BM1688 · Validated VLM performance
Last passing: 6 channels
Scope: the validated VLM performance boundary comes from 60-second short-run steps. Each added route passes task-local readiness before the executed 0.1 FPS-per-channel, 80% achievement gate. This is not maximum-capacity, long-run, or production-profile certification.
| Channels | Hold | Target FPS/ch | Current new-route FPS | Minimum active-route FPS / target ratio | Avg discard | Accelerator | CPU | Memory | New-route readiness | Executed gate | Failure reason |
|---|---|---|---|---|---|---|---|---|---|---|---|
| 1 | 60 s | 0.1 | 0.09 | 0.09 / 87.1% | 0% | 60% | 3% | 35% | PASS | PASS | — |
| 2 | 60 s | 0.1 | 0.1 | 0.1 / 104.4% | 0% | 0% | 6% | 34% | PASS | PASS | — |
| 3 | 60 s | 0.1 | 0.09 | 0.09 / 87.1% | 0% | 54% | 7% | 35% | PASS | PASS | — |
| 4 | 60 s | 0.1 | 0.09 | 0.09 / 86.8% | 0% | 100% | 8% | 34% | PASS | PASS | — |
| 5 | 60 s | 0.1 | 0.09 | 0.09 / 87.1% | 0% | 100% | 10% | 35% | PASS | PASS | — |
| 6 | 60 s | 0.1 | 0.1 | 0.09 / 86.6% | 0% | 100% | 13% | 35% | PASS | PASS | — |
| 7 | 60 s | 0.1 | 0.1 | 0.07 / 69.5% | 0% | 90% | 13% | 35% | PASS | FAIL | minimum active-route FPS ratio 0.695 was below the 0.8 gate |
Minimum end-to-end acceptance
| Model load | Task creation | Valid inference | Event/alarm | Task recovery after restart | Result |
|---|---|---|---|---|---|
| PASS | PASS | PASS | PASS | PASS | PASS |
Frozen identities
| Candidate package SHA-256 | VLM model SHA-256 | Tokenizer SHA-256 | Engine SHA-256 | End-to-end evidence SHA-256 | Evidence bytes |
|---|---|---|---|---|---|
| d9ef69d53eb3c7a55c281dcdf3eabeb97766d653af67806257654c3ad66b50a4 | 31a03f4845d612b3c9e946468c2032254c42be54594a0848eb5938fc0f4bd767 | cb16258a9df7062f4c4d291a1501b07fb0750c9a8b247c5639dba0369e42f0cb | 33dc3f74c7dd656e6980c97d83181cc660ee72ce73719ff0750e5ef509503bdb | 9750d12e35f54470bb0a349dc4ef7ffb523e9dd35bb0acef28b77871afeae833 | 5551 |
Test source: f4ec2fc51f1753231165f143d14522effbb77125 · tree 2ed191ac67df2fc23c36ffa9933a387683f51732 · canonical data · methodology