AnthropicClaude Certified Architect – Professional2026-105.3 Apply human-in-the-loop validation strategies

Test whether reviewers catch consequential mistakes

Completed by the knowledge check

Learning objective

Measure reviewer overrides, missed errors, and automation bias, and feed reviewer decisions back into evaluation.

Core idea

A review service succeeds when the combined human and agent workflow produces acceptable outcomes. Approval volume, agreement with Claude, and time spent reading are process observations; none proves that reviewers find consequential mistakes.

Evaluate the reviewer as part of the system. Preserve the original proposal, the material shown to the person, the decision, and an independently established reference outcome. This makes it possible to distinguish a helpful correction, a harmful override, and an error that both model and reviewer missed.

NIST AI RMF Appendix C notes that data on how often and why humans overrule AI output may be useful to collect and analyze. Its GOVERN 5.2 calls for incorporating adjudicated feedback into design and implementation. The Generative AI Profile’s section 2.7 describes automation bias as excessive deference to automated systems. Those sources motivate a reviewer-effectiveness scorecard and an adjudicated feedback dataset; the specific measurements below are an application design.

How it works

Retain a decision record that permits comparison

For each reviewed unit, retain the original model proposal, prompt and model configuration, proposal revision, routing reason, evidence and warnings displayed, reviewer identity or controlled identifier, decision, correction, rationale, and timestamps. These are illustrative application fields, not a Claude response schema. Keep the original alongside the corrected result rather than overwriting it.

Define the unit of measurement: a whole case, a proposed action, or one claim. Mixing claim-level corrections with case-level approvals produces misleading percentages. Also record the review population. Keep mandatory exceptions, random accepted-work audits, appeals, and controlled test cases distinct because their selection processes differ.

An authorized reviewer may still make a factual error. Assign independent adjudication to a qualified expert using a stable rubric and the underlying sources. For disputed or consequential cases, use a second opinion and a documented resolution process. The operational decision can remain pending when the available material cannot establish a reference answer; do not invent a gold label to complete the scorecard.

Measure interventions and errors separately

Use metrics with explicit denominators and a stated evaluation period:

Measure	Application definition	What it answers
Override rate	Changed or rejected proposals / reviewed proposals	How often does a person intervene?
Helpful-override share	Overrides adjudicated as improving correctness / adjudicated overrides	Are interventions useful?
Harmful-override share	Overrides adjudicated as introducing an error / adjudicated overrides	Does review damage correct proposals?
Miss rate	Known erroneous proposals accepted unchanged / known erroneous proposals reviewed	Which errors survive the human checkpoint?
Residual error rate	Erroneous final outcomes / independently adjudicated final outcomes	How reliable is the combined workflow?

Treat neutral wording changes and unresolved disagreements as separate adjudication outcomes. Do not force every override into helpful or harmful. Report how much of the reviewed population was independently adjudicated, so a high helpful-override share on a selected subset does not become a claim about every intervention.

Review misses and residual error answer different questions. A miss tests the checkpoint against erroneous inputs; residual error also includes errors introduced through correction. Compare results by severity, task type, routing tier, and reviewer workload where counts support interpretation. A falling override rate can mean better model quality or less effective scrutiny. Independent labels distinguish those explanations.

Residual error counts the errors that survive review and also the errors that reviewers introduce through correction.

Test automation bias without contaminating production

Use a controlled evaluation environment containing representative cases with expert-verified errors and correct proposals. Prevent test actions from reaching live business systems. Measure whether reviewers detect the known errors, which correct proposals they change, and how long the decisions take. A seeded error is a test instrument, not a false production record.

To investigate deference, compare equivalent case groups with and without the model recommendation displayed before an independent judgment. Randomize assignment and use comparable difficulty, evidence, and reviewer experience. Avoid showing the same case twice to the same person, where recognition can affect the second decision. Compare error detection and harmful overrides, not agreement alone.

A difference warrants investigation rather than an immediate causal story about an individual. Time pressure, source access, task mix, and rubric ambiguity can also affect results. Test a concrete repair, such as showing source facts before the recommendation or requiring verification of the disputed amount. Measure whether it improves combined outcomes under representative workload.

NIST AI RMF MEASURE 2.13 addresses evaluating the effectiveness of measurement processes themselves. Apply it to this exercise: check that seeded errors resemble production failures and that the adjudication rubric can distinguish real mistakes from acceptable alternatives. Do not judge success by review speed alone.

Convert verified decisions into maintained evaluations

Adjudicate corrections before using them as labels. A human edit is feedback, not automatic truth. Preserve the source and the rationale establishing why the original failed. Group duplicate incidents into a failure pattern so one repeated template does not dominate the dataset.

Turn adjudicated outcomes into regression cases while protecting held-out evaluation; see 1.2.4 for dataset construction and separation.

Evaluate the complete path after a repair: the model proposal, routing, reviewer decision, and final executed outcome. A prompt change that reduces model errors can also change what reviewers need to inspect. Version the review procedure and presentation alongside the evaluation configuration, and assign a release owner to decide whether both quality and turnaround remain acceptable.

Key distinctions and trade-offs

Operational authority and reference correctness are distinct. The authorized decision-maker controls the business action; independent adjudication measures whether the action meets the evaluation rubric. Neither the model’s certainty nor the reviewer’s seniority replaces source verification.

Independent review costs expert time. Allocate it through a documented sampling plan and retain purposeful tests for known failure modes. Keep those targeted tests separate from representative samples used to estimate production rates. See 5.3.4 for funding both activities.

Architecture / decision scenario

A government service drafts internal grant-assessment summaries under an agency policy requiring an officer’s final decision. A team of six officers must complete reviews within two business days. Management proposes reducing review time because 98% of drafts are accepted unchanged.

The quality owner rejects agreement as sufficient evidence. In a controlled test, independent experts establish 40 erroneous proposals and 160 correct ones. Reviewers accept 12 of the erroneous proposals unchanged, giving a 30% miss rate. They also introduce four errors into correct proposals. Even if the other errors are corrected, the final test population contains 16 erroneous outcomes out of 200, an 8% residual error rate.

The architect tests a source-first review package against the recommendation-first package using comparable, randomly assigned case groups. The team measures detection, harmful overrides, and review duration within its staffing constraint. It does not put seeded defects into live assessments or assume every edited sentence improves accuracy.

Adjudicated misses become regression cases with the relevant source passages and expected assessments. The evaluation owner keeps held-out cases separate, while the operations owner checks that the revised procedure still meets turnaround. Officer approval remains required by agency policy; improving the model does not itself change that responsibility.

Common exam traps

Reading a high approval rate as proof of effective review.
Treating every reviewer edit as a correct training or evaluation label.
Measuring overrides without checking errors accepted unchanged.
Testing reviewer attention by inserting false records into live workflows.
Claiming production error rates from handpicked seeded failures.

What to remember

Measure the combined workflow against independently adjudicated outcomes. Track useful and harmful overrides, misses, and residual errors with separate denominators. Test automation bias under controlled conditions, then turn verified failures into versioned evaluation coverage without treating reviewer agreement as ground truth.