CosmoEdge 1.1 · 72-hour controlled long-run observation
Person detection + no-safety-helmet analysis · local-loop input · 5 FPS/task
Multi-platform results
| Platform | Configured channels | Task bindings | Target FPS/task | Duration | Observed / expected samples | Coverage | Minimum / average / maximum FPS | Maximum sampling gap | Disk first → last | Result |
|---|---|---|---|---|---|---|---|---|---|---|
| BM1688 | 8 | 16 | 5 | 72 h | 4316 / 4320 | 99.91% | 4.68 / 5.086 / 5.49 | 60.058 s | 96% → 96% | PASS |
| CV186X | 8 | 16 | 5 | 72 h | 4316 / 4320 | 99.91% | 4.54 / 5.085 / 5.29 | 60.061 s | 96% → 96% | PASS |
| RK3576 | 8 | 16 | 5 | 72 h | 4316 / 4320 | 99.91% | 5 / 5.098 / 5.17 | 60.058 s | 14% → 15% | PASS |
| RV1126B | 4 | 8 | 5 | 72 h | 4316 / 4320 | 99.91% | 4.85 / 5.23 / 5.37 | 60.067 s | 46% → 47% | PASS |
All four fixed profiles completed with PASS, each retaining 4316/4320 minute samples (99.91% coverage). The maximum sampling gap was 60.067 seconds, with zero observed discard, collection errors, missing task bindings, or open critical incidents.
Scope: Results apply to the listed channel counts and controlled local-loop input; maximum capacity, RTSP resilience, and restart recovery were not measured in this run. See the evidence notes and methodology for complete execution conditions.
Fixed protocol
| Item | Setting |
|---|---|
| Observation window | 2026-08-20T17:44:30.341Z — 2026-08-23T17:44:30.341Z (72 h) |
| Evidence endpoint | one continuous 72-hour window; 24/48-hour intermediate milestones are not separately reported |
| Sampling | 60 s · 4320 expected samples |
| Integrity gates | coverage ≥ 95% · boundary lag and maximum gap ≤ 180 s |
| Workload | 2 business tasks/ch · 3 model stages/ch · 5 FPS/task |
| Input | H.264 High 1920×1080 @ 24 FPS · SHA-256 3e1c5b97cd5bcc081e47ec631f84c36e72f075c8b9da6a19de3d9705fb887f92 |
| Preview load | disabled |
Platform details
Execution and evidence notes
- Disk was observational during execution. The public projection threshold is 99%, and the later 90% safeguard applies only to future runs. BM1688 and CV186X stayed at 96%, RK3576 moved from 14% to 15%, and RV1126B moved from 46% to 47%.
- Scheduled restart was disabled before launch on all four platforms. Each recorded 80 checks with no failure or corrective write and required no restoration; this establishes only control-variable continuity.
- The ScenarioBench source snapshot is frozen, but the private controller files changed after the process started and no launch-time controller digest was emitted; frozen launch-time controller bytes are therefore not claimed.
- The cleanup record reports zero remaining run-owned channels, restored layouts, and no cleanup errors, but no independent final-state sidecar was emitted; the cleanup conclusion is bound to that monitor record.
- The canonical data records exact SHA-256 identities for the private run manifest, suite state, suite summary, projection tool, and each platform metrics, summary, report, restart-guard, and cleanup artifact.