| Prepared for | Gallery Condominiums Board |
|---|---|
| By | Elberta Labs |
| Date | 25 August 2026 |
| Part of | Managed Camera Intelligence pilot package |
This appendix sets out how the pilot runs and how it will be judged. The aim is a decision a board can make on evidence, not impression. Every measure below has a baseline, a target, and a method, and each is agreed before launch.
| Included in the pilot | Coordination only | Future roadmap |
|---|---|---|
| Selected existing camera feeds and IntuVision alerts, connected into one reception queue with AI-assisted prioritization, staff training, and reporting. | Camera, network, and access-control vendors, engaged when their work affects the pilot. Elberta does not hire or pay these vendors. | Access-door event analytics and broader consolidation across the building's systems. |
Access control appears only as coordination and future roadmap. It is not a pilot deliverable.
| Phase | What happens |
|---|---|
| Weeks 1 to 2 | Baseline measurement and shadow mode. AI recommendations are recorded but do not change the alerts staff act on. |
| Weeks 3 to 4 | Supervised prioritization and reception training. The queue goes live under review. |
| Weeks 5 to 7 | Measured live operation against the scorecard. |
| Week 8 | Results, capacity findings, handover package, and a board go/no-go report. |
Sign-off is by the board or its designee, on the Week 8 report. Where a result is inconclusive, the measurement window is extended within the pilot term at no additional fee before a go/no-go call is made.
| Measure | Baseline | Target | Method |
|---|---|---|---|
| Alert volume per camera-day at the front desk | Measured in weeks 1 to 2 | At least 50% lower after prioritization | Queue counts, before and after |
| Downgrades judged correct by a human reviewer | Not applicable | At least 95% of a sampled set | Weekly human review of sampled downgrades |
| Concerning events wrongly downgraded | Not applicable | Zero tolerance; any miss triggers a tuning review | Human review plus incident check |
| Staffed shifts using the dashboard | Measured at launch | At least 80% | Dashboard session logs |
| Median acknowledgment time by severity | Measured in weeks 1 to 2 | Improves against baseline | Queue acknowledgment timestamps |
| System availability during staffed hours | Not applicable | At least 95% | Health monitoring uptime |
| Maximum camera load | Not applicable | 12 cameras at 1080p and about 5 frames per second | Load test on the pilot appliance |
| Staff feedback after training | Not applicable | Net positive | Structured survey |
Before the pilot, Elberta ran a short proof of concept at no charge on the Peachtree pedestrian gate to confirm that AI screening works on the existing cameras without per-camera license fees. Over a four-day sample the camera produced about 40 alerts, roughly 10 per day, most of them routine gate traffic and maintenance activity. AI screening separated the routine from the worth-a-look without discarding anything.
These are illustrative observations from one camera over a few days. They are not a general conclusion. The 60-day pilot, measured against the scorecard above, is the actual validation.