Human-in-the-loop review puts a person at a decision point that automation should not resolve alone. In production AI systems, review can gate a consequential action before execution or supply judgments that calibrate evaluation and validate test data. section 6.7.3 section 8.3.4
A human gate intercepts a high-stakes tool call before it reaches the external system. The platform stores the pending action, notifies an appropriate approver, and returns a holding response. Execution resumes after explicit approval; a timeout can expire the request or escalate it, according to policy. section 6.7.3
Listing 6.25 Human approval gate configuration
platform.guardrails.add_policy(
name="high-stakes-approval",
type="human_gate",
rules=[
{
"tool": "healthcare.clinical.record_diagnosis",
"approval_role": "physician", #A
"timeout_hours": 24,
"on_timeout": "expire_and_notify_user"
},
{
"tool": "healthcare.billing.process_refund",
"condition": "amount > 500", #B
"approval_role": "billing_supervisor",
"timeout_hours": 48
},
{
"tool": "healthcare.scheduling.book_procedure",
"condition": "procedure_category == 'surgical'",
"approval_role": "scheduling_coordinator",
"timeout_hours": 4,
"on_timeout": "escalate_to_manager" #C
}
]
)The policy assigns approver roles and timeouts to diagnosis, refund, and procedure-booking actions; conditions limit some gates to higher-risk cases. section 6.7.3
The gate belongs in the platform or client, not merely in the tool transport. MCP recommends human approval for tool invocations but leaves enforcement to the client. Approval complements other guardrails, including permission checks and action-consistency policies. section 6.4.4 section 6.7.3
Human evaluation serves a different purpose: reviewers label responses rather than approve actions. Because review is expensive and less scalable than automated scoring, a small production sample can provide ground-truth labels to calibrate automated scorers and LLM judges. A layered example uses automated scores on all traffic, model judges on 5–20%, and human review on 1–5%. section 7.6.1
Listing 8.6 Annotation queue for human review
queue = platform.experiments.create_annotation_queue(
name="intake-quality-review",
workflow_id="patient-intake",
rubric=[ #A
{"name": "correctness", "type": "categorical",
"options": ["correct", "partially_correct",
"incorrect"]},
{"name": "helpfulness", "type": "numeric",
"min": 0.0, "max": 1.0},
{"name": "add_to_dataset", "type": "boolean", #B
"description": "Should this be added to the "
"eval dataset?"},
],
routing_rules=[
{"condition": "score.helpfulness < 0.5",
"source": "model_judge"}, #C
{"condition": "tag contains 'guardrail_triggered'"},
{"condition": "random_sample", "rate": 0.02}, #D
],
reviewers=["sarah@healthfirst.com",
"dr.patel@healthfirst.com"],
)An annotation queue routes low-scoring, guardrail-triggered, and randomly sampled traces to reviewers using a rubric. section 8.3.4
Reviewers inspect the input, output, spans, and existing scores, then annotate against the rubric. Their scores are marked HUMAN and can be compared with automated judgments to detect calibration drift. Review can also validate production traces before they become evaluation test cases, preventing unsuitable examples from polluting the dataset. section 8.3.4 section 8.3.1



























