for irreversible account closures is pre-action approval: block the closure until a person signs off, accepting the latency because the action cannot be undone. If closure volume is high enough that pre-action approval will not scale, the answer is to tighten the routing rule so only genuinely high-stakes closures reach the gate - not to downgrade the placement to an audit the stakes cannot tolerate. Sampled review would be worse still, since a specific wrongful closure might never be sampled.

Common misreadings to avoid

Misconception

A post-action audit that worked for one workflow is a safe default for another.

What's actually true

Post-action audit only fits reversible, lower-cost decisions, because the action has already taken effect when it is reviewed. Applying it to an irreversible, high-cost action means auditing damage that can no longer be undone. Placement must follow the new decision's stakes.

Misconception

Sampled review protects a specific flagged high-stakes case.

What's actually true

Sampled review monitors aggregate quality by checking a fraction of decisions, so a specific case may never be sampled. It watches the system, not the individual outcome, and cannot substitute for pre-action approval when a particular high-stakes decision must be guarded.

How this shows up on the exam

Domain 5 items describe a routed decision or a change of workflow and ask where the human should sit. The reliable method is to read the stakes - reversibility and cost - and match them to a placement: irreversible and high-cost demands pre-action approval; reversible and lower-cost can use post-action audit; system-wide quality monitoring uses sampled review. Reject any placement carried across a change in stakes, and never treat sampling as protection for a specific case.

This choice follows the stakes-based routing rule, which decides what gets routed, and rests on the three variables that set decision stakes. Whatever placement you choose, what the reviewer's screen must show determines whether the review is accurate, and over-using pre-action approval leads to consent fatigue and over-routing.

Check your understanding

A post-action audit works well for a reversible content-moderation workflow. The team wants to reuse it for a new workflow that performs irreversible account closures. What is the correct assessment?

People also ask

What are the options for placing a human reviewer?
Pre-action approval (blocks until sign-off), post-action audit (runs then reviews), and sampled review (checks a fraction to monitor quality) - each trading safety against throughput.
When is post-action audit appropriate?
For reversible, lower-cost decisions, because the action has already taken effect when reviewed. It should not be used for irreversible, high-cost actions.
Does sampled review protect a specific flagged case?
No. It monitors aggregate quality and may never sample a given case, so it cannot guard an individual high-stakes decision the way pre-action approval does.

Watch and learn

Official Anthropic Academy lessons first, then hand-picked walkthroughs. Videos load only when you press play.

No videos curated for this concept yet

We are still curating the best official and community videos for this topic.

Official prep for this domain

Anthropic's own free prep module for this part of the syllabus, on the official prep course. Free with an Anthropic Academy sign-in.

References & primary sources

Adaptive study

Master this concept with Archie

Practice it inside an adaptive study session. Archie, your Socratic AI tutor, tracks your mastery with Bayesian Knowledge Tracing and schedules the perfect next review.

Start studying