05 · Why this exists
Warehouse CV · release qualification bench
A model is not ready when training stops.
It is ready when conversion, device latency, artifact size and lineage can all say yes—and the system can explain a no.
Loading observed run evidence…
QUALIFICATION / LIVE EVIDENCE—
- 01Checkpointbest.pt
- 02ConvertONNX → INT8
- 03Measurex86 + ARM64
- 04GateQ1—Q5
- 05Promoteonly if every gate passes
01 · The decision
Make the release system say no.
Change a hypothetical budget. The observed measurements stay fixed; only your scenario changes.
Scenario simulation—
Waiting for evidence.
02 · The model, in your hands
Bring a warehouse frame.
The image goes to the live scale-to-zero ONNX API, is processed in memory and is not retained.
Choose an image to begin.
Your frame and detections appear here.
03 · The evidence chain
Every parent of the verdict, on screen.
Open any link in the chain to see the full immutable identifier behind the short label.
04 · The fleet
One model. Different runtime consequences.
Reference and trend lanes stay visibly separate; ARM64 container timing is not presented as physical-device latency.
| Target | Lane | Quality | p50 | p95 | Size |
|---|
06 · Method and scope
What this demo proves.
Observed numbers come from versioned artifacts. Missing evidence renders a dash, never a convenient fallback.
Developer contract
curl -s https://peregrine.devopsdive.com/api/healthz
curl --data-binary @warehouse.jpg -H 'Content-Type: image/jpeg' \
'https://peregrine.devopsdive.com/api/predict?confidence=0.25'