Caged AI privacy inspection component study
Study record: privacy-inspection-2026-10-03
Measured: October 3, 2026

Configurations: pattern and language analysis; local AI model; combined inspection.
Test hardware: Apple M2 Max, 96 GiB RAM, ARM64, CPU inference.
Runtime: Python 3.11.13. Fixed detector and model configuration; verified model assets; offline analysis.
Mode: monitor. One inspection at a time. Three warmup calls per configuration, then three measured passes.
Timing workload: 100 synthetic English texts (33 short, 34 medium, 33 long); 300 calls per configuration; 900 retained samples.
Order: seeded text and configuration shuffles, seed 20261003.
Clock: monotonic time.perf_counter_ns. Median uses the two middle values for even sample counts; p95 uses linear interpolation at index (n - 1) * 0.95.
Text lengths: short under 120 characters; medium 300-600; long 1500-2500.

Detection check: 50 synthetic labelled values. A catch needs the expected entity type and a detected character span overlapping the labelled substring.
The follow-up corpus changes only ten reserved email domains from .invalid to example.com. Both corpora, both scored results and 200 repeated per-case observations are retained.

Measurement scope: warm CPU privacy inspection in monitor mode using synthetic text. Network transport, authentication, routing, masking and model-provider responses are separate stages of an AI request.
The report records host load, initialization and warmup times alongside each measured configuration.
Detection counts refer to the seeded values in the supplied labelled synthetic datasets.

Use the supplied files to review sample counts, timing summaries, detection scoring and input changes.