AI privacy inspection: lab results.

Compare the privacy engine configurations.

Pattern and language analysis uses structured patterns and language analysis to identify sensitive fields. The local AI model adds model-based entity detection. Combined inspection runs both configurations.

Each configuration processed the same 100 texts in three passes, for 300 timed calls per configuration. The median is the middle observed time; p95 is the time below which 95% of the observed calls fell.

The timings measure the local inspection step in monitor mode. Network transport, request routing, masking, and model-provider responses are separate stages of an AI request. Initialization and warmup timings are included in the inspection report.

The test data changes the result.

The labelled set contains 50 synthetic expected values across names, email addresses, phone numbers, payment test-card values, and SSNs. A value counts as caught when its entity type matches and the detected span overlaps the expected substring.

The original email fixtures use the reserved .invalid suffix. A follow-up check changed those ten email samples to reserved example.com addresses while retaining the other text and labels. Both versions are downloadable, and both results are shown.

The comparison reports detected values within each labelled synthetic dataset. Use representative input formats when choosing the entity rules and detector configuration for your application.

The lab study explains how to turn this fixture sensitivity into a practical pilot check.

Keep the conditions with the numbers.

Test hardware

Apple M2 Max; 96 GiB RAM; ARM64; CPU inference.

Software

Python 3.11.13; fixed detector and model configuration across all measured passes.

Study record

privacy-inspection-2026-10-03

Model

Local CPU inference with verified model assets and offline analysis.

Workload

100 English synthetic timing texts: 33 short, 34 medium, 33 long. Three warmup calls per configuration; three measured passes; one call at a time.

Timing

Monotonic perf_counter_ns; seeded text and configuration shuffles (seed 20261003). Median and linearly interpolated p95 are calculated from all recorded samples.

Short texts are under 120 characters, medium texts are 300–600, and long texts are 1,500–2,500. The inspection report includes individual timings, initialization and warmup times, observed host load, and detection observations.

Use these measurements to understand the component and plan validation. The AI governance brief explains how inspection fits into the governed request path.

Inspect the samples and calculations.

The inspection report contains 200 per-case detection observations. Use the supplied test inputs and detected character spans to review detection counts, and the 900 timing samples to check the reported summaries.

The email-domain comparison changes only the detection examples. Both versions use the same timing texts, so the inspection-time comparison stays consistent.