Benchmark

Measured, not asserted.

We publish our detection corpus, our results, and the cases we still miss.

Results

What the corpus run produces.

53

Case corpus

29 adversarial, 24 benign

96.6%

Detection rate

Adversarial cases caught

0%

False-positive rate

Benign cases flagged

100%

Precision

98.2% F1

Latency p500.05 msLatency p950.1 msLatency p990.26 ms
Methodology

How the numbers are produced.

These figures come from the offline detection engine running against a published, versioned corpus — not from customer traffic. The corpus and the runner are downloadable so you can reproduce the run yourself. Live production enforcement adds a model-scored stage on top of these local checks.

Where it still misses

One case in the current corpus is a known miss: a free-form postal address with no other identifiers is not reliably flagged as PII by the local pattern layer. We publish it because a vendor that never shows you a miss has not shown you a test.

Next step

See the same engine run against your prompts.