Understand detection benchmarks
Documentation for Tutela v3.9.0.
Some detection controls use benchmark results supplied with the software release. Those results describe how the classifier performed on a curated test set; they do not describe the sensitive information in your organization's traffic.
Prerequisites
- Access to the relevant control in Evidence > Compliance.
- The release version of your deployment when comparing benchmark results.
What these controls measure
A benchmark runs known examples through the classifier and measures its detections. Recall is the proportion of relevant examples that it detected. A passing benchmark-backed control means the measured result met that release's threshold for the tested data class.
For example, a cardholder-data detection control can use benchmark accuracy as evidence of a platform capability. Every deployment using the same benchmark-bearing release receives that result. It is not a count of cardholder data found in your workspace and does not guarantee detection of every future input.
What to check for your own traffic
- Open the control and identify whether its source is a release benchmark or observed workspace activity.
- For observed activity, open Findings and the relevant Activity session. Check the source, time, data classification and recorded policy decision.
- If there are no detections, determine whether the connection has recent inspected traffic at all. Use Verify your installation with a synthetic example and an expected policy result.
- If a known synthetic example produces an unexpected result, record the version, request ID and policy scope and report it. Do not send actual sensitive customer content to demonstrate the problem.
No detections can mean that the data class was not present, that traffic was not inspected, or that evidence is missing. A passing benchmark alone cannot distinguish those cases.
Validate
- Benchmark evidence and workspace activity are identified separately.
- An absence of detections is not interpreted as complete protection.
- Unexpected results have a reproducible synthetic test and a recorded version.
Next steps
Model changes and historical results
The Laya stable release's focused category acceptance is integration evidence, not a replacement general accuracy benchmark. Do not attribute an earlier detector's benchmark to Laya merely because the older evidence remains available. Compare the model artifact, source version, dataset and measurement date before reusing a result.