Leak test
The project’s own 12-prompt test planted 35 entities and observed 0 reaching Upstream. It is not an adversarial audit.
Trust
Every public claim maps to an implementation test or a named synthetic evaluation.
The project’s own 12-prompt test planted 35 entities and observed 0 reaching Upstream. It is not an adversarial audit.
On a 220-sentence synthetic project-authored set, PERSON recall measured 96.9% and precision 61.9%.
Measure rule and model overhead on your own deployment using bench_scrub.py; results depend on hardware and workload.