Adaptive sampling

Human review that adapts as the agent proves itself.

A new agent version starts with every decision reviewed. Review then runs at the validation rate you set until the agent reaches its confidence target, and in steady state the rate follows the width of the confidence interval. A new release, a model change, a drift alert, or an incident raises it again.

Agent version 3.2New versionEarning confidenceSteadyHuman review coverageConfidence
Illustrative example. Actual rates follow the burn-in count, validation rate, confidence target, and surge rate configured for the agent.

Every agent version earns its own review rate

Review begins at New version, runs at the validation rate while sampled verdicts build confidence, then settles at a derived rate.

  1. New version

    Every decision reviewed

    Each execution of a new version is sampled until the burn-in count you configured is reached.

  2. Earning confidence

    Validation rate

    Review runs at the configured validation rate. Sampled verdicts build the accuracy estimate until its confidence interval reaches the target.

  3. Steady

    Derived rate

    KLA derives the rate from the width of the confidence interval, capped at the validation rate. Spot checks continue so confidence stays current.

02Surge controls

Change raises the review rate

While an enabled trigger is active, KLA raises human review to at least the surge rate you configured. When the trigger clears, the phase-derived rate resumes.

New release. A new agent version or manifest is deployed. Model change. The underlying model or provider is switched. Drift alert. Assurance detects a change in outputs or decision mix. Incident. A recent incident is open on the agent or its process.

A rate with a clear basis

The rate combines observed accuracy, configured bounds, and risk factors attached to the execution.

Sampled reviews. Sampled executions land in Decision Desk as review items. Each verdict records the reviewer, the time, the outcome, and the rationale, and feeds the accuracy estimate.

Confidence interval. In steady state the review rate follows the width of the confidence interval on observed accuracy. A wide interval raises it. A narrow interval lowers it.

Configured bounds. The burn-in count, validation rate, confidence target, and surge rate are set per agent. The derived steady rate never exceeds the validation rate.

Risk weighting. Configured risk factors multiply the base rate, up to reviewing every decision, so a payment release is sampled more often than a document classification at the same confidence.

Oversight you can prove

Each execution records its sampling phase, final rate, reason codes, and risk factors. Each completed review records the reviewer, time, outcome, and rationale. Both export with the execution evidence.

Illustrative sampling record

execution 8F31B2 · agent version 3.2 · 09:47:12 UTC

phase Earning confidence · final rate 40%

reason codes validation, deterministic_sampled

risk factors none

Illustrative review verdictcorrect

review RV-2048 · 09:53:40 UTC

A. Martin · Risk Operations

Rationale: Decision matched the approved policy and source data.

See it on your own agents.

Bring one agent and one process. We will show you the review rate it would start at, and what it takes to bring it down.

Adaptive Sampling: Human Review That Scales With Agent Confidence | KLA