01:56 · repliedI've updated your doc ([AMS 구축 제안서](https://claude.ai/code/artifact/802ee2da-8525-48b5-bcbc-a761b814d36e)).
**26% figure:** the bullet on the run without the initial report is gone. The throughput section's source note now just calls it a separate run of the same 300 cases with the same model; it's only used there for timing.
**Video PII in route A now uses blackbox-pi:**
- **How it works:**
- The whole clip is de-identified inside KB first: faces and plates are tracked across the video and masked with a solid fill that can't be recovered.
- Anything detected once is masked for its entire time on screen, so nothing gets blurred late or flickers.
- The frames sent out are extracted only from that de-identified clip. The original stays inside for adjusters and the CV measurements.
- The dashcam overlay (speed, time, GPS) is still handled separately: speed and relative time go out as text, and the band itself is blacked out.
- **The diagram** now shows a "영상 비식별화 (blackbox-pi)" lane in place of the per-frame masking box. The render checks out.
- **Readiness, taken from blackbox-pi's own handoff notes:**
- The air-gapped batch mode is built.
- Recall on real dashcam plates hasn't been validated. Running it on the KB 300 and reviewing the debug overlay is now part of the pre-build verification.
- The detectors are AGPL (ultralytics and the YOLO weights). Installing it at KB as delivered software means making the source available. The alternatives are a commercial license or swapping in permissively licensed models, which only touches two files. I pointed to your existing license review.
- **Cost:** de-identification throughput on an L4 hasn't been measured. I budgeted one extra L4 per 1,000 cases a day. That puts route A at $168–268k a year for 1,000 cases a …